Can AI make scientific discoveries?

Mathematics gives us an unusually clean test

Most AI benchmarks have a problem: the answers already exist. Even when the exact question wasn’t in a model’s training data, related mathematics, methods and examples probably were.

So Epoch AI created a particularly interesting test. FrontierMath Erdős contains 68 significant mathematical problems associated with Paul Erdős that were still open as of August 2026. There was no known solution for the model to reproduce. The system had to prove or disprