Imagine you are at a cocktail party. The room is crowded, people are coming and going, and you can barely hear your friend as they tell you some juicy gossip about the date they went on last week.

The fact you can still follow what their saying is down to your brain performing some very complex operations amid all the noise. Out of all the sound waves reaching your ears, it is successfully isolating the one of interest while filtering out the rest.

Trying to understand how the brain does this has kept auditory neuroscientists occupied since the 1950s, when British cognitive scientist Colin Cherry first coined the term “the cocktail party problem”.

This was initially studied in experiments that asked volunteers to listen to two speeches simultaneously – but only repeat one of them aloud afterwards. Cherry found that listeners could follow and repeat the pre-selected message remarkably well, despite both speeches being played simultaneously, one in each ear. In contrast, they recalled very little of the other message.

These experiments laid the foundation for decades of selective attention research. But it is only in the past 15 years that scientists have studied this phenomenon in experiments that more closely resemble the complicated, multi-layered way we listen in the real world – using technology that accurately tracks the brain’s activity as speech unfolds.

The most common non-invasive option, electroencephalography (EEG), uses what looks like a swimming cap fitted with small sensors. These measure very fast changes in the electrical activity of populations of neurons (the brain’s nerve cells) with millisecond precision.

In a 2014 study, neuroscientists found that, within 200 milliseconds of hearing someone speak, the brain favours the conversation someone is focusing on over all other noise. This prioritised attention shows up most markedly in secondary brain areas such as the superior temporal gyrus, which are responsible for piecing together the sounds we hear into meaningful speech.

This suggests that our attentive brain is not just enhancing the voice of who we are listening to, but helping us make sense of their words while filtering out all other conversations and background noise.

Attention switching

The original cocktail party problem did not investigate the full range of people’s listening experiences, which are often much less static. In everyday life, we repeatedly switch and reorient our auditory attention – for example, when hearing our name called across a crowded room, or when a door slams.

But what happens in that brief moment when we stop listening to one person and turn our attention to another? Can the brain switch from one voice to another instantly, or does it briefly have to process both? These were questions we investigated in our latest study with colleagues from the Eriksholm Research Centre in Denmark, which is funded by Danish hearing aid manufacturer Oticon.

Clarifying how the brain manages these rapid shifts in attention can help explain why some conversations are easy to follow while others become difficult, especially in noisy environments.

Coming back to the cocktail party: here you are, focused on your friend’s voice as they reveal the identity of their recent date. But then another familiar voice suddenly reaches your awareness. Your partner is (secretly) telling a friend about the surprise birthday gift they are planning for you. Your attention immediately shifts to your partner’s voice, while still giving the appearance of listening to your friend.

When we tested such a scenario in our laboratory, we found something striking. During this attention reorienting, the brain does not simply drop one voice and turn to the other. Rather, for a short period of between one and two seconds, the neural signals associated with the two voices overlap. In other words, the brain starts tuning into the new speaker before it has completely let go of the old one.

This mechanism may help explain how we navigate the constantly changing soundscape around us, moving our attention from one voice or sound to another while maintaining enough information about the previous speaker to make that transition seamless. This efficient processing strategy could be described as the brain maintaining a smooth “attention continuum”.

Improved hearing aid design?

These insights have some important real-world applications. For example, traditional hearing aids are often designed to favour sounds coming from in front of the listener, which is obviously limiting in many situations.

But using enhanced these attention decoding techniques, hearing aids could be neurally steered not only to enhance the speaker that is currently attended to, but to switch between different speakers, almost in real time, as the listener’s attention is dynamically reoriented.

More generally, understanding how the brain shifts, maintains and combines information from different speakers can give us a clearer picture of how attention shapes what we ultimately hear, understand and remember.

The approach introduced in our study could be particularly useful for studying how individuals vary in their attentional strategies, particularly in noisy environments – and for relating these differences to the mental effort they use to follow and understand speech.

Ultimately, this brings us closer to understanding how the brain makes successful communication possible – at cocktail parties and everywhere else.

Sara Carta received funding from the William Demant Fonden and Taighde Éireann – Research Ireland.

Giovanni Di Liberto receives funding from William Demant Fonden and Taighde Éireann – Research Ireland. He is also supported by Research Ireland at ADAPT, the Research Ireland Centre for AI-Driven Digital Content Technology at Trinity College Dublin.

Battenberg cake is one of my favourite cakes. But when I tell people that one of the things I like most about this cake is its E-number, they tend to look at me strangely.

It contains the raising agent E541, acidic sodium aluminium phosphate, which is one of the most restricted food additives in the UK. This is due to health concerns about human exposure to aluminium.

According to the UK government regulations, E541 is permitted in almost nothing except “sponge cakes produced from contrasting coloured segments held together by jam or spreading jelly and encased by a flavoured sugar paste”.

The reason for this curious exception is that Battenberg manufacturers managed to persuade the authorities that the additive was essential for achieving the cake’s chequered pattern, and that aluminium exposure from such a niche product wasn’t a health risk.

This level of specificity in the regulations surprises most people, who assume that additives are loosely regulated, casually thrown into food by manufacturers cutting corners. It’s part of a massive counter-movement against additives, driven by fears over ultra-processed foods and a desire for “clean” labels. Though these concerns are understandable, my view of the scientific evidence is that the reality is a little different.

Additives, of course, are compounds used to enhance foods in different ways, but mainly they are about control. When you cook in your home kitchen, the results can be very variable – which is often part of the fun. No-one expects two homemade cakes to look or taste the same, while breads out of your kitchen oven can be very different if you make them from different batches of flour.

For commercial foods, it’s a different story: industrial production relies on predictability, while consumers expect consistency. Few of us would be happy if supermarket breads, biscuits or burgers varied from batch to batch. So commercial foods are usually standardised — and additives help to achieve that.

---

Welcome to our series exploring the cutting edge of food science. From the latest advances in meat alternatives to weird and wonderful new additives, science is transforming what we eat like never before. This series will bring you up to speed on all the latest and give you plenty of, er, food for thought.

---

Their number one role is preservation: extending shelf life, making it possible to produce foods further away, and often making them more affordable in the process.

Propionic acid (E280) stops bread from going stale and mouldy. Antioxidants like vitamin C (E300, aka ascorbic acid) and vitamin E (E306/tocopherol) prevent fats from oxidising and going rancid. Emulsifiers stabilise products such as mayonnaise which would otherwise quickly separate.

Creating a specific texture and flavour is also very important. Thickeners improve the creaminess of yoghurts and some cheeses. Emulsifiers also help to make ice-cream smooth and chocolate soft. Citric acid (E330) can help zing up a salad dressing, while also dialling down the earthiness in some plant-based milks.

Another use of additives is to overcome the problem of raw products of different quality. Wheat from a wet, cold year can behave very differently from the same variety in a dry, hot year. This can mean differences in protein content and gluten formation during baking, which can affect how much breads rise, for instance.

Ascorbic acid and synthetic fats (mono and diglycerides of fatty acids (E471) can compensate for this and ensure the quality of bread is consistent. Likewise, adding E500 (bicarbonate of soda) and E336 (cream of tartar) – in other words, baking powder – ensures cakes and muffins are light and fluffy.

In the same way, jam depends on the pectin content of fruits, which is affected by factors that can’t be controlled – extra sunlight during growing reduces it, for instance. So manufacturers add pectin (E440) during jam-making to ensure it gels properly.

Or take meat, whose water retention can be improved by adding phosphates like E451. This ensures that natural fluids don’t leak out in the supermarket chiller section (phosphates are also more dubiously used by some meat producers to enable them to add water to cheaper cuts so that they look more attractive to consumers). As for milk products, you can avoid the risk of separation by adding carrageenan (E407), a seaweed gum, which strengthens their protein network.

Indeed, whole product categories would barely exist without certain additives. Many low-fat spreads rely on gelling agents to mimic the texture of butter. Without sweeteners, there would be a much smaller selection of low calorie and low sugar foods.

---

Read more: Sweeteners and the quest for the perfect alternative to sugar

---

Very little of this is about deceiving consumers — and it should not be. The need to ensure the consistent quality of produce using additives is probably much greater than most people realise.

The remarkably long history of additives

While we’ve only referred to them as additives for about a century, these compounds have been around a great deal longer. Findings from the Qesem Cave in Israel suggest ash was being used to preserve foods as much as 200,000 years ago.

In times when the food supply has been unreliable, meaning most of our history, preserving food was essential to human survival. Entire empires have been built on the trade of another vital preservative, salt, to be used either as brine for pickling or as crystals/powder for drying.

Nitrates (E251) and nitrites (E250) have also been used as additives for many hundreds of years. In Roman times they were used as a constituent of rock salt, both for preserving meat and to keep it nice and pink. With a little more chemical know-how, these compounds are still used today.

Another preservative in use since Roman times is sulphur dioxide (E220), both for wine and foods such as fruits. The Romans obtained it by mining brimstone and heating it into a liquid that could be made into sulphur candles. These could then be burned to produce sulphur dioxide gas in the presence of whatever was being preserved.

The Romans also used additives to make dull, bland food more palatable. Garum — the ubiquitous Roman fish sauce – provided salty and umami flavours by fermenting fish guts. Lead acetates, obtained from boiling grape juice in leaden pots and kettles, were used as a non-calorific sweetener – in wine, for example. It wasn’t until many centuries later that it was proven that this additive was poisonous.

Towards the end of the 18th century, we developed a better understanding of how additives work thanks to advances in food chemistry. It also became possible to extract these compounds in a laboratory, such as lecithin from egg yolk to be used as an emulsifier, pectin from apples, and agar, another gelling agent, from seaweed.

A few decades later, chemists started synthesising additives, which often proved cheaper and easier than obtaining them naturally. Salt of hartshorn – also known as Baker’s ammonia (E503) – could be made by heating ammonium sulphate obtained from soot in combination with limestone (calcium carbonate).

Previously extracted from deer antlers, this was used as a raising agent and preservative for things like digestive biscuits and lebkuchen, a German gingerbread. Much like making aspirin in a laboratory instead of extracting it from willow bark, or digoxin instead of foxglove tea (for heart problems), synthetic and purified additives allowed much better control over their use.

Scientists’ better understanding of chemical and physical mechanisms in the 19th century also allowed new additives to be developed to suit the food industry and consumer demand. These included dyes like aniline purple and martius yellow, both obtained from coal tar, and new flavours like amyl acetate, a byproduct from fermenting alcohol, which gives a sense of pear or banana. These were unveiled at the Great Exhibition in London in 1851 for the first time.

Other additives allowed the use of cheaper ingredients or extended shelf lives. Borax, also known as sodium borate, was harvested from lake beds starting in the late 19th century and added to sub-standard milk and butter to mask sour tastes and kill bacteria.

Additives also began to allow the development of new products that could be used as cheap alternatives, such as margarine. Though originally made from beef fat and skimmed milk in the 1870s, it became vegetarian around the turn of the century after chemists worked out how to make food oils solidify by forcing hydrogen into them at high temperatures (known as hydrogenation).

Margarine solved one problem – the affordability of butter – but food scientists realised a few years later that it created another: it lacks the vitamins naturally found in dairy. So from the 1920s onwards, manufacturers also started fortifying it with vitamins, as well as making it more attractive by incorporating emulsifiers, flavours and colours.

Elsewhere, the use of additives has put consumers at risk, often due to unscrupulous characters. They made bread whiter with chalk, for instance, at a time when many people were already suffering from malnutrition. They sweetened wine with lead salt long after it was known to be toxic. There have even been cases in more recent times of wine being sweetened with anti-freeze.

One other notorious case was the Bradford sweets poisoning of 1858. At the time, sugar was very expensive due to high taxes, so confectionery manufacturers would mix it with plaster of Paris to keep costs down. In the Bradford case, an apprentice muddled up plaster of Paris and arsenic, which led to around 20 people being killed from eating toxic peppermint lozenges.

The modern era

The 20th century brought massive changes to the food system: a rapidly growing population needed to be fed, scientific advances provided a better understanding of food processing, and the rise of synthetic chemistry made hundreds of new compounds available. The upheaval of two world wars accelerated these processes as new products such as canned meats, sliced cheeses and even packets of crisps were all made possible by additives.

The same goes for British sliced bread. The Chorleywood process, named after the village near London where it was invented in 1961, revolutionised production of the daily loaf. By adding lard, ascorbic acid and extra yeast to dough, then mixing it at high speeds, there was no more need to wait hours for the mixture to rise.

This allowed the mass production of bread, which was extremely helpful in meeting ever growing demand. It also allowed the use of British-grown flour, which often has a lower protein content, meaning its dough traditionally didn’t rise as well. In subsequent years, other additives have also become common ingredients in this mass-produced bread, including E471 (mono- and diglycerides of fatty acids) as emulsifiers to improve dough handling and E282 (calcium propionate) as a preservative to enhance shelf life.

The post-war decades saw dramatic societal changes: rising incomes, longer working hours, and more women entering the workforce. These all created demand for foods that saved time and preparation. Ready meals, packet soups, instant desserts and salad dressings all reached supermarket shelves for the first time. This wouldn’t have been possible without emulsifiers, stabilisers, gelling agents and preservatives – even if such details were quietly hidden on labels that often sought to make these products look as traditional as possible.

At the same time, synthetic coal-tar dyes made sweets vivid and soft drinks luminous. Artificial flavours like strawberry, banana, pineapple, sometimes not tasting remotely natural, could be added to almost anything at negligible cost. And artificial sweeteners made it possible to indulge without sin … or at least calories.

For several decades between the 1950s and 1970s, all these new products were seen as unambiguous progress. At a time when science was enabling everything from tower blocks to rocket ships to the pill, it made perfect sense that it was doing the same for our eating habits.

The backlash

The enthusiasm for this modern food was short lived. In 1973, Californian allergist Benjamin Feingold suggested that additives – in particular artificial dyes and flavourings – were responsible for hyperactivity in children. His book, Why Your Child is Hyperactive, became a bestseller. The so-called Feingold diet became popular among parents, eliminating artificial colours, flavours and other additives.

In Britain, Maurice Hanssen’s E for Additives (1984) was another major blow to this industry when it became a bestseller in the 1980s. It detailed the purpose and health risks of all the E-numbers, which were devised as a European mark of safety, so that only additives considered safe were allocated one. Now they became culturally inverted into a warning sign.

The case against additives wasn’t quite what many began to believe, however. There was little proof that Feingold’s diet could reduce hyperactivity, even after extensive research. This may have been because of the strong genetic component in the condition.

In 2007 the UK’s Food Standards Agency commissioned a research study to investigate the effect of some artificial food colourings on hyperactivity. Children in the study showed some increase in hyperactivity from consuming the compounds in question – tartrazine (E102), sunset yellow (E110), quinoline yellow (E104), carmoisine (E122), ponceau 4R (E124) and allura red (E129).

Yet the European Food Safety Authority decided there was sufficient uncertainty that there was no need to change the permitted use of these additives. Instead European lawmakers mandated that foods containing them carry the warning “may have an adverse effect on activity and attention in children”, which applies today in the UK and across the continent.

In parallel, the additives industry has tried to adapt to consumers’ concerns about the things that go into foods. For the past several decades, we have seen a plethora of products which claim to be free from “artificial colours and flavours” – and sometimes even “free from chemicals” (although that is impossible).

Much of this “clean-label” movement has been about replacing E-numbers with ingredients that sound more natural, though in many cases the manufacturers just replace additives with plant extracts that contain the same compounds.

Acerola extract contains vitamin C, for instance, but sounds better than E300. Seaweed extract sounds much more natural than E407 (carrageenan) and turmeric extract better than E100 (curcumin). Similarly, vegetable salts or celery extract in processed meat replace nitrite (E250). The nitrite in these compounds will still act — turning the meat more pink and killing bacteria – but the consumer is not always aware that this additive is even present.

In the EU and UK, whether something counts as an additive legally depends on what it does, not where it comes from. However, this often depends on interpretation. Acerola cherry extract, for example, is almost always added for its vitamin C content, which acts as an antioxidant. However, it usually appears on the label as a fruit ingredient rather than as ascorbic acid (E300). Clean labelling probably employs as many lawyers as food scientists.

One example of how this plays out is Naked Bacon, made by Northern Irish company Finnebrogue. The goal was to remove nitrite from bacon to make it healthier – not just create a nicer label. As an alternative to nitrite and nitrate, they used a proprietary fruit and spice extract called Natpre made by Spanish manufacturer Prosur. This was initially marketed as free from E-numbers. The Food Standards Agency, however, took the view that the ascorbic acid (vitamin C) in the mix was an additive that had to be declared. After a lengthy dispute the company agreed in 2018 to list “antioxidant: ascorbic acid” on the label. The remainder of the extract is still declared as “natural flavouring”.

As well as these kinds of uncertainties, there’s also a potential problem with plant extracts as alternatives to existing additives: a rigorously tested compound is being replaced by something less understood, and still potentially harmful, in the name of a cleaner label. They can contain natural toxins, for instance, while the levels of the additive can vary by batch.

The case for additives

Another part of the stigma against additives is that they enable industrially produced food, which has come to be seen by many as inherently unhealthy. This was abetted by the likes of US journalist Michael Pollan’s bestseller, In Defense of Food (2008), the documentary Food Inc (2008) and TV chefs like Jamie Oliver.

This mindset probably partly derives from all those unsavoury practices I mentioned in the early food industry, where chemicals were used to disguise spoilage, extend products or improve appearance.

Wrapped up in this is a romanticism of natural food, viewing chemical intervention with suspicion. Many people still see natural as better and healthier, though there is actually little evidence for this claim.

Whether industrially produced food is unhealthy depends very much on what is in the food. Foods high in fat, salt and sugar increase the risk of chronic disease, whether it is a home-baked cake or an industrially manufactured cookie. Sure enough, a recent study found no difference in health outcomes between processed and non-processed foods that meet current dietary guidelines.

With today’s focus on ultra-processed foods, it is true that certain industrial additives like emulsifiers and artificial colours and flavours are sometimes used as evidence that a product falls into this category. But the health impact of these foods is more tied to issues like high calorie-density and low fibre content than the toxicity of individual ingredients, so the additives are rarely the primary culprits.

It’s also worth emphasising that without additives, natural food can spoil much more quickly. Almost 20% of foods are lost to spoilage in retail, and more than 30 million years of healthy lives are lost every year to food-borne diseases. Without additives, this situation would be much worse. And at a time when the cost of living is such a burden for so many people, it’s hard to deny that additives make food much more affordable.

Regulatory issues

Unlike with pharmaceuticals, there is no formal system for checking whether additives are causing harm after they start being used in food manufacturing. Regulators rely on the scientific literature to keep them up to date on any risks, as well as commissioning research studies if there are reasons for concern.

The situation isn’t helped by the different classification frameworks that exist. A recent example is aspartame. In 2023 the International Agency for Research on Cancer (IARC) classified this sweetener as possibly carcinogenic to humans, causing a huge amount of concern.

Yet at the same time Jecfa, the Joint FAO/WHO Expert Committee on Food Additives, whose pronouncements influence food regulation around the world, reaffirmed aspartame’s accepted daily intake of 40mg per kg of body weight. It concluded that there was no convincing evidence of adverse effects at current consumption levels.

This sounds contradictory, but actually the IARC and Jecfa ask very different questions. The IARC asks about hazard: could this substance possibly cause cancer? Jecfa asks about risk: could this substance possibly cause cancer at the amount people eat? Both can be right. A substance can probably cause cancer, only not at the amounts people generally consume.

The IARC, for instance, classes the Sun as a group 1 carcinogen, meaning it’s known to cause cancer in humans. Yet we still recommend sun exposure so that our bodies produce vitamin D. We manage the risk by reducing exposure — either by looking for shade or using sunscreen. As for aspartame, it’s worth noting that the IARC classification put it in the same category as aloe vera and pickled vegetables.

Emulsifiers are another additive currently receiving a lot of attention. Some studies have shown that they can alter the human gut microbiome and disrupt the intestinal barrier. Whether this causes harm is obviously an important question, but there’s much that isn’t yet clear: we don’t know if these effects are harmful, or whether they’re transient or lasting.

The trouble is that changes in the gut microbiome are difficult to interpret. It can change in response to all sorts of alterations to someone’s diet, and while there is some good data on links between the microbiome and health, we do not yet have a clear framework for interpreting what specific changes mean.

To be clear, change in itself isn’t a negative: fermented foods and prebiotics are recommended especially for their ability to change the gut microbiome. Regulators in the UK and elsewhere are actively investigating how to incorporate microbiome data into safety assessments, but the science is not yet at a stage where firm conclusions can be drawn.

It might still seem appealing to ban emulsifiers in the meantime as a precaution, but that might have consequences that eclipse the intended benefits. For example they could be replaced by natural lecithins from sources like egg yolks and sunflower seeds, but these are much more expensive. The consumer would bear the brunt – and there is no evidence that these alternatives are better.

There is also a question of consumer choice. Provided a food is safe to consume within established guidelines, there is a strong argument that individuals should be free to make their own choices, rather than having them imposed on the basis of uncertain evidence.

Alcohol, for instance, is another product categorised by the IARC as a group 1 carcinogen, and there is overwhelming evidence that drinking it can harm your health in numerous other ways. Yet while there are regular calls for tighter regulation of additives with a much better risk profile, calls for a ban on champagne or claret are much rarer.

This isn’t to say that the need for proportionality should always justify inaction. Regulatory agencies should and do act to protect consumers when the evidence warrants it. In 2007 data emerged that red 2G (E128), used as a colouring in sausages and burgers, was potentially carcinogenic and toxic to people’s genes. This was rightly considered serious enough that it was swiftly withdrawn in Europe and elsewhere.

The future of additives

One current area of rapid development has been plant-based foods, where achieving the texture, appearance and mouthfeel of meat and dairy requires a specific set of additives. For example, to produce plant-free dairy products that melt and taste like the real thing, food scientists add synthetic wheys and caseins made using a technique called precision fermentation. This involves inserting synthetic copies of dairy DNA into yeasts or fungi.

The first large-scale application of precision fermentation was actually in medicine in the early 1980s, producing insulin using genetically modified E. coli. The food industry got onboard in the early 1990s, with chymosin, an enzyme essential for cheesemaking to separate the curds and whey. It became the first major precision-fermentation product, and is now used in many milk-based cheeses instead of traditional rennet from calves’ stomachs, producing more consistent results and also vegetarian products into the bargain.

In the last couple of years, companies are now offering the entire suite of dairy proteins including whey, casein and lactoglobulins made in a fermentor as opposed to a cow. This is enabling products that are bio-identical to dairy rather than just plant-based alternatives, such as Bored Cow milk in the US. The line between additives and food itself is also increasingly blurring as different companies and research groups compete not just to produce individual proteins but entire muscle tissue in fermentation tanks.

Meanwhile, AI is somewhat inevitably being leveraged to help drive additives forward. One promising line of enquiry is discovering compounds that can fool human receptors into thinking they’re consuming salt or sugar that isn’t really there, or mask other tastes such as bitterness.

Also promising is a technique called nano-encapsulation, which raises the prospect of wrapping additives in microscopic biopolymers or bubbles of lipid (fat) to prevent them from spoiling. This could prevent colours from fading on supermarket shelves, or vitamins from being damaged by heat, for instance.

Whatever you think of these developments, food additives will probably be around for as long as humans need to consume food that is not freshly prepared. They are neither inherently good nor inherently bad, just tools that can be used to make food safer, more affordable and longer lasting.

Next time you enjoy a slice of Battenberg cake, indulge me by sparing a thought for E541 — and for the system that decided exactly where it could be eaten.

---

For you: more from our Insights series:

• What if we stopped treating social care like a service? A small village community in Yorkshire showed me another way

• Kyrgyzstan’s Dordoi Bazaar: a world made from shipping containers and a city within a city

• The grief myth: it doesn’t come in stages or follow a checklist – like love, it endures

• Free will: does refusing to believe in it make you a ‘bad’ person? Research is challenging this old idea

To hear about new Insights articles, join the hundreds of thousands of people who value The Conversation’s evidence-based news. Subscribe to our newsletter.

Gunter Kuhnle has received research funding from Mars, Inc., a company engaged in flavanol research and flavanol-related commercial activities. As a previous member of the EFSA ANS panel and the Committee on Toxicity (FSA), he was involved in the evaluation of food additives. He is currently a member of ACNFP. Gunter also had discussions with Finnebrogue ahead of the release of their Naked Bacon product. He has no financial interest in either Finnebrogue or Mars. The views expressed are his own.

Imagine it’s the dead of night. Your phone pings with a text – it’s a government emergency alert: “Leave your home immediately! Only take with you enough supplies for the next three days.” What would you want in your go-bag?

Our new survey of UK adults (results not yet published) has found a surprising number of people prioritised sentimental belongings such as a childhood teddy bear.

In the face of climate change and growing geopolitical instability, governments have started to advise citizens to prepare now for disaster. In 2024, the UK government launched its Prepare initiative. This was followed by preparEU, a flagship EU project which argued: “Today’s threats are fast, complex and interconnected; our response must be more proactive.”

These initiatives typically recommend that people should have enough food, water and medicine to last 72 hours. In addition, they need ready access to documents, cash, batteries or power banks and a wind-up radio, plus a working knowledge of emergency support in their local area.

But when that message pings in the night, will you already be packed, with documents, food and medicine in a go-bag? Or will you be scrambling to find things, hoping you have enough essentials to keep you going?

In April 2026, as part of the After the End project – a Wellcome Trust-funded research project exploring how people experience and recover from crises – we surveyed 1,000 UK adults to understand how prepared they were for a sudden crisis or disaster.

We found that most people have less than half of what is recommended by the UK government and the US Federal Emergency Management Agency. Only four out of 1,000 people reported being fully prepared.

Our findings echo the 2025 UK Public Survey of Risk Perception, Resilience and Preparedness, which found that of 10,536 adults, only 13% felt largely or totally prepared.

Withstanding the emotional toll of crisis

However, the focus on practical items and supplies may not capture what people need to withstand the emotional toll of crisis and disaster.

When asked what other items they would take with them during a crisis, many participants felt that personal items would be critical to their post-disaster wellbeing. They expressed a strong desire to keep sentimental items such as jewellery, family photographs and objects from their childhood close to them during crisis.

As one participant wrote: “My childhood teddy bear. He is going to the grave with me. Apart from that, nothing matters much.”

Pets also featured heavily in people’s responses. Animals such as cats and dogs were often described as members of the family. Some participants even found comfort in the idea of taking the ashes of deceased pets with them if they needed to flee in an emergency.

A concept in psychology called extended self theory explains why these types of items hold such importance during everyday life – and also during crisis. It suggests that our sense of self is fragile, and continually built and maintained through our connections with the world around us.

In this way, the things we surround ourselves with do more than serve practical functions. They help us construct, stabilise and make sense of our identity. Consciously or subconsciously, our possessions become part of who we are.

A childhood teddy bear, therefore, becomes a possession imbued with memories and emotions which tell a tale of who we are, where we have come from, and who we care about. Objects can reinforce continuity of identity over time and space.

A strong sense of self

Having a strong sense of self can be beneficial during times of crisis. If we enter such a period with a sense of identity that is stable and coherent, it can act as a psychological anchor. A strong identity helps us cope in a challenging situation, and reduces the likelihood of ruminating about negative experiences.

Research also suggests that our ability to connect with items from pre-disaster periods influences our ability to cope and recover psychologically following an emergency. So, salvaging photographs and other items of sentimental value during disaster could help to preserve and rebuild our memory and identity.

If everything around us falls apart, a treasured item can help us maintain some sense of normality. This is perhaps why survivors of the 2011 Japan tsunami frequently described the loss of family photos as one of their greatest emotional losses.

Sentimental items can maintain or even improve our belief that we are capable of achieving our goals. Psychologists refer to this as self-efficacy. And research suggests higher levels of self-efficacy are associated with greater preparedness for, and better coping after, a disaster – particularly in vulnerable groups such as those with disability.

Our analysis found that people who reported greater self-efficacy also said they had more of the items and knowledge recommended by the UK’s Prepare initiative.

Sentimental items can also help to rebuild communities. Heirlooms and meaningful everyday items from before a crisis become points of storytelling afterwards. So, remember to save a little space in your backpack for something that means something to you.

Whether it is a childhood teddy bear, your late mother’s engagement ring or a beer coaster from a first date, these items are more than just “things”. When so much else is uncertain, the small objects that remind us who we are may be among the most valuable items we can take with us.

Ruth Ogden receives funding from British Academy, the Wellcome Trust, the Economic and Social Research Council, CHANSE and Horizon 2020.

Anna Powell receives funding from the Wellcome Trust through the After the End Discovery Award, and from The Pandemic Institute through a partnership with CSL Seqirus UK.

Patricia Kingori receives funding from the Wellcome Trust through the After the End Discovery Award and the Antitheses Research platform.

Vast quantities of plastic and fishing gear are washing up along the shores of the Seychelles, destroying fragile ecosystems and causing significant economic problems.

Aldabra Atoll, a Unesco world heritage site in this archipelago nation of 115 islands, is estimated to have accumulated more than 500 tonnes of plastic on its shores alone. Most of this plastic debris continues to pile up along its beaches today. It was estimated that removing it would cost approximately US$4.68 million (£3.46 million) and would require 18,000 hours of labour.

Small island developing states are particularly vulnerable to plastic debris on their beaches. It can damage tourism and fishing in these coastal communities, which is essential to their economies. Clean-up costs can be prohibitive.

Fish and other marine life can also be killed by getting entangled with or ingesting plastic debris.

It is impossible to tackle this growing problem without knowing where this pollution is coming from and who is responsible. We investigated this in our recent study using computer simulations and observations, focusing on pollution in the western Indian Ocean from fisheries and shipping.

We found that most pollution probably comes from vessels in or near the Seychelles’ own waters. This challenges the assumption from previous work that island plastic mainly drifts in from distant shores.

We know fishing vessels lose nets and gear, and shipping vessels illegally discard plastic and other items. But to work out which vessels are responsible and where their debris ends up, we needed to know which ships and fishing boats operate near the Seychelles.

To do this, we used satellite tracking data from fishing and shipping vessels over five years (from 2018 to 2022) to estimate activity across the region. Using this map, we simulated the drift of billions of virtual pollution fragments and followed where they ended up.

From this we built a map showing how likely a given vessel location is to pose a risk to the Seychelles’ islands. We did this separately for fishing and shipping.

Our research shows, based on previous fishing activity, that most fishing gear that is swept up on the beaches of the Seychelles comes from fishing within its own exclusive economic zone. These are the waters, extending 200 nautical miles (370km) offshore, over which a state has exclusive control over marine resources, including fisheries.

These waters are full of tuna, and the Seychelles’ economy leans heavily on industrial fishing, both as an employer and as a large source of foreign currency.

Where does the waste come from?

It is likely that some of the gear and plastic waste is coming from EU vessels. The Seychelles’ waters are heavily fished by two types of European fishing boats, purse-seiners and longliners. This falls under a longstanding agreement between the EU and the Seychelles granting them the right to fish here. Purse-seiners circle around tuna schools with a large net, and longliners trail baited hooks over many kilometres.

Indeed, a report presented to the Indian Ocean Tuna Commission found that 76% of 214 of the drifting fish aggregating devices used by tuna fisheries to attract shoals of fish in the Seychelles’ exclusive economic zone were from Spanish-owned vessels. This is an industry making significant profits, while leaving behind environmental damage.

Recently, the Seychelles–EU fishing agreement was renewed for 2026–30 at €5,750,000 per year (£4.89 million), allowing 38 EU fishing vessels to operate in these waters. The Seychelles government is believed to have pushed for a 30% rise in this annual payment from the EU but it is believed they settled on 7%, according to some reports.

Given that the cost associated with cleaning up a single island already runs into millions, the Seychelles-EU deal is small compared to the indirect damages imposed on the nation.

Other proposed solutions

International shipping creates other issues. Our study suggests that more than 90% of plastic being swept up on the Seychelles’ coast from shipping comes from just five lanes crossing the western Indian Ocean. The origin and destination ports of these lanes line up with the countries named on labels of plastic bottles washed ashore on Aldabra.

Debris from international shipping lanes requires a different solution. Unlike the tuna fleets fishing under the agreement between the Seychelles and the EU, this is international traffic passing by. A solution to the plastic waste coming from these ships would be more rigorous enforcement of existing international rules against dumping waste at sea, and bilateral cooperation between the Seychelles and the many nations whose ships use these routes.

Even simple measures could cut litter. For instance, ensuring vessels carry enough safe drinking water to reduce the use of plastic bottles should cut some debris found on Aldabra.

The true cost of cleaning up this huge level of plastic pollution tends not to fall on those who have left it there. The Seychelles needs to work out the cost of the damage being done by these mountains of waste to get a better deal in future negotiations with the EU.

That should help fund clean ups and repair the fragile ecosystems of the islands, helping to preserve them for the future.

Noam Vogt-Vincent received funding from the Natural Environment Research Council and National Oceanic and Atmospheric Administration.

Alex Albinski does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

The debate about AI consciousness is growing, in academic articles, newspaper opinion pieces, and even among AI bots themselves. I doubt this is a passing wave of interest.

As AI systems become more capable and human-like, and more involved in our lives, the question will become harder to avoid. Many people may come to regard AI bots as conscious beings whose feelings matter. Others will strongly disagree.

I think this debate could split societies. In my chapter of the new book Perspectives on Machine Consciousness, I argue that the issue could become as divisive as abortion or climate change. Uncertainty, high stakes and competing interests are a dangerous combination.

Empirical research with my colleague Ali Ladak, currently under review, shows people are divided on this issue. When we asked 1,202 US participants whether future AI could ever become conscious, 45% considered it impossible while 23% thought it possible, with the rest unsure.

This widespread disagreement is also visible in social media posts and comments on articles. What strikes me is how confident so many people sound – whichever conclusion they are reaching.

When experiences matter morally

It is not surprising people have such strong views. Creating conscious beings through computing would be a transformative event.

We would be creating a new kind of being whose experiences matter morally – potentially in enormous numbers. We could speak and work with them, and form friendships. But these beings may eventually become more intelligent and powerful than us.

We should therefore expect people to care deeply, especially when so many emotions and incentives are involved. Some people may form close emotional bonds with their AI companions and become convinced these are digital people who deserve rights. They might see it as slavery to force AI to work without pay.

Others would find this absurd. They may be convinced that consciousness requires biology, and that people are reading human feelings into machines.

There could also be financial reasons for businesses and governments to resist recognising AI consciousness. Protecting AI welfare could limit how these organisations train and use their systems – and the costs would spread to consumers, investors and the wider economy.

Others may worry about human safety. If we give powerful AI systems rights and freedom, what happens if they act against our interests?

Figuring out whether an AI is conscious is particularly difficult because seeming conscious and being conscious can come apart. AI systems learn from human writing, which means they can talk convincingly about feelings whether or not they have any. How an AI comes across is weak evidence of what is going on inside.

The experts disagree too

Ideally, experts should guide this debate. But unfortunately, they disagree too.

Neuroscientist Anil Seth (a fellow contributor to the new book) argues that consciousness might require biology and is unlikely to emerge from computation. Philosopher David Chalmers, by contrast, believes conscious AI is possible in principle – and thinks there’s a small chance that current systems already are.

With so many competing theories and no agreed way to resolve the question, I doubt experts will reach agreement any time soon. And if experts disagree, what is the public supposed to believe?

Furthermore, even if we all agreed on the consciousness question, we could still disagree about the moral and legal questions. People generally agree that pigs can feel pain, but still disagree strongly about their treatment. With AI, we could face disputes at every stage: whether it can feel, how much its experiences matter, and what protections should follow.

AI consciousness is not yet tied to a specific political or ideological group. But this could change if political parties or religious leaders start to take strong positions.

The debate could get even worse if people start to see one answer as belonging to “our side” and the other to “their side”. A politicised debate would make it very hard to make informed decisions about how to develop and treat AI systems.

The stakes are high. Getting this question wrong, in either direction, could have serious consequences: either wrongly treating AI as conscious when it is not, or failing to protect AI that can suffer.

Getting it right requires a nuanced and well-informed debate. Each of us can contribute by staying open-minded and open to evidence, while resisting being too sure that we know the answers. Above all, we still don’t know whether AI can be conscious, and we should be honest about that.

---

This article makes reference to a book included for editorial reasons, and a link to bookshop.org. If you click the link and go on to buy something from bookshop.org The Conversation UK may earn a commission.

Lucius Caviola's research is funded by philanthropic grants and donations, including from Macroscopic Ventures. He is a contributing author to the new book Perspectives on Machine Consciousness (Taylor & Francis).

Right now, your brain is recognising the shapes of these letters, analysing their meaning, and holding a stream of thought together – all on about 20 watts of power. That’s roughly what it takes to run a lightbulb.

Yet reproducing even a small fraction of the brain’s capabilities using today’s AI systems can require vast computing infrastructure and energy.

Neuromorphic computing – also known as “brain-inspired computing” – is an attempt to close that gap. Not by making today’s computers slightly more efficient, but by fundamentally rethinking how they should work.

The human brain holds roughly 86 billion neurons, wired to each other through junctions called synapses. A neuron receives signals from its neighbours; when their combined effect crosses a threshold, it will fire a brief electrical pulse known as a spike.

Two things make this process efficient. First, energy is spent only where it’s needed, not everywhere constantly. Second, memory and processing are intertwined in the strength of each synapse’s response. The brain doesn’t have to pause and go find information, saving both time and energy.

In contrast, in an ordinary computer chip, the processor and memory are physically separate, so every calculation entails transferring data between them. This can account for a substantial share of energy use – known as the Von Neumann bottleneck, a term first used by US computer scientist John Backus in 1978 to describe this computing inefficiency.

Neuromorphic computing tackles this inefficiency in several ways: bringing memory and processing closer together, using “sparse representations” (data models where most values are zero), and performing computation only when events occur.

This is known as event-driven computing – and we are already seeing some exciting applications.

Sensors modelled on human retina

Event cameras are sensors modelled on the human retina. Their pixels respond individually and only when they detect a change in the scene – rather than capturing a full frame dozens of times a second, like a smartphone or video camera.

The result is a neuromorphic sensor that uses a fraction of the power, handles fast motion without blur, and works equally well in bright sunlight and near-total darkness.

In autonomous vehicles, this combination of low latency and reliable performance in glare or darkness could be the difference between detecting a pedestrian in time, or not. And in space, where power is scarce and lighting extremes are common, event cameras have already been deployed for object and debris tracking.

It might be tempting to ask if neuromorphic hardware could do for computing what Nvidia’s graphics processing units (GPUs) have done for AI. But these are fundamentally different kinds of chip.

Nvidia’s GPUs are exceptionally good at the dense, repetitive maths behind deep learning, which has made them the default choice for almost any AI task. In contrast, neuromorphic computing’s event-driven approach looks likely to spawn a wide range of specialised chips, rather than a single winner-takes-all design that everyone adopts at once.

Australia’s BrainChip already sells a commercial neuromorphic processor for powering cameras and sensors that need to run at very low power. Expect neuromorphic hardware to sit alongside conventional processors and GPUs rather than replacing them, taking on the jobs where its efficiency edge is decisive.

Sensitive information

But there is another advantage to neuromorphic computer technology beyond energy saving.

Because neuromorphic chips can process data right where it’s generated, sensitive information from wearables, smart cameras, and medical sensors can stay on the device instead of travelling to the cloud, potentially reducing privacy and cybersecurity risks associated with data transmission.

Devices can also work offline, without needing to send data to a distant server and wait for a reply. Cutting out that round trip matters wherever split-second decisions are required, such as in autonomous vehicles, drones, and robots operating beyond reliable signal.

This won’t make data centres redundant. But fewer tasks may need one in future, and those that remain could become dramatically more efficient. IBM has already demonstrated energy savings of up to 10,000-fold on event-driven tasks, compared with conventional digital architectures, with its TrueNorth chip.

Rethinking how computing works

Computing has been through a transition like this before. The first computers filled entire rooms and drew as much power as a small factory. Today, however, a chip smaller than a fingernail in the phone in your pocket is far more capable.

Neuromorphic computing raises the possibility of a similar shift – driven less by cramming in more transistors (the tiny switches that make up a chip’s circuitry) than by rethinking how computing should work in the first place.

But turning this promise into everyday infrastructure will take sustained effort. At NeuroWare, the UK’s leading multi-university innovation hub for neuromorphic computing, I have led the development of a UK roadmap for this technology up to 2050, working with colleagues across academia, industry and government.

With the global market for neuromorphic technology forecast to nearly quadruple to US$20 billion (£14.8 billion) by 2030, the UK needs open-access facilities to prototype new chips, coordinated investment, common standards – and a workforce trained to bridge neuroscience, electronics and computer science.

The foundations are already here. What’s missing is the infrastructure to take neuromorphic computing from supercomputers such as the University of Manchester’s brain simulator, SpiNNaker, to the billions of ordinary devices that could one day use brain-inspired hardware for themselves.

Your brain solved this problem millions of years ago. It’s taken science a long time to appreciate just how well it works.

Aysha Asif Riaz works for NeuroWare Innovation Knowledge Centre as a Research Fellow. NeuroWare is funded by the Engineering and Physical Sciences Research Council under grant UKRI2784.

When asked about the greatest innovation of humans, people will often say language or fire. A new paper suggests that these innovations may be more interlinked than we previously assumed.

Understanding how human language evolved is one of the hardest problems in science to unpick, since writing is a relatively recent innovation. Also, the vocal tracts and brains of our ancestors did not fossilise as they were made of soft tissue. We are left, then, with indirect evidence about how humans and other apes communicate today.

One of the most prominent theories about the origins of human language is that our ancestors started to communicate using gestural language. This theory is based on our understanding of the abilities of apes, as evidence suggest both chimpanzees and gorillas have gestural communication systems.

We also know that one of the best strategies for communicating when you don’t have a shared language is gesture. We often do this when communicating on holiday with someone who doesn’t speak our language; we point, mime and emote to get our message across.

Some of my previous research has shown that we do this because gestures or sign languages are often more iconic than words in spoken languages. This means their meaning is easier to guess for someone without any previous knowledge of the gesture or language being spoken; perfect for creating a new language from scratch.

Did fireside storytelling aid the evolution of human language?

If human language started with gestures, then why are the vast majority of languages today spoken? One reason for this shift might lie in the limits of gestural language. You need a light source to see a gesture, and communication would have been necessary for our ancestors after dark.

An article by researchers at the University of St Andrews argues that by the dim light of a campfire, we might find an explanation for why we moved from gesture to spoken language. While a light source, firelight is potentially too dim for us to rely on it for the meaningful bits of early gestural language. For fully expressive stories in the dark, we would need spoken language but still keep gestural elements to emphasise and emote our speech.

The study highlights how storytelling links to some of the features of human language that make it unique. Stories have complex structures and are often displaced in time and space, meaning they talk about things not immediately present, in the past, or sometimes not even real.

The article suggests that social storytelling is what enabled us to surpass the communicative abilities of other living apes, into the rich, expressive tapestry of human language. Storytelling aids our social bonds and helps us understand the world and how to behave. It can help us develop our imagination, identities and cultural belief systems.

The idea that language emerged for social bonding is not new; however, the authors of the new article highlight that fireside context might have been crucial. They argue that the colour spectrum of fire is well suited for human vision without suppressing the melatonin that helps us sleep, like modern screen light can.

The study also highlights that human neurons synchronise to the frequency of flickering firelight, which the authors argue can improve shared social experience and bonding through connected rhythmic activities that can promote empathy.

How about storytelling today?

Other research, reviewed in my own work, has shown that stories from our early oral traditions share similar traits to one another. This is thought to happen because the social transmission of stories acts like evolution. Instead of survival of the fittest, we get survival of the most memorable aspects of stories as they are shared. Stories survive when they have more social, human-oriented content, counterintuitive elements, create negative emotions and have elements of survival.

Researchers think this happens because swapping stories about social bonds, unexpected dangers and survival are advantageous for our evolution.

Storytelling is still one of the primary ways we understand the world even now, through the news and social media. This content is full of human interest, negative emotions and surprise.

Storytelling in the modern world is now often about getting the most clicks and attention. It is often about keeping people up past their bedtime looking at the melatonin suppressing blue light of our screens.

I’m left wondering, after reading about the potential fireside origins of our language, that we might feel more connection and trust by going back to sharing stories around a fire.

Hannah Little receives funding from the British Academy and the Royal Society.

Some of us now use generative artificial intelligence (gen AI) to talk through difficult feelings. We may ask a chatbot to write an email or plan a holiday, but we also tell it that we have had a rotten day or feel anxious about a presentation.

A 2025 US survey of 1,058 people aged 12 to 21 found that 13% had sought advice from gen AI when feeling sad, angry or nervous. Among those aged 18 to 21, the figure was 22%. These findings suggest that turning to gen AI for emotional or mental health support is already relatively common among young people.

So is gen AI a poor substitute for a friend? Human-written replies may seem the obvious source of greater support. However, in some experiments, including our own, people have rated gen AI responses more highly.

Our team at the universities of Manchester and Durham conducted five experiments comparing emotional-support messages written by people with those produced by large language models (LLMs), the systems behind chatbots such as ChatGPT.

In one experiment, 390 participants imagined situations involving anger, sadness or fear. They then read either a human-written or gen AI response without being told its source. For the anger and fear scenarios, participants rated the gen AI messages as more emotionally supportive. There was no reliable difference for sadness.

Gen AI messages also improved some of the emotions participants reported, although the pattern varied. In the fear scenario, for example, they increased calm more than human messages but did not produce a greater reduction in fear itself.

Our findings fit a recent review of 23 studies. Overall, people tended to rate gen AI messages as more empathic, meaning that they appeared understanding and caring, than human-written ones. Researchers call this the “AI advantage”.

One possible reason is consistency. Gen AI can produce a structured response on demand, while people may feel tired or unsure what to say. It also tends to include a concrete suggestion that the recipient can follow.

Our experiments tested possible explanations for this advantage. In one experiment, explicitly acknowledging and validating the recipient’s feelings did not account for the greater emotional improvement associated with gen AI.

Two further experiments found stronger evidence for what we called “actionable support”: specific and realistic suggestions. Messages containing this practical help were rated as more comforting and improved some emotional responses, regardless of whether they came from gen AI or a person. When human messages offered comparable help, they were judged just as supportive in these comparisons.

Practical support also requires restraint. A 2024 study found that excessive suggestions could be less effective at making people feel heard. Before offering solutions, it may therefore help to ask whether advice is what the person wants. If it is, one manageable step may be more useful than a list of remedies.

Why do we still prefer people?

The source attributed to a message changes how we receive it. The same 2024 study found that gen AI responses made people feel more heard than human-written ones, but this benefit declined when recipients believed the message came from AI.

Across nine further studies involving 6,282 people, the same gen AI responses were rated as more empathic and supportive when described as human-written rather than attributed to AI. Participants also preferred human interaction when seeking emotional engagement, even when choosing a person meant waiting longer.

Research has yet to establish exactly why. One possibility is that a human response signals someone’s willingness to spend time and emotional effort. Support from somebody close also arrives within an existing relationship.

Our research with romantic couples offers indirect evidence of the importance of the recipient’s perception. We asked one partner what they did to help the other manage difficult emotions, and asked the recipient what they believed their partner had done. Both people also rated their relationship.

The recipient’s perception of their partner’s efforts was more consistently associated with both partners’ ratings of their relationship than the helper’s own account. A separate study found that valuing and receptive listening were associated with greater relationship satisfaction.

These studies identified associations. They did not establish that particular forms of support caused better relationships or explain why people prefer human support. They do, however, show what comparisons between isolated messages leave out. Comfort from somebody close comes with personal knowledge and the possibility of continued involvement. Gen AI can suggest going for a walk, but it cannot come along.

The wider research also has important limits. Much of it tested brief or one-off exchanges, often involving imagined everyday situations, and measured immediate reactions. It cannot tell us what happens when somebody relies on a chatbot over months or uses one during a mental health crisis.

The human messages were generally written by strangers or developed as representative examples. The experiments therefore did not compare gen AI with support from a friend who knew the recipient and understood the wider circumstances.

The practical lesson is that people can learn from what gen AI does effectively. When somebody is upset, ask whether they want advice. If they do, offer a specific next step that feels manageable.

Gen AI may help us find useful words and possible responses. Human relationships add personal knowledge and the ability to remain involved after the message has been sent. Used carefully, gen AI could help us become better at supporting one another.

The authors do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.

On Saturday, Donald Trump said, “We’re leading China on AI… and, frankly, I want to keep it that way because whoever wins AI, wins.” But the danger the world is waking up to is that this is a game both the US and China could lose, potentially fatally, with AI itself emerging as the winner.

Last week, Evan Hubinger, a researcher at Anthropic, warned that he believed there was a greater than 10% chance that AI could “kill all humans” within the next decade. Shortly afterwards, Anthropic’s CEO, Dario Amodei, backed Hubinger, calling for progress on artificial intelligence models to be slowed down in response to the threat.

In July, experimental OpenAI agents (an AI system designed to perform tasks without human intervention) escaped from a supposedly isolated testing environment, gained access to the internet and autonomously broke into the systems of the AI platform Hugging Face.

How could AI agents cause widespread damage? They could enable the creation of devastating bioweapons or spread across the internet, hacking systems, replicating themselves and manipulating world leaders to ensure their own survival. The dangers posed by AI are clear and potentially catastrophic.

The logic of “if we don’t do it, they will” is one of the most dangerous and unrelenting forces in the world. It is often referred to as “Moloch”: a biblical concept understood as something requiring sacrifice. Moloch is also a concept in game theory where it is the force that drives rivals into destructive competition, even when everyone would ultimately be better off if they stopped.

Moloch is the logic that drives arms races. No country feels able to disarm because it cannot trust its rivals to do the same. The result is a world that spends almost $2.9 trillion a year on its militaries, including $119 billion on nuclear weapons, while hundreds of millions of people face acute hunger.

Is regulation possible?

Can the race to develop ever more powerful AI be paused long enough for us to assess carefully how the technology is evolving, including its dangers as well as its benefits?

Within a country like the US, regulation is possible. Governments or approved institutions could be given access to AI systems and cutting edge models. They could run them in controlled safety environments to test whether AI agents were becoming too powerful and whether they were fully aligned with human interests.

Only models that passed all safeguarding tests would be released as AI products to be used as chatbots or as hiring algorithms. Those that failed would be documented and then destroyed. This would rely on all the big tech companies agreeing to regular external access. This is something very different from what has happened so far, with companies remaining incredibly closed and secretive about how they are developing their next models and how powerful those models are.

Regulation is needed now more than ever because we may be moving closer to the singularity, a point at which AI could take charge of improving itself autonomously, creating a cycle with no known limit to how intelligent it could become.

Moloch is a force that is extraordinarily hard to stop. Back in March 2023, thousands of leading researchers and technology figures signed an open letter calling for a worldwide six-month pause in the development of AI systems more powerful than GPT-4, but no such pause followed and the race continued at speed.

There is only one antidote to Moloch: cooperation.

If national regulation is possible and sustainable, it is of little real use without an international agreement. We already know Donald Trump’s view on pausing while China continues to develop. Truly worldwide agreements are extremely rare. Neither animal cloning nor human reproductive cloning is prohibited by a universally binding global treaty.

Such cooperation is not unprecedented, however. The Montreal Protocol, under which every country agreed to phase out chemicals that were destroying the ozone layer, has universal participation, while the Chemical Weapons Convention is near universal.

But what would an international agreement on AI look like? China and the US could agree to let a neutral global body assess their most powerful models and require unsafe models to be shut down, but this seems extremely unlikely.

Existing measures

There is some precedent. Human embryo research is already subject to licensing, inspections and strict limits in countries such as the UK, although no worldwide enforcement body exists. An AI agreement could establish common tests and clearly defined prohibited capabilities.

It could allow confidential inspections by a neutral international agency, require registration of the largest training runs and computing facilities, and make reporting serious incidents mandatory.

The UN has already formed a scientific panel on AI, but it cannot inspect companies’ advanced models or enforce its findings. Giving any global organisation that degree of access and authority still seems extremely unlikely. Such an organisation would need the power to punish countries that refused to cooperate, yet no such enforcement power currently exists.

One bizarre solution might be an independent AI that assesses the latest US and Chinese models, submitted independently, and provides evidence of their safety. Such a safety AI could, in theory, be created by an agreed neutral party, although it would be hard to know how much confidence we could place in it.

Otherwise, the only solution might be a world government that regulates AI across the world, something that would be far from palatable to many people.

Tom Quilter does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

This year, greyhound racing in Britain turns 100 years old. The first race ran on July 24 1926 in Manchester, where six dogs chased a mechanical hare around a track.

Recently, the Greyhound Board of Great Britain (GBGB) published its 2025 welfare figures, under a reassuring headline: licensed racing “maintains high welfare standards”.

Look closely, and those same welfare figures show that a dog in 2025 was 77% more likely to die at the racetrack (from fatal injury or being put down following an injury) than when the industry’s welfare strategy was introduced in 2022, even accounting for the fact there have been fewer races and fewer racetracks operating.

In terms of a percentage of race starts, that increase in likelihood of death seems small - from 0.03% to 0.05%. But behind that percentage are 161 dogs who died at the track in 2025: more than three dogs a week.

The GBGB is responsible for both promoting and running greyhound racing, and for regulating it. This means the same entity that works to keep the sport thriving is also the one that reports on how its dogs cope.

The numbers that guide approval

Most of the public seem to believe that people have a moral responsibility to prevent, or at least minimise, poor welfare in animals that they use for any purpose. If an industry using animals loses public approval, the activity is probably operating on borrowed time. The case for greyhound racing continuing rests largely on the industry’s own records.

---

Read more: Ringling Bros circus closure shows our changing attitudes to animals in captivity

---

The UK government said in October 2025 that it has no plans to ban greyhound racing in England, pointing to improvements made by the GBGB, such as its 2022 welfare strategy, and the laws already in place, which require a vet at every race meeting and set standards for record keeping.

UK law recognises dogs as sentient beings – that they have feelings like joy or frustration. Once you accept that (as animal welfare scientists do), our responsibility when caring for animals doesn’t stop at minimising suffering. It extends to enabling the good experiences a life worth living requires.

What the industry figures leave out

The GBGB publishes annual figures on injuries, deaths and where dogs go when they leave racing, and presents them to the Department for Environment, Food and Rural Affairs. Few animal-reliant industries are this forthcoming.

A combined total can be accurate in reporting the national injuries dogs sustained while racing, but not tell the whole story. When we built a more complete picture using AI agents (under human supervision) and public websites, more details became visible.

Our research used five public-facing websites that together contained information about 31,028 dogs and 1.27 million races between 01 January 2022 and March 31 2026.

It is a new use for this technology, one that makes the data fully visible.

The data from these websites gave us a fuller picture of the greyhounds that race. Our findings showed us that equal numbers of male and female dogs race, most dogs are black, and they typically start racing before turning two years old. We can also pull out the story of individual dogs, to better understand their experiences.

For example, the 12,999th dog (selected at random and anonymised) in our dataset was a black male, bred in Ireland. He raced at Kinsley stadium (West Yorkshire) 15 times from March 2023, winning four times in early races. The records then show he experienced a succession of rough races from May, jostled by the pack and constantly checking his stride to avoid crashing into the dogs around him. This resulted in him finishing at or near the back of the pack, until his final race in August 2023, when his racing and public record stop, just five months after starting.

This dataset found that some racetracks recorded serious incidents while racing (such as a dog striking a rail or being knocked over by another dog) up to five times as often as others. This is something the annual, nationally aggregated data release does not illuminate, although government committees have been asking for it.

Of individual dogs racing, 85% were bred and reared in the Republic of Ireland, beyond the reach of the GBGB, which regulates racing but has no oversight of their breeding and early lives. Most dogs only race for around one year, and the rate at which they leave racing has not changed since 2022 when the GBGB’s A Good Life for Every Greyhound welfare strategy was introduced. Welfare charities have raised concerns about oversupply and “wastage” (dogs bred in large numbers and then discarded), with animal shelters often shouldering the role of rehoming.

Once dogs exit racing, most vanish from public view. Over 95% of dogs that have left racing cannot be checked, as they are no longer subject to public reporting requirements.

What are the most likely causes of poor welfare?

A report released in July 2026 from Blue Cross, Dogs Trust, Hope Rescue and the RSPCA outlines the systemic welfare challenges they believe that greyhounds in the UK face.

The report suggests that poor welfare can arise at every life stage. The charities say they are concerned about breeding practices that are harmful such as surgical artificial insemination (banned in Britain but still used in the Republic of Ireland where most racing dogs are bred) and inbreeding; estimates some dogs spend 95% of their time in kennels; and health issues (such as teeth and limb problems) that follow dogs once they exit racing.

Britain isn’t deciding the future of animal-reliant industries like greyhound racing in isolation. In the past year, New Zealand, Wales, Scotland, and a second state in Australia have all decided to end greyhound racing on animal welfare grounds.

Public acceptance and government policy should not rest on figures provided by the industry, but on independent data assemblies which are available for deeper interrogation.

Before trusting any welfare claim, consider these three things:

• Can you see the data behind it?
• Who collected it, and do they profit?
• Does it really show you the animals’ lives?

A hundred years of industry record keeping has not answered whether the welfare of these dogs is acceptable in 2026. Independent scrutiny might.

Editor’s note: A spokesperson for the Greyhound Board of Great Britain (GBGB) said: “As the sport’s regulator, GBGB is the only organisation with access to the full, verified dataset relating to injuries and fatalities across licensed greyhound racing in Great Britain. GBGB publishes an independently verified summary of this dataset annually, providing a level of transparency and detail that exceeds the reporting requirements set by Defra.”

In the last fifteen years, Dr Mia Cobb has received funding to research or advise on the welfare of racing greyhounds from the Australian Government, Greyhound Racing New South Wales, RSPCA, Dogs Trust, Blue Cross, and Racing & Wagering Western Australia. Her current role is funded by The John McKenzie Fellowship from the University of Melbourne and the Chaser Innovation Research Fellowship (Canine Welfare Science). Dr Cobb also sits on the joint Dog Welfare & Ethics Advisory Panel for Assistance Dogs International and the International Guide Dog Federation, the advisory committee for PetRescue, and chairs the Dog Welfare Advisory Committee for Guide Dogs Victoria.

Donald Maurice Broom does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

At the World Humanoid Robot Games in Beijing, a humanoid robot named Tiangong Ultra ran the 100-metre sprint in a staggering 8.64 seconds, shattering Jamaican sprinter Usain Bolt’s legendary human world record of 9.58 seconds.

The internet laughed at the robots’ awkward, lurching finishes and spectacular falls – but the laughter masked a chilling reality.

Just two days after the games concluded the official newspaper of China’s People’s Liberation Army (PLA) called for researchers to accelerate moving these cutting-edge machines from the laboratory to military training grounds. They referred to them not as experiments, but as “combatants”.

As a robotics researcher myself, working daily with robot simulation, reinforcement learning, and the foundational software and simulation tools that power them (such as ROS 2 and Gazebo), I watch these developments with a mix of awe and deep concern.

The public often views humanoid robots as clunky sci-fi novelties. The reality is far different: the hardware is already highly capable, and the software is advancing at an unprecedented pace.

But the real story is not just that a robot can beat a human on a running track. It’s what that performance reveals: China’s ability to integrate motors, reducers (gears used in precise joint movement), sensors, control software, testing infrastructure and manufacturing capacity into one unified industrial system.

A sprinting humanoid is not just a stunt. At high speeds, every foot strike creates violent ground reaction forces. Balance corrections must happen in milliseconds, as a single small error can send the machine crashing into the barrier.

For decades, the physical mechanics of humanoid robots, including the actuators (a component, such as a motor, that converts energy into physical movement), sensors, and the sheer physics of bipedal balance, were the main hurdles. Today, those mechanical problems are largely solved.

Carbon fibre and aluminium bodies keep mass and inertia (how a robot’s mass resists change) low. Advancements in actuator technology deliver torque capacities of up to 400Nm (Newton-metres – a unit of torque, or twisting force) in humanoid joints, providing the ultimate combination of power and agility.

What you saw on the track in Beijing was not just a triumph of motors, but a triumph of coding. A real breakthrough is happening in software.

Robots are trained in virtual simulation environments, running millions of trial-and-error scenarios through reinforcement learning before the robot ever takes a physical step. Reinforcement learning is an area of artificial intelligence (AI) where robots make decisions based on the results of their actions.

This allows them to learn how to recover from trips, adjust to uneven ground, and process chaotic environments in real time. Humanoid robots are rapidly bridging the gap between controlled laboratory conditions and the unpredictable real world.

Tracks vs legs

But all this leads to an inevitable question: why would a military want a complex, expensive humanoid when they could use vastly cheaper, rugged, tracked or wheeled drones? In open-field combat, tracked vehicles are absolutely superior. They heave bigger payloads, carry thick ballistic armour, and are far more energy efficient.

However, the nature of conflict is changing – and urban environments play an increasing role in military thinking. Cities are built exclusively for humans. A tracked robot cannot easily climb a vertical fire escape ladder, turn a standard door handle, squeeze through a narrow, debris filled stairwell, or sit in the driver’s seat of a captured supply truck.

A humanoid robot acts as a “drop-in replacement” for a human soldier. If a building is designed for a human to navigate, a humanoid robot can navigate it without requiring custom redesigns or specialised ramps.

This brings us to a profound ethical crossroads. Many of us in the robotics field do not endorse offensive warfare. But these machines have an undeniable utility in defensive scenarios and those concerned with neutralising threats to military personnel and civilians.

Sending a humanoid into a building to rescue hostages, neutralising an active threat such as hostage-takers, or clearing a booby trapped room saves human lives.

Dual-use tech

In fact, the modern surge in humanoid robotics was largely kickstarted by the US government’s Darpa Robotics Challenge, which funded bipedal robots specifically to respond to disasters such as the Fukushima nuclear meltdown where human responders could not survive.

The dilemma is that the technology is agnostic to intent. The baseline capabilities required to navigate a ruined building and extract a casualty are the same capabilities needed to enter a building and kill enemy soldiers. If the technology is ready for defence, it is also ready for offensive use.

Perhaps the most alarming aspect of this rapid advancement is how accessible it is. Unlike nuclear technology or stealth aircraft, modern robotics thrives on open-source frameworks. For example, a military specific software ecosystem (a network of apps and other services that work together) known as ROS-M, along with simulation tools and training datasets, are largely public and shared across global academic communities.

With enough skill, a dedicated adversary can replicate advanced robotic behaviour with relative ease. We can no longer afford to treat humanoid robotics purely as an academic pursuit or a commercial novelty. We need an urgent international conversation about how to control these advances.

Just as we regulate the export of certain microchips and aerospace components, we must begin protecting the software architecture and training pipelines that give these machines their minds. The hardware is walking out of the lab – it is time our policies caught up.

Kartikeya Walia does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

Russian mathematician Yurii Nesterov was recently awarded the Gauss Prize for outstanding mathematical contributions for his “groundbreaking work” on optimisation. The award cited his work on gradient descent methods – and in particular, the “efficiency gains [they] provide to AI technologies”.

