The nonprofit ARC Prize Foundation on May 1, 2026, released the results of a new benchmark: a test of an AI system’s ability to solve a game. The results were striking – humans scored 100%, while the most advanced AI systems scored under 1%.

At first glance, this may be surprising to users of AI who are impressed by its polished essays, codebases and multistep projects generated in seconds. How can these brilliant AI systems struggle with these simple Tetris-shape puzzles?

That confusion points to a risk: AI is becoming integrated into everyday life faster than people can make sense of it.

We are cognitive psychologists who study how to teach difficult concepts. To recognize the limits and risks of today’s AI agent systems, it’s important for people to grasp that the systems can both accomplish superhuman feats and make mistakes few humans would. To that end, we propose a new way to think about AIs: as button-pushing explorers.

Mental models for AI

We teach college students, a group rapidly incorporating AI tools into their daily routines. That gives us regular opportunities to ask what they think is going on with AI. The answers vary widely. One student said that someone at OpenAI or Anthropic is reading and approving every response the system generates. Another, more succinctly, said, “It’s magic.”

These responses illustrate two tempting ways of making sense of AI. At one extreme, AI is treated as an inscrutable black box – a powerful but ultimately mysterious force. At another, people explain it using the same assumptions they use to understand other humans: that its outputs reflect reasoning or judgment.

The worry is that these misinterpretations don’t go away as users gain more experience interacting with AI, and they might get reinforced. When AI performs well, its output can feel like evidence of understanding or confirmation that it really is something like magic. That apparent success makes it harder to question what the system is actually doing. Biases can seem logical or inevitable; harmful behavior can look like a deliberate choice or even fate, as if it could not have gone any other way.

Saying that AI models are shaped by patterns in data, training processes and system design is true, but that’s too abstract to tell people when to trust the systems’ outputs or when they might fail. To help people avoid misplaced trust in AI, AI literacy efforts will need to include some mechanistic understanding of what produces their behavior – explanations that are perhaps not perfectly accurate but useful. Statistician George Box once wrote, “All models are wrong, but some are useful.”

Researchers have come up with several mental models for large language models. One is “stochastic parrot,” which shows that the models use statistical methods – stochastic refers to probabilities – to mimic responses with no understanding of meaning. Another is “bag of words,” which emphasizes that the models are collections of words – for example, all English words found on the internet – with a mechanism for giving you the best set of words based on your prompt.

These ways of thinking about large language models were never meant to be complete accounts of the systems. But the metaphors serve an important cognitive purpose: They push back against the idea that fluent language is necessarily caused by humanlike understanding.

But as the AI systems people use are increasingly powerful agents capable of stringing together actions on their own, it’s important for people to have a different kind of mental model: one that explains how they act. One place to find such a model is in earlier research on AI systems that learned to play Atari 2600 games. These systems didn’t understand the games the way humans do, but they still managed to rack up a lot of points.

The simple loop: Act, observe, adjust

Imagine a neural network, a relatively simple kind of AI model, placed into a video game it has never seen before. It does not “understand” the game like a human would. It has no idea whether it’s shooting space invaders or navigating an ancient pyramid. It doesn’t know the goals or rules.

Instead, it learns to play through a simple loop: Take an action – move left, jump, shoot – observe what changes, and then adjust. If an action leads to a good outcome, such as gaining points, it adjusts to become more likely to take similar actions in similar situations. If it leads to a bad outcome, such as losing a life, it adjusts in the opposite direction.

Even this simple mechanism can produce surprisingly capable behavior. Over time, by repeating this loop, the neural networks learned to play a wide range of Atari games – but not all games.

There is one game that famously stumped these early neural networks: Montezuma’s Revenge. To make progress, a player must carry out a long sequence of actions – climbing ladders, avoiding obstacles, retrieving keys – before receiving any reward at all. Unlike simpler games, most actions offer very little immediate feedback. The game required something like goal-directed, long-term planning.

Early neural networks would try a few actions, receive no reward and fail to make further progress through Montezuma’s underground pyramid. From the system’s perspective, all actions looked equally useless. But researchers made a breakthrough by changing the feedback signal. Instead of rewarding only success, they also rewarded the system for doing something new. The rewards were for visiting parts of the game it had not seen before or trying actions it had not previously taken. This tweak encouraged exploration.

With that change, performance improved dramatically. The neural network began navigating obstacles, taking multiple steps toward goals and adapting when things went wrong. From the outside, this kind of behavior can look like planning or problem-solving. But what looks like planning was not caused by sophisticated planning abilities. The underlying mechanism is still the same simple loop: act, observe, adjust.

This kind of system isn’t a stochastic parrot or a bag of words. It’s closer to a button-pushing explorer: something that doesn’t understand the world in a human sense but moves forward by pushing buttons, seeing what happens and adjusting what it does next.

From video games to modern AI agents

Today’s AI systems can do far more than play games like Montezuma’s Revenge. They can coordinate tools, write and run code, and carry out multistep projects. The range of possible actions is much larger, and the environments in which they operate are increasingly complex.

But these agents are still fundamentally button-pushing explorers. The behavior can be sophisticated, but the process that produces it is not. Humans can often infer how a new environment works after just a few observations. Systems that rely on these feedback loops cannot. They need to try many actions and see what happens before they can make progress.

This helps explain both the strengths of these AI systems and some of their most concerning failures. What these agents learn depends on what is being rewarded. And in real-world systems, those reward signals are often imperfect.

AI systems that conduct negotiations aim to maximize their client’s interests, sometimes with deceptive tactics. Rental pricing software used by landlords ends up price fixing. Marketing tools generate persuasive but misleading reviews.

These systems aren’t trying to be evil or greedy. They are adjusting to the signals they are given. From the button-pushing explorer perspective, these failures are downright predictable.

Effective AI literacy means holding two ideas at once: These systems can do surprisingly complex things, and they are not doing them the way humans do. If AI is seen as humanlike or magical, its outputs feel authoritative. But if it is understood, even imperfectly, as a button-pushing explorer shaped by feedback, people are likely to ask better questions: Why is it doing this? What shaped this behavior? What might it be missing?

That’s the difference between being impressed by AI and being able to reason about it.

Ji Y. Son receives research funding from the Gates Foundation and Valhalla Foundation.

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

The nonprofit ARC Prize Foundation on May 1, 2026, released the results of a new benchmark: a test of an AI system’s ability to solve a game. The results were striking – humans scored 100%, while the most advanced AI systems scored under 1%.

At first glance, this may be surprising to users of AI who are impressed by its polished essays, codebases and multistep projects generated in seconds. How can these brilliant AI systems struggle with these simple Tetris-shape puzzles?

That confusion points to a risk: AI is becoming integrated into everyday life faster than people can make sense of it.

We are cognitive psychologists who study how to teach difficult concepts. To recognize the limits and risks of today’s AI agent systems, it’s important for people to grasp that the systems can both accomplish superhuman feats and make mistakes few humans would. To that end, we propose a new way to think about AIs: as button-pushing explorers.

Mental models for AI

We teach college students, a group rapidly incorporating AI tools into their daily routines. That gives us regular opportunities to ask what they think is going on with AI. The answers vary widely. One student said that someone at OpenAI or Anthropic is reading and approving every response the system generates. Another, more succinctly, said, “It’s magic.”

These responses illustrate two tempting ways of making sense of AI. At one extreme, AI is treated as an inscrutable black box – a powerful but ultimately mysterious force. At another, people explain it using the same assumptions they use to understand other humans: that its outputs reflect reasoning or judgment.

The worry is that these misinterpretations don’t go away as users gain more experience interacting with AI, and they might get reinforced. When AI performs well, its output can feel like evidence of understanding or confirmation that it really is something like magic. That apparent success makes it harder to question what the system is actually doing. Biases can seem logical or inevitable; harmful behavior can look like a deliberate choice or even fate, as if it could not have gone any other way.

Saying that AI models are shaped by patterns in data, training processes and system design is true, but that’s too abstract to tell people when to trust the systems’ outputs or when they might fail. To help people avoid misplaced trust in AI, AI literacy efforts will need to include some mechanistic understanding of what produces their behavior – explanations that are perhaps not perfectly accurate but useful. Statistician George Box once wrote, “All models are wrong, but some are useful.”

Researchers have come up with several mental models for large language models. One is “stochastic parrot,” which shows that the models use statistical methods – stochastic refers to probabilities – to mimic responses with no understanding of meaning. Another is “bag of words,” which emphasizes that the models are collections of words – for example, all English words found on the internet – with a mechanism for giving you the best set of words based on your prompt.

These ways of thinking about large language models were never meant to be complete accounts of the systems. But the metaphors serve an important cognitive purpose: They push back against the idea that fluent language is necessarily caused by humanlike understanding.

But as the AI systems people use are increasingly powerful agents capable of stringing together actions on their own, it’s important for people to have a different kind of mental model: one that explains how they act. One place to find such a model is in earlier research on AI systems that learned to play Atari 2600 games. These systems didn’t understand the games the way humans do, but they still managed to rack up a lot of points.

The simple loop: Act, observe, adjust

Imagine a neural network, a relatively simple kind of AI model, placed into a video game it has never seen before. It does not “understand” the game like a human would. It has no idea whether it’s shooting space invaders or navigating an ancient pyramid. It doesn’t know the goals or rules.

Instead, it learns to play through a simple loop: Take an action – move left, jump, shoot – observe what changes, and then adjust. If an action leads to a good outcome, such as gaining points, it adjusts to become more likely to take similar actions in similar situations. If it leads to a bad outcome, such as losing a life, it adjusts in the opposite direction.

Even this simple mechanism can produce surprisingly capable behavior. Over time, by repeating this loop, the neural networks learned to play a wide range of Atari games – but not all games.

There is one game that famously stumped these early neural networks: Montezuma’s Revenge. To make progress, a player must carry out a long sequence of actions – climbing ladders, avoiding obstacles, retrieving keys – before receiving any reward at all. Unlike simpler games, most actions offer very little immediate feedback. The game required something like goal-directed, long-term planning.

Early neural networks would try a few actions, receive no reward and fail to make further progress through Montezuma’s underground pyramid. From the system’s perspective, all actions looked equally useless. But researchers made a breakthrough by changing the feedback signal. Instead of rewarding only success, they also rewarded the system for doing something new. The rewards were for visiting parts of the game it had not seen before or trying actions it had not previously taken. This tweak encouraged exploration.

With that change, performance improved dramatically. The neural network began navigating obstacles, taking multiple steps toward goals and adapting when things went wrong. From the outside, this kind of behavior can look like planning or problem-solving. But what looks like planning was not caused by sophisticated planning abilities. The underlying mechanism is still the same simple loop: act, observe, adjust.

This kind of system isn’t a stochastic parrot or a bag of words. It’s closer to a button-pushing explorer: something that doesn’t understand the world in a human sense but moves forward by pushing buttons, seeing what happens and adjusting what it does next.

From video games to modern AI agents

Today’s AI systems can do far more than play games like Montezuma’s Revenge. They can coordinate tools, write and run code, and carry out multistep projects. The range of possible actions is much larger, and the environments in which they operate are increasingly complex.

But these agents are still fundamentally button-pushing explorers. The behavior can be sophisticated, but the process that produces it is not. Humans can often infer how a new environment works after just a few observations. Systems that rely on these feedback loops cannot. They need to try many actions and see what happens before they can make progress.

This helps explain both the strengths of these AI systems and some of their most concerning failures. What these agents learn depends on what is being rewarded. And in real-world systems, those reward signals are often imperfect.

AI systems that conduct negotiations aim to maximize their client’s interests, sometimes with deceptive tactics. Rental pricing software used by landlords ends up price fixing. Marketing tools generate persuasive but misleading reviews.

These systems aren’t trying to be evil or greedy. They are adjusting to the signals they are given. From the button-pushing explorer perspective, these failures are downright predictable.

Effective AI literacy means holding two ideas at once: These systems can do surprisingly complex things, and they are not doing them the way humans do. If AI is seen as humanlike or magical, its outputs feel authoritative. But if it is understood, even imperfectly, as a button-pushing explorer shaped by feedback, people are likely to ask better questions: Why is it doing this? What shaped this behavior? What might it be missing?

That’s the difference between being impressed by AI and being able to reason about it.

Ji Y. Son receives research funding from the Gates Foundation and Valhalla Foundation.

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

Some clinical trials aren’t designed to answer scientific questions. They’re designed to market drugs. In our recently published research, my team and I analyzed over 34,000 industry-funded trials and found that hundreds of studies across seven medical fields were likely designed to promote a drug to physicians rather than to generate scientific data. For some fields, nearly 1% of clinical trials were for marketing purposes.

Known as seeding trials, these studies prioritize marketing over science while disguising their commercial purpose as legitimate research. Pharmaceutical companies use them to familiarize physicians with new products under the guise of data collection. Participants sign consent forms, believing they are contributing to medical knowledge.

In reality, patients are absorbing risks that serve corporate interests rather than resolving genuine uncertainty about the therapeutic potential of a drug.

The term seeding trial first entered the medical literature in 1994, when then-commissioner of the Food and Drug Administration David Kessler and his colleagues described such studies as attempts to entice doctors to prescribe new drugs through trials that appear to serve little scientific purpose.

Three decades later, the problem of seeding trials persists.

How seeding trials work

While the structure of a seeding trial looks similar to legitimate clinical trials on the surface, the objectives are different.

In a typical clinical trial, researchers recruit patients across clinics and hospitals to test whether a treatment is safe and effective.

In contrast, the pharmaceutical company behind a seeding trial enrolls large numbers of physicians at many sites, each seeing only a few patients. The goal is exposure: getting doctors to prescribe the drug, not generating robust data. Doctors may be selected based on their prescribing volume rather than their research credentials.

In a legitimate trial, the number of study sites reflects the number of patients needed to answer a scientific question. In a seeding trial, the number of sites reflects the number of doctors the company wants to reach.

Seeding trials often target drugs already on the market and operate as Phase 4, or postmarketing, studies. These types of studies are typically conducted after a drug has been approved to monitor its long-term safety or effectiveness. This trial stage receives less regulatory scrutiny than trials for initial drug approval, and the aims of the study may have limited relevance to actual patient care. For example, a seeding trial might measure whether patients prefer the taste of a new formulation or how quickly a drug dissolves in the stomach, rather than whether it actually improves health outcomes.

Legitimate trials also have independent oversight, with committees of scientists and ethicists who monitor the study’s progress and can halt it if patients are being harmed.

In a seeding trial, this oversight is often minimal. The sponsor of the study – typically the pharmaceutical company funding the research – maintains heavy control over the trial’s design and conduct.

Cases that exposed seeding trials

Seeding trials had attracted little public attention until litigation in the 1990s forced open the internal files of two major pharmaceutical companies, revealing that studies presented as science had been designed as marketing campaigns.

The most notorious example is Merck’s ADVANTAGE trial for the painkiller Vioxx (rofecoxib), which was first approved in 1999. The company presented the study, which ran from 1999 to 2001, as scientific research, but internal documents revealed that its primary purpose was to encourage physicians to prescribe Vioxx to their patients.

Meanwhile, Merck was accused of downplaying the significant cardiovascular risks associated with the drug. The consequences were severe: Approximately 30,000 lawsuits and nearly $5 billion in compensation followed Vioxx’s withdrawal from the market.

Parke-Davis’ STEPS trial for the painkiller Neurontin (gabapentin) – first approved in 1993 for epilepsy – followed a similar pattern of disguising marketing as research. Internal documents showed that the trial, which ran from 1996 to 1998, aimed to disseminate marketing messages through the medical literature and encourage clinicians to prescribe the drug off-label for conditions it was not approved for, such as neuropathic pain and bipolar disorder.

Unlike Vioxx, gabapentin was never withdrawn. The trial’s commercial legacy outlasted its scientific one.

These cases came to light only because litigation forced the release of internal company documents. Without that exposure, they would have remained indistinguishable from ordinary research.

How common are seeding trials?

My team and I study how pharmaceutical firms innovate and respond to regulations. To estimate the prevalence of seeding trials, we analyzed nearly 34,400 industry-funded Phase 3 and Phase 4 studies that posted results on ClinicalTrials.gov between 1998 and 2024. The trials covered seven therapeutic areas where researchers had previously documented seeding trials, including major depressive disorder, epilepsy, Type 2 diabetes and rheumatoid arthritis.

We screened these trials for criteria that prior research has identified as hallmarks of a seeded trial, such as low patient-to-site ratios and limited independent oversight.

Ultimately, we identified 204 trials – 0.59% – that had characteristics consistent with marketing-driven study design. The prevalence of these probable seeding trials in different disciplines ranged from 0.15% in osteoarthritis to 0.98% in rheumatoid arthritis.

These figures might understate the true scope of marketing-driven research. The criteria we used capture only the most identifiable cases of studies driven by marketing purposes. Definitively identifying seeding trials requires access to internal sponsor documents revealing the intent of the study, and those documents surface only through litigation or whistleblowers.

Many trials occupy an ambiguous middle ground, generating useful data while simultaneously serving promotional objectives. Without systematic surveillance, the full extent of marketing-driven studies remains unknown.

The criteria to identify seeding trials also require careful interpretation. A low patient-to-site ratio, for instance, can reflect the practical difficulties of enrolling patients in studies of drugs already on the market, such as trials testing new drug combinations or new uses for an existing treatment. These markers are best understood as signals of possible marketing intent warranting closer scrutiny, not proof of marketing intent.

Whether the prevalence of seeding trials has shifted with the expansion of transparency requirements over the past decade cannot be determined from existing registry data.

What can be done

Seeding trials may be uncommon, but they are not accidental. They reflect structural incentives in a system where the companies that fund research also stand to gain from its results. Strengthening transparency in clinical trial registration, funding disclosure and oversight would help ensure that clinical research serves patients first.

Along with other researchers, we’ve proposed reforms that cluster around two areas. The first is standardized reporting that discloses trial funding, investigator payments, enrollment criteria and the rationale for site selection. The second is independent oversight, such as committees funded through pooled industry levies, which are fees collected from pharmaceutical companies to finance independent monitoring. Random audits with publicly available results are one form of such oversight.

Some infrastructure for tracking financial relationships between industry and physicians is already in place. In the U.S., the Open Payments database allows public tracking of industry payments to physicians. But regulatory variability across countries creates openings for companies to conduct marketing-driven trials in jurisdictions with weaker oversight, particularly in low- and middle-income countries.

Clinicians can protect themselves and their patients by screening for a set of red flags before agreeing to participate in or cite a trial in their research. These include unusually low patient-to-site ratios, selecting investigators based on prescribing volume, sponsor-dominated oversight and study endpoints of limited clinical relevance. Consent forms are among the few documents patients see before enrolling, and clearer disclosure of the commercial and scientific purpose of a study is among the reforms we have called for.

For patients, clinicians and regulators alike, the question to ask of any trial is the same: Whom does it really serve?

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

Rotavirus is a highly contagious virus that spreads easily and can make babies and young children very sick. This year, doctors have been seeing more cases earlier in the season than usual.

Data from the Centers for Disease Control and Prevention shows that almost 8 in 100 people tested for rotavirus have the virus. This is only a little higher than last year at this time, when about 7 in 100 tests were positive. However, doctors are concerned because rotavirus cases started rising earlier than usual – in January – which means more children are getting sick over a longer period of time.

Often referred to as a stomach flu or stomach bug, rotavirus infection can cause extreme diarrhea, leading to severe dehydration and hospitalization. Just like measles and whooping cough, infectious diseases that are also on the rise, rotavirus can be prevented with a safe and highly effective vaccine. But vaccination rates in the U.S. have fallen since 2018.

The Conversation asked epidemiologist Annette Regan to explain why this virus is on the rise and what families can do to protect themselves from the illness.

What is rotavirus and why is it dangerous?

Rotavirus, first identified in 1973, affects the gastrointestinal system – that is, the stomach and the intestines.

Rotavirus spreads from person to person, often when germs from poop get on hands or surfaces and then into the mouth. But a person can also become infected by touching a contaminated surface and then touching their mouth, or by drinking or eating contaminated food or water.

Rotavirus causes sudden diarrhea, vomiting and fever that can cause rapid dehydration, which can lead to death if left untreated. There is no medicine to cure the virus. Doctors can only help by giving fluids and watching closely for dehydration. Babies who lose too much fluid may need care in the hospital.

Rotavirus most often affects infants and young children. Without vaccination, nearly all children have a rotavirus infection by age 5.

The virus causes most instances of hospitalization due to severe diarrhea and is the leading cause of death due to diarrhea in children under 5. Older children and adults typically experience more mild infections, but the virus can cause severe illness in people with weakened immune systems and those over 65.

A safe and effective vaccine

Safe and effective vaccines against rotavirus have been available in the U.S. since 2006.

U.S. regulators approved an early rotavirus vaccine, but it was taken off the market the next year after doctors learned that, in very rare cases, it could cause a serious bowel problem. The rotavirus vaccines used today are different. Studies in more than 70,000 babies show that these vaccines are safe and work well.

Before vaccines were introduced, rotavirus accounted for more than 400,000 medical visits, including 200,000 emergency room visits, 70,000 hospitalizations and 20-60 deaths in the U.S. each year.

Annually, vaccination prevents an estimated 40,000-50,000 hospitalizations of infants in America. Since 2006, hospitalizations due to rotavirus have dropped by 80% and emergency room visits by 57%.

Recent rotavirus surge

Rotavirus is a springtime illness in America. Cases usually increase over the winter and reach their highest point around April or May, then drop off as the weather gets warmer in the summer.

Since January 2026, doctors have been seeing more rotavirus in babies and young children than usual. According to CDC data, about 3% of rotavirus tests in January were positive, when normally only about 1% of tests are positive. That rate is now peaking at nearly 8% of tests.

Scientists have also found more rotavirus by monitoring community sewage to track how germs are spreading. The levels of virus in sewage have gone up by about 40% since February. Together, this tells doctors that rotavirus is spreading more widely and lasting longer than it usually does, which is why they are watching it closely.

Rotavirus vaccine rates in the U.S. have been declining – 77% of children received the full vaccine series by 8 months of age in 2018 compared to 74% of children in 2024. That leaves more infants susceptible to infection. Rotavirus surges are generally shorter in areas where more people are vaccinated against it, meaning they could last longer in areas with lower vaccination coverage.

In January 2026, the Department of Health and Human Services shifted rotavirus vaccination from a universal recommendation to a decision to be made by families and their health care providers. Although this change was recently paused by a U.S. judge, this has left public health officials increasingly concerned that rotavirus vaccination rates could continue to decline.

Preventing rotavirus infection

Proper hand-washing can help reduce rotavirus transmission, but because rotavirus is highly contagious, preventing the disease through vaccination is the most effective form of protection.

There are two oral, live‑attenuated rotavirus vaccines available for infants in the U.S. The first dose must be given before 15 weeks of age, and all doses must be completed by 8 months of age.

Rotavirus vaccines reduce the risk of severe disease in infants by 85% to 90%. This means fewer hospital visits, less risk of dehydration and more babies staying healthy at home.

But these benefits last only when most babies get vaccinated. When vaccination rates drop, rotavirus can spread more easily, and more infants, especially the youngest ones, can get seriously ill. Keeping vaccination rates high helps protect individual babies and keeps the whole community safer.

Annette Regan receives research and related funding from the National Institutes of Health, Pfizer Inc, Moderna, and Merck Sharp & Dohme paid to her institution. She consults for the Pan American Health Organization and is affiliated with Kaiser Permanente Southern California.