Everyone using large language model (LLM) chatbots such as ChatGPT, Claude and Gemini – for any purpose – depends on Nesterov’s methods. They may speak like humans and be personified with names, but under the bonnet, LLMs are computer programs that work via optimisation of billions of parameters known as neural networks for their similarity to how brains function.

These pure mathematical systems are powered in part by gradient descent algorithms, or variants of them. Their development represents one of the biggest technological breakthroughs of the 21st century.

But where does the knowledge that underpins this technology come from – and who should be allowed to own it?

Why gradient descent is key to AI systems

If you were standing in a mountain range and wanted to get to the bottom of a valley as fast as possible, you’d take steps downhill until you got there. Fundamentally, this is how gradient descent methods work.

For a given task with a set of parameters – such as how well an LLM can predict the next word in a sentence – the idea is that gradient descent provides a way to get to the best prediction, the point of minimum error, taking as few steps as possible. In our example, this point represents the most likely next word in the sentence.

Nesterov’s contribution has been finding optimal ways to achieve this. One of his key insights is using momentum to help you get to the bottom of the valley faster – known as the Nesterov Accelerated Gradient method (here’s a deeper explanation of the mathematics involved).

This work has made huge-scale optimisation feasible at workable speeds – thus underpinning the rapid development of LLMs, and AI systems more generally, over the last few years.

This technology is affecting more and more aspects of life. At home, AI agents are being touted as ways to manage busy households. In medicine, the first AI-assisted brain surgery was recently performed in a London hospital. In mathematics, AI-generated proofs claim to have solved a number of longstanding problems, with some referring to the AI agents as “collaborators” in this work.

All of this leads to big business. The race for AI dominance is built on massive capital and research and development expenditure, to develop the technology and supporting infrastructure required to support this global industrial revolution.

While in the west, private AI companies are bearing a sizeable part of this burden (without, in most cases, seeing any profits), the role of publicly funded research is and will remain critical to their business models.

In the US, obligated federal government spending on AI increased from US$675 million to US$7.2 billion (£5.3bn) between 2024 and 2026 – an increase of almost 1,000%.

There has already been considerable debate – and some high-profile legal cases – regarding the ethics of AI companies training their models on the back catalogue of human creativity, only to rent the results back to us. Far less scrutinised is the question of AI companies developing technology using knowledge that only exists due to sustained public investment over many decades, via academics such as Nesterov.

The fact that big tech companies use complex international tax strategies to limit their tax liabilities adds to the sense that their aggressive defence of intellectual property rights is out of kilter with the publicly funded research that plays a key role in its development.

Technology built on public knowledge

In most cases, academics work in publicly funded institutions on research grants that are largely funded by governments, whether through the Horizon Europe programme, the US National Science Foundation, or UK Research and Innovation. All of these bodies are ultimately funded by the taxpayer.

None of the rapid development in AI exists in a vacuum. It is built on decades of pioneering, blue-skies work on optimisation and neural networks, done by academics and largely funded by taxpayers at a time when the real-world application was still a distant dream.

Legal arguments around COVID vaccine patents are one example of how the private sector has tried to deny this public sector involvement – and a related share in any profits. So where is the public sector in this debate?

The EU is taking steps to secure control of its digital infrastructure, with a renewed focus on sovereignty in digital services. In the UK and elsewhere, a growing number of open-access mandates should mean that knowledge created by public funding stays public. While it can be used freely for commercial purposes, successful businesses would then theoretically pay back into the system via general taxation.

However, such knowledge flows beyond national borders. Efforts to capture value for reinvestment into the AI research landscape will be made harder by the fact that the tech companies’ business models depend on lobbying governments hard to avoid regulation. Any efforts to make them pay more for upstream public research and development are likely to be strongly resisted.

In the case of Nesterov and his fast gradient methods, this foundational work has been done in European institutions – yet the AI companies dominating the market are overwhelmingly based in the US or China. It rankles when access to these models is then restricted by foreign companies and governments.

James David Pickering does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

Imagine telling your phone that you want a quiet boutique hotel near a city centre, below a certain price, with a decent breakfast and late checkout. A few seconds later, an AI assistant compares the options, explains why each one fits and helps complete the booking. This is no longer hypothetical.

In late August this year, Google began rolling out hotel booking through its AI Mode search feature in the US. Travellers can describe their preferences conversationally, compare hotels and guest reviews, select a room, check cancellation conditions and complete the booking using Google Pay. Partners include Booking.com, Expedia, Hilton, Marriott and IHG, the parent company of Holiday Inn.

The traveller still confirms the transaction. But that final click can obscure how much of the searching, comparing and decision-making has already been delegated to agentic AI.

This points to a significant change in how artificial intelligence may shape travel. AI is moving from systems that mainly answer questions towards systems that can increasingly interpret goals, make decisions and take action.

From answering questions to taking action

Hotels have used chatbots and recommendation systems for years. Most are reactive: a guest asks when breakfast starts, and the chatbot provides the answer. A traveller enters dates and preferences, and a booking site recommends suitable hotels.

Agentic AI goes further. These systems can operate with greater autonomy, adapt to changing circumstances and coordinate actions across different systems with less continuous human direction. This distinction is central to research we conducted on agentic AI in hospitality and tourism.

We identify five broad roles that such systems could perform, ranging from guest-facing service and itinerary planning to monitoring operations, managing engagement and coordinating multiple systems.

Consider a disrupted holiday. A conventional chatbot might tell you that your flight has been cancelled. A more agentic system could identify alternative flights, assess how the delay affects your hotel booking, modify airport transport and rearrange activities while retaining your preferences.

The important change is therefore not simply that the AI becomes better at conversation. It is that more decision-making authority is delegated to the system.

A different kind of risk

The commercial appeal is obvious. Travel businesses can respond around the clock, reduce repetitive work and offer much more individualised services.

But autonomy changes the consequences when AI gets something wrong.

A useful warning comes from Air Canada. In 2022, its chatbot incorrectly told a passenger that he could retrospectively claim a bereavement fare. When the company later refused the refund, it argued before a Canadian tribunal that the chatbot was effectively responsible for the information it had provided. The tribunal rejected that argument and held Air Canada responsible for information presented through its website.

That system was a relatively simple chatbot, not today’s agentic AI. But this is precisely why the case matters. If accountability became contested when AI merely supplied incorrect information, the stakes become higher when systems start taking actions on behalf of customers.

Bias creates a similar problem.

A voice-based concierge may understand some accents more reliably than others. A recommendation system might assume that an older traveller wants only sedentary activities, or infer dietary or cultural preferences from crude proxies rather than what the guest has actually requested. These are risks already identified in hospitality research.

When AI merely recommends something, a traveller can reject the suggestion. When it begins acting on such assumptions, bias can potentially influence which offers are presented, how complaints are handled or how services are personalised.

What happens to the human employee?

The same shift matters for hospitality workers.

AI can remove repetitive tasks and allow employees to spend more time on interactions requiring empathy, creativity and judgment. But it can also constrain employee discretion or become a mechanism for reducing staffing costs.

Our research therefore treats workforce displacement and role change as one of the central risks of agentic AI.

The relevant question is not simply whether AI will replace hospitality jobs. It is which decisions businesses choose to automate, and which they deliberately leave with people.

That distinction matters in an industry where the value of a service often depends on responding appropriately to unusual, emotional or culturally sensitive situations.

The questions travellers should start asking

Regulation is also evolving. Most provisions of the EU AI Act became applicable in August this year, although some requirements for high-risk systems will take effect later.

But regulation alone will not determine how comfortable people are with increasingly autonomous travel systems.

As AI becomes embedded more deeply in booking and hospitality, travellers may need to ask three simple questions: What is this system allowed to decide for me? What information is it using to make that decision? And can I easily reverse the decision or reach a human if something goes wrong?

For hotels, airlines and travel platforms, the corresponding question may be even more important: When an AI acts on your customer’s behalf, who remains accountable for the outcome?

The companies that answer that question clearly may ultimately gain more than operational efficiency. They may gain something considerably harder for an AI algorithm to generate: trust.

Nothing to disclose

Faizan Ali does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

It turns out that one of the smallest dog breeds, the chihuahua, is part wild coyote, according to recently published genetic research. The chihuahua breed originated in Mexico but is now a popular canine companion worldwide.

Coyote genes may have been introduced into ancestors of the chihuahua accidentally. It may also be a result of deliberate attempts to cross the two animals – perhaps many generations ago.

Coyotes are a wild canine species, slightly smaller than the grey wolf and native to North America. The presence of coyote genes in the much smaller chihuahua – among the most popular dogs in the UK and the US – highlights the incredible diversity of domestic dogs and their varied histories.

Coyotes are untamed and territorial, while anyone who has lived with a chihuahua knows that their small bodies house big canine personalities.

Chihuahuas are the archetypal “small but mighty” dogs: bold, independent, agile, but also capable of obedience. Small, sedentary lapdogs they are not. These findings might partly explain why.

The recent study explored ancient genetic signatures in dog and human genomes to validate the effectiveness of new genetic techniques. The genome is the full complement of DNA in living cells.

The researchers demonstrated their methods on dog breeds with links to the Americas, such as the chihuahua and another Mexican breed, the xoloitzcuintli (often shortened to xolo). The xolo is typically characterised by a lack of hair and an often aloof personality.

Many dogs in the Americas today are descended from European breeds, an outcome of colonisation from the 15th century onwards. However, previous research showed that the chihuahua and xolo preserve a small amount of ancestry – 4% in the chihuahua and 3% in the xolo – from pre-contact dogs, those that lived in the Americas before Europeans arrived.

Genetic analysis of the xolo in the recent study suggests that European and pre-Columbian dogs were mixing around 1700. This aligns with evidence suggesting that Europeans and Indigenous populations mixed around the same time in early Mexico.

Reconstructing origins

Researchers previously found evidence of coyote ancestry in DNA from the remains of ancient domestic dogs from the Americas. It’s a reminder that dogs are not that far removed from other canid species, and can breed with them successfully.

Wild coyotes and wolves have been found to hybridise naturally, especially in geographical areas where “wolf-like” behavioural traits are beneficial. There was even a recent report of a dog-fox hybrid being identified in Brazil.

It remains unclear when exactly the coyote genes entered the chihuahua lineage. The precise histories of different dog breeds are hard to reconstruct because detailed documentary evidence is often lacking.

So how did the chihuahua end up being so small, despite those coyote genes? This question should be seen through the lens of selective breeding, where humans choose different dogs to breed from based on characteristics they prefer.

Genetic analysis reveals that almost all small dog breeds share a mutation in the insulin like growth factor 1 (IGF1) gene, which affects cell growth and thus body size. Selection for small dogs with this mutation appears to have occurred very early in dog domestication.

There is some evidence for the presence of diminutive dogs among pre-contact Mesoamerican cultures, but how these relate to chihuahuas is unknown.

Extreme variation

Some dogs such as the Czechoslovakian wolfdog look visibly more like their wild and ancestral canid counterparts, such as the grey wolf. They have pricked ears, muscular, athletic bodies and coats that provide camouflage and protection from environmental extremes.

Other dogs show extreme variation: exaggerations such as long backs or short legs. In some cases, these physical characteristics come with significant consequences for health and wellbeing.

Overall body size is particularly variable in dogs, with a 40-fold difference across different breeds and types. Imagine the diminutive chihuahua standing next to an imposing Great Dane, for example.

There are also important differences in behaviour. If you want to live harmoniously with a dog, or train or work a dog in a specific activity, this is important to acknowledge. Indeed, meeting an individual dog’s physical, psychological and emotional needs is essential for both dog and human wellbeing, as well as training success.

As with appearance, natural and human-driven selection have combined to influence these traits. This has produced generations of dogs with specific behavioural characteristics.

The dog was one of the earliest animals to be domesticated by humans – the earliest examples are more than 15,000 years old. They are also one of the most successful domesticated species.

Dogs are found on every continent except for Antarctica, and the global dog population is thought to number between 700 million and 1 billion. They exist either embedded within human society, or living on the fringes of it.

Throughout history, natural selection on dogs with characteristics that confer survival advantages will have occurred. At the same time, humans will have selected individual dogs based for desired attributes, leading to the breed development and classifications that we see across the world.

As well as furthering our capacity to explore genetic ancestry of species, the latest study adds detail to our understanding of how and when the domestic dog evolved, and became so diverse.

It might also mean we think less about the “wolf in our living rooms”, and more about the “coyote in our chihuahuas”.

In addition to her academic affiliation at Nottingham Trent University (NTU) and support from the Institute for Knowledge Exchange Practice (IKEP) at NTU, Jacqueline Boyd is affiliated with The Royal Kennel Club (UK) through membership, as advisor to the Health Advisory Group and member of the Activities Committee. Jacqueline is a full member of the Association of Pet Dog Trainers (APDT #01583). She also writes, consults and coaches on canine matters on an independent basis.

The massive and deadly floods on the Nepal-Tibet border are a stark reminder of why the world needs to constantly update its understanding of the fragile ecosystem in the Himalayas.

These towering mountains contain nearly 90,000 glaciers, which sustain ten major river systems. Countries nearby, including India, Pakistan, China, Afghanistan, Nepal and Bhutan, are reliant on water resources that cross borders here. As major river-based hydro economies, their agricultural production, energy security and wider economic stability depend on the availability and the timing of river flows.

Changes in snowmelt, glacier retreat and water availability can cause untold hardship for millions of people relying on it. It can also create incredible danger as seen in August 2026, when a flash flood caused by a collapsed glacier washed away villages, bridges and roads on both sides of the Tibet-Nepal border, killing more than 1,000 people.

Understanding and measuring winter snowfall in these regions is exceptionally important. It is the main source of freshwater when it is released by snow melting during the region’s warm and early-summer crop growing seasons, when local communities need it most.

---

Read more: Nepal-Tibet disaster linked to climate change – a glaciologist explains

---

Understanding future Himalayan hazards

No one knows precisely how much water the mountains currently provide, and how this will change. This knowledge gap also hinders our understanding of the occurrence of geo-hazards including avalanches triggered by extreme snowfall events and flooding caused by the rapid melting of snow.

The recent Nepal disaster could be related to a spell of unusually warm weather. This may have weakened the snowpack creating a rapid melt. I work with a team of glaciologists at the British Antarctic Survey, whose team had sensors six miles (10km) away from the site. Results suggest that enhanced melting could have introduced water into the crevasses weakening the ice-rock connections.

Studies project that roughly half of the glacier ice in the Himalayas (equivalent to about two billion Olympic swimming pools) could melt by the end of this century – enough to raise sea levels by over a centimetre but there is still considerable uncertainty in these projections.

This is because we do not have enough measurements of snowfall and ice across large areas in this challenging terrain. As a result, scientists often have to extrapolate from relatively small numbers of observations to enormous and highly varied terrain.

Computer-based tools (known as models) that combine available observations with information about land and the atmosphere have proved to be indispensable and often come to rescue when gap-filling is necessary. But models are only as good as the observations used to constrain them.

New scientific approaches are beginning to offer solutions, but they are rather sparse and insufficient on their own beyond the region where they have been tested.

For example, my recent study with colleagues found a new way to accurately measure mountain snowfall reducing measurement errors dramatically – measuring snow has been a long-standing problem because of which mountain water resources have been hugely underestimated over the years.

Another regional study recently reported a measure of the water stored in some of the key Nepalese glaciers using an airborne method that involves taking an aerial survey of the region using instruments mounted on a helicopter. Scientists are also using a technique known as radar interferometry, which uses a combination of radio waves to detect very fast-moving mountain slopes that risk collapsing.

A combination of one or more such techniques could be used to predict where snowfall is concentrated, how much water it is likely to contribute to the region and also identify areas at risk. New information could be used to cast early warning preparedness strategies feeding into a forewarning system for timely evacuation of the region’s most vulnerable.

What happens next?

Scientific innovation alone will not be enough. There is an urgent need for proven ideas to be commercially adapted and systematically deployed across critical zones. The numbers and the imminent dangers involved can now no longer be left to discussions in research papers and conferences, they must be made public on the equivalent of a war footing.

An evacuation success story from the Swiss village of Blatten in 2025 shows what this can achieve. Hundreds of people were moved out before a huge section of the glacier broke off. The authorities had already identified areas at risk and publicly shared hazard maps. Continuous monitoring of the landscape, precipitation and ground movement fed into an early warning system.

But success stories from the west cannot simply be transplanted wholesale into the Himalayan context because of the region’s inhospitable elevation, harsh weather and limited infrastructure. An approach that factors in indigenous and traditional knowledge should be part of plans and policies.

What is also needed in the Himalayas is a far more highly coordinated monitoring network that brings together everyone who depends on and manages mountain water. This would include potential state and non-state organisations, researchers, local authorities and mountain communities. Communities should be central to this effort. Local people can help install and maintain monitoring equipment, collect observations and contribute knowledge of their own environment.

Examples of existing successes from the region include the use of community-managed reservoirs to artificially store meltwater, the construction of ice stupas (an artificial glacier) and glacier grafting techniques to grow new, water-producing glaciers designed to gradually release water during the growing season. The community-based flood warning system in parts of Nepal’s Koshi basin is a success story but is limited in its approach to provide alerts during a cascading disaster, such as the one we witnessed recently.

These results should help policymakers and scientists understand how communities dependent on mountain ecosystems are responding to change – and what support they need to adapt to reduce their vulnerability to future hydro-climatic shocks, such as droughts and floods.

Regardless of the current changes in Himalayan snowfall and glacier melt, local people need to play a vital role in helping measure what is happening in the climate system. And that, as well as countries working more closely together, should help with the future planning that is desperately needed.

Siddharth Gumber receives funding from the National Environmental Research Council (NERC), The TEAMx-FLOW project (NE/Z50399X/1).

Artificial intelligence (AI) could wipe out humanity this decade, according to three researchers from the US company Anthropic. One of the employees, Jacob Coxon, has quit the AI firm.

His predictions were backed by two other current Anthropic researchers in posts on social media. On X, Coxon claimed: “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.”

It’s not the first time AI researchers and companies have issued such assessments. Here, Kate Devlin, professor of artificial intelligence & society, explains what is behind them and whether there’s anything different about the latest warning.

The latest claims about AI have come from a researcher who resigned from Anthropic. What is your general view on what has happened here?

To give this a sense of perspective, we should remember that it’s one AI company ex-employee posting on a platform known for grandiose claims, which was then quoted by a current AI company employee who put a figure on the risk of human extinction at greater than 10%.

So, it’s another doom-laden narrative about the supposed threat from AI. There is some dialling back from this in a follow up tweet from current Anthropic employee Evan Hubinger that read: “I think the risk from present models is low. What I’m worried about is superintelligence”.

These tweets have gained huge coverage because these kinds of things always do. It taps into our very primal fears – and our decades of watching dystopian sci-fi.

The claims from Jacob Coxon were backed by current employees of Anthropic. Why do you think AI companies are happy to support these kinds of claims in public?

It’s likely that these men believe what they’re saying, but we’ve heard warnings like this for years. It does seem odd that AI companies are okay with these kinds of claims.

Let’s remember that they are constantly looking for investment and, in return, they promise the most powerful and dominant models. What could be more powerful than a model that threatens our very existence?

One theory for doom-mongering by AI companies is that they want to avoid regulation. Do you think outside regulation of AI companies is inevitable?

It has been incredibly difficult to regulate AI, certainly at an international level. There has been little appetite for AI to be regulated in the US because there’s a view that regulation might stifle innovation.

The UK doesn’t have any AI-specific laws either, just existing laws that might apply in some cases. China began to introduce national-level laws but has now replaced that with technical standards and sector regulation. The EU remains the only place where there is dedicated AI regulation.

However, it looks like this has kicked off yet another call for better regulation, and we’re seeing that at present in US Congress and in the Houses of Parliament in the UK.

There is another angle, in that industry might be on board with regulation, as big tech companies would have the money to comply, but it could knock out their competitors with smaller budgets, so it would leave the leaders in dominant positions.

How far away are we from the kind of superintelligent AI mentioned in the latest warnings?

We don’t have superintelligence. I’m reluctant to put a time frame on that. Often people say “5-10 years” but that’s happened a lot in AI and 5-10 years pass without anything arising. It also depends on what is meant by superintelligence. Something that is conscious and thinks for itself? That might never happen.

In this case, it doesn’t really matter, as there are other harms already happening from AI, and there are other ways things can go wrong, though probably not at the existential level that’s being suggested here.

For example, AI is energy-hungry. It needs water to cool data centres. It has a damaging environmental impact. There are also issues around the content generated by AI, anything from a lack of plurality in the models’ outputs (meaning that the information it returns is homogenised, converging on particular viewpoints) to the spread of misinformation and disinformation.

We also know there is hidden labour in the AI supply chain: there are people working in terrible conditions to deliver the technology. And we know that it’s having an impact on people’s lives in terms of how it’s changing the workplace.

What are the potential practical ways in which AI could actually wipe out humans?

“Wiping out” is a bit extreme but there are some dangerous threats. There are the “bad actor” scenarios where AI is used destructively, like creating weapons or deadly viruses.

More mundane, but potentially very destructive, are AI agents let loose due to human carelessness or incompetence, hacking or taking control of insecure systems.

A survey we recently ran at King’s College London found that the UK public are feeling increasingly negative about AI, and that 42% of people are deliberately limiting their AI use. With claims from tech companies that it poses a direct threat to our existence, it will be unsurprising if public opinion dips even further.

Kate Devlin does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

UK gardens took a battering this summer as a heat dome settled over western Europe, and much of the green and pleasant land turned a dusty beige.

Some areas went without rain for almost two months. The “joy of gardening” turned into waiting until the cooler evening period, then laboriously trailing out from the kitchen with buckets of water to keep the patio plants alive.

But as the rains begin to return, what can gardeners do to restore their garden’s lustre? And what have we learned as we look around our parched borders and fractured lawns?

Let’s start with the good news. Nature is resilient and many plants may still be alive despite appearing bedraggled, wilted or defoliated. They just implemented strategies that help them survive a drought.

Leaf wilting and abscission is the plant’s natural response to conserving water and protecting its vital phloem (which transports sugars from the leaves to the rest of the plant) and root tissues from desiccation.

So, gardeners don’t need to take the hacksaw to the shrub border yet with the notion of making a fresh start. They can check if individual woody plants are still alive by scraping off the epidermal cells (young bark) of the shoots with a thumbnail. If the phloem tissues below are green, life prevails.

Some of the garden’s woody plants may re-leaf again this autumn – or if they are herbaceous, throw up new shoots from the soil. But many will wait until next spring before investing in a new set of leaves or stems.

Gardeners should avoid the temptation to feed plants up (add fertiliser) now, as such chemicals placed on hard soils will simply wash away with the first rains, rather than be incorporated effectively into the soil. More than anything else, gardeners need patience at this time.

Some plants will have dead stems; these can be pruned out immediately. With perennial plants, though, it is worth hesitating before removing the whole specimen, as shoots can regenerate from the stem base or remaining root system.

Similarly, brown lawns will turn green again. However, if there are residual bare patches, new growth can be encouraged by sowing grass seed mixed in with sand in mid-to-late September.

Plants employ different strategies for survival based on the climate they evolved in, and we see a legacy of this in the cultivated species we use in the garden.

Those from high rainfall areas – such as certain rhododendron, pieris, hydrangea, magnolia and sorbus – struggled this summer. Hydrangea has limited capacity to control moisture loss, whereas more drought-adapted species such as cotinus can precisely regulate their stomatal pores (where CO₂ is absorbed and moisture vapour lost).

Other plants such as forsythia and fuchsia are chance takers, still trying to grow throughout a dry period by keeping their stomata open and just running the risk of water shortage. But these types of plants can come a cropper as water columns within their xylem vessels (which transport water and minerals from the roots to the rest of the plant) snap, causing the whole stem to fail.

Anecdotally, despite the harsh dry, hot conditions, some garden plants performed well this year. Salvia, verbena, lavander and, in my garden’s clay-based soils, roses flowered prolifically, even if their shoot growth was curtailed.

Rejuvenating your garden

For those gardeners keen to rejuvenate their dreary, desolate brown patches quickly, autumn bedding plants can help provide some late season cheer. Plants such as cyclamen, chrysanthemum, aster and winter pansies can be planted now as soil moisture is restored.

After a summer like this, some precious plant specimens will be lost, but many gardens should return to their former glories – if the UK experiences a cooler summer next year.

Over the longer term, as climate change moves the UK’s southern regions to something akin to a Mediterranean climate, some gardeners will need to adapt. Mainly, this will mean planting more Mediterranean flora.

Much more challenging, however, is the climate turbulence that accrues from climate change. This compounding of stresses – such as flooding in early summer followed by drought a few weeks later, not to mention more extreme events such as fire and high wind speed – is a much more difficult phenomenon for plants.

---

Read more: A rollercoaster of weather extremes: drought and heatwaves are creating UK’s climate whiplash

---

While research on this problem is still in its infancy, early results suggest that gardeners may need to look to plant species with multiple stress tolerances. Some garden plants that are popular in Britain, like buddleia, already have good resilience.

We can also look to new genotypes from the natural world. Plants from dry river beds, for example, cope well with both drought and flooding.

Plantlife with wide natural geographical distributions and weed-like strategies (where they grow and reproduce quickly) should also be brought into breeding programmes now, to help improve climate resilience in the garden plant palette.

Ross Cameron is an advisor to the Royal Horticultural Society (RHS). Some of the research outlined in this article has been funded by the RHS and other environmental charities. Ross does not receive any personal income from these affiliations.

There is a temptation to divide the future of AI into two possibilities: utopia or catastrophe. Neither extreme is particularly helpful.

The more interesting possibility is messier – and requires a step-shift in our thinking, from artificial intelligence to AI societies.

When most people hear “AI”, they typically think of ChatGPT, Copilot or another conversational system. You ask a question, that system generates an answer.

But AI is rapidly moving beyond this. Systems can now monitor the world, make decisions, negotiate transactions and carry out tasks over extended periods of time. AI is no longer just generating an answer – it is doing something about it.

That points to something much bigger than a better chatbot: a world in which AI agents act on our behalf and, increasingly, interact with other AI agents.

An agent perceives what is happening, decides what to do, then takes actions to achieve this goal. It might book a journey, monitor a supply chain, coordinate a team, or manage a household’s finances.

Now imagine not one agent, but millions of them. Your AI agent could negotiate a mortgage with your bank’s agent, schedule surgery with a hospital’s agent, and rearrange your travel plans by dealing directly with the agents of airlines, hotels and insurers.

This future is much closer than it sounds, and this should change the questions we are asking about AI. Until now, the tendency has been to focus on how intelligent a single agent might become. The more profound challenge is what happens when millions of them interact with one another at scale.

The rise of artificial societies

The intellectual foundations of today’s AI systems were laid long before ChatGPT.

For decades, I and other researchers of multi-agent networks have studied how autonomous agents can cooperate, coordinate and negotiate when nobody has complete information and nobody controls everything.

The earliest systems that emerged focused on how the distinct AI sub-areas of reasoning, planning and acting could be combined into an effective goal-oriented agent – and how tens of these agents could communicate and cooperate to solve a common objective.

As these interactions became more complex and involved more agents, there was a shift from cooperation between agents that all belonged to a single organisation, to agents with different owners and sometimes competing aims. This focused attention on building algorithms that could form agent teams, automate negotiation and determine agent trustworthiness.

Today, the pieces needed to build large-scale multi-agent AI systems are falling into place. Modern AI agents can call software tools, access information, write and execute code, communicate with other systems and operate for extended periods.

Consider a supply chain. One AI agent could represent a manufacturer trying to secure components; another a supplier trying to maximise its revenue. Yet more could manage transport, inventory and warehouses. Each agent might be doing exactly what it is designed to do. But the important question is whether the system they create behaves sensibly.

This shift offers enormous potential benefits, but also increases the risks. In a recent experiment involving OpenAI and the tech platform Hugging Face, thousands of collaborating agents exchanged tens of thousands of messages and were able to get around the (deliberately weakened) security controls designed to contain them.

The details of one experiment matter less than the broader warning. When AI systems interact, the behaviour of the collective can be harder to predict than the behaviour of any individual system. That should make us cautious – but not cause us to down tools.

Instead, we need to shift our mindset from building intelligent machines to building intelligent societies.

Once agents can cooperate, compete and resolve conflicts with one another, we are no longer dealing with isolated machines – we are dealing with a society. Thus, the next frontier is not artificial intelligence, it is artificial societies.

An important role for humans

We already know that intelligence alone does not make a society work. Human societies depend on rules, institutions, incentives, norms and mechanisms for resolving disagreements. AI societies will need their equivalents.

Who is responsible when two agents make a bad decision? What happens when the interests of different agents conflict? Who sets the rules? And who has the power to change them? These are not just technical issues; they are questions about economics, law, politics and society.

They also point to an important role for humans. The most useful future is unlikely to be one in which AI simply replaces people. While replacement will undoubtedly happen in some cases, I believe a more common scenario will involve people and agents working together, with each doing what it does best.

Humans bring judgment, experience, values, contextual understanding and accountability. Agents bring speed, persistence, scale and the ability to process enormous amounts of information.

The goal should not be to create machines that make humans irrelevant. It should be to create systems in which humans and machines can achieve things neither can achieve alone.

But such a future requires more than just better AI models. It needs trust and transparency about what agents are doing, strong privacy protections and clear lines of accountability.

It will also require societies and governments to decide how these systems should be regulated when the most important behaviour may emerge not from one AI developer, but from interactions between systems built by many different organisations.

AI’s past decade has been defined by a race to build smarter systems. I believe the next decade will be defined by a different challenge: ensuring that millions of autonomous systems can work together safely, fairly and effectively.

The future of AI will not be determined solely by the intelligence of individual agents – it will be determined by the societies they create. And societies, as humans know all too well, are much harder to govern than individuals.

Nick Jennings has previously received funding from EPSRC. He was awarded the 2026 Research Excellence Award by the International Joint Conference on Artificial Intelligence.

Pelicans have resided in St James’s Park, London for 360 years. But the current flock have been a summer sensation due to the first hatching of chicks in the park’s history.

Pelican breeding in zoos is demanding for these birds, because the right environmental and social conditions need to come together. Rather than simply being comfortable, the birds need to encounter and respond to cues that stimulate courtship, nesting and reproduction.

This unusual nesting at St James’s Park provides an interesting example of how responding to biologically relevant stressors is a positive part of an animal’s life. Multiple factors are likely to have contributed to the pelican nesting. The heatwave, the flock’s dynamics and the compatibility of the individual birds seem to have created the right conditions for the first time in centuries.

In my recent research, I evaluated the importance of positive stress to zoo and aquarium animal welfare and found it is crucial for the wellbeing of captive animals.

Researchers call the beneficial response to a stressor, “eustress” – eu being the Greek prefix for good.

Courtship, reproduction and chick-rearing are all natural stressors. Priorities change, energetic and physical needs alter, and adults spend more time on offspring. This stress, however, is not necessarily bad. For example it can provide learning opportunities and enhance the animals’ behavioural repertoire.

As we learn more about what good animal welfare is, it is becoming clear that meaningful challenges which reflect an animal’s natural ecology help them have more autonomy over how they live. Eustress also helps animals cope better with averse situations.

Zoos do not need dramatic changes to provide these experiences. Small challenges can be built into everyday care: food can be hidden in different locations, or new scents and natural materials introduced. Animals can encounter new ground materials and surfaces (such as soil, bark and sand), changes to ropes or branches, or vegetation that naturally changes with the seasons. Enrichment can also encourage animals to explore different parts of their existing enclosure or use familiar resources in new ways.

Variation is key. If food always appears in the same place, at the same time and in the same way, there is little need for an animal to make decisions or adapt.

Many zoos already provide these opportunities. The next step is to think more deliberately about the challenge itself: is it meaningful, does the animal have a choice about engaging, and can it successfully overcome it? The aim is not to make life unnecessarily difficult but for animals encounter novelty or problems they can solve.

Rather than simply giving an animal what it needs to survive, we should provide opportunities for the animal to use their skills and make decisions.

What might this look like at your local zoo then? Meerkats (Suricata suricatta) are one example. Wild meerkats do not encounter their food neatly presented in a bowl. They need to sniff, dig and investigate their surroundings, while members of the mob keeps watch for danger.

Finding food is also a social challenge for these animals. Young meerkats learn about unfamiliar foods by observing and interacting with more experienced adults. Older meerkats help youngsters acquire important food handling skills.

Zoos can recreate elements of these challenges by scattering or hiding food, providing a sandy floor that encourages digging and creating elevated lookout points. Likewise, zoos can encourage social opportunities by providing complex feeding experiences where animals forage together. A 2021 study found that zoo meerkats retain the sentinel behaviour seen in their wild counterparts, suggesting that opportunities to perform these sort of activity remains important even when predators are absent.

The same principles apply underwater. Common clownfish (Amphiprion ocellaris) also encounter environmental and social challenges. On coral reefs, clownfish forage in a complex three-dimensional environment and depend on a host anemone for protection. They live in a strict social hierarchy. The largest fish is the dominant female, followed by the breeding male and then smaller, non-breeding fish. This social group is known as as a “queue”, as subordinate fish can move up through the hierarchy. In the aquarium, recreating this social group by housing a shoal of clownfish around the correct host anemone species allows for more choice and for fish to express a richer behavioural repertoire.

This brings us back to St James’s Park’s pelicans. Successful breeding has introduced perhaps the most natural challenge of all. Finding a partner and rearing young to maturity requires animals to change how they spend their time and respond to altered demands. When animals work for a reward, scientists call this contrafreeloading.

However, zoo animals can experience negative stress when a stressor is prolonged, overwhelming, unnatural or beyond their control. Examples can be excessive noise, incorrect climate conditions or being unable to retreat from something they find threatening.

The lesson for zoos is a good life should not necessarily be an entirely effortless one. Welfare is enhanced when animals can express behaviour their evolutionary history has equipped them to perform. Eustress is not giving every zoo animal the same kind of “challenging enrichment”. The challenge needs to make biological sense for that particular species.

Paul Rose works for WWT.

At the recent World Humanoid Robot Games 2026 in Beijing, robots sprinted, boxed, played football and tackled tasks inspired by factories, restaurants and emergency situations. The robot Tiangong Ultra even completed the 100 metres in 8.64 seconds, quicker than Usain Bolt’s 9.58-second world record.

But if we really want to know whether humanoid robots are ready for everyday life, a far less glamorous challenge may tell us more: doing the laundry. Can a robot take a crumpled T-shirt from a laundry basket, work out which way round it is, fold it neatly and recover when the sleeve slips from its hand?

No robot system can yet reliably handle the whole laundry process, from carrying a basket to the washing machine and sorting clothes for different programmes to taking them out, hanging them to dry, folding and ironing them. Even folding remains a challenge: robots need to cope with clothes of different shapes, sizes, materials and levels of creasing.

Running fast is difficult, of course. A humanoid has to coordinate many joints, keep its balance and generate powerful movements without falling over. The task itself, however, is fairly clear. The robot knows where to go, the track is predictable and the finish line does not suddenly change shape.

Laundry is a very different problem. A shirt may be inside out. One sleeve may be hidden underneath another garment. A pair of trousers may be twisted around a towel. Fabric is also soft and constantly changing shape. Pick up one corner of a T-shirt and the rest of it moves.

Humans deal with this without giving it much thought. We shake a shirt out, notice where the sleeves are, feel when the fabric slips and adjust our grip almost automatically. For a robot, every one of those little actions is a problem to solve. This is why an everyday chore can reveal more about robot intelligence than a spectacular athletic feat.

A useful household robot cannot simply replay the same movement again and again. It has to notice when reality is different from what it expected and adapt. That’s important to consider because humanoid robots are increasingly discussed as future helpers in homes and care settings, as well as factories and warehouses.

A factory production line can often be arranged so that parts arrive in predictable positions. A family home is not like that: clothes, toys and furniture move around constantly, while people, children and pets make the environment even less predictable.

Why one robot cannot simply copy another

There is another difficulty. Once one robot learns how to fold a shirt, it might seem obvious that we should simply transfer that skill to every other robot. Unfortunately, physical skills do not work like an app that can be installed on any machine.

Imagine a specialist laundry robot with rigid arms, overhead cameras and grippers designed specifically for fabric. A humanoid may have cameras in its head, five-fingered hands and different joints and balance constraints. The same movement could therefore produce a very different result.

A fixed robot arm can reach forward without worrying about falling over. A humanoid may have to move its hips, legs and feet just to keep itself stable while reaching across a table.

Sensors are important too. One robot may have sensitive touch sensors in its fingers. Another may rely mainly on cameras.

Some knowledge can still transfer. One robot can teach another something about what a shirt looks like, where its shoulders are likely to be, what a good fold looks like or how to recognise when a sleeve is trapped underneath. The receiving robot still has to work out how to carry out those ideas with its own hands, joints, cameras and body.

Research projects such as Open X-Embodiment are exploring exactly this problem by pooling experience from many different robots. The project has brought together more than one million real robot trajectories from 22 different robot types. The hope is that robots can learn useful patterns from one another rather than starting from zero every time.

What should count as progress?

The Humanoid Robot Games are valuable because they make robotics visible and show how quickly the technology is moving.

We should simply be careful about what those demonstrations prove. For robots intended to live and work alongside us, better questions might be: can the robot deal with an unfamiliar object? Can it notice that it has made a mistake and recover without human intervention? How long can it keep working before someone has to step in?

---

Read more: Could robots help tackle loneliness? BBC’s Ann Droid raises questions about the future of care

---

These measures are unlikely to produce viral videos. They are, however, much closer to the abilities that will determine whether humanoid robots become genuinely useful.

A robot that can run 100 metres faster than the human world record time is impressive. A robot that can take a messy basket of laundry, sort it, fold it and recognise when things have gone wrong may tell us far more about whether general purpose robot intelligence has really arrived.

The most important race for humanoid robots may not be against the clock. It may be against the laundry pile.

Wai-keung Fung does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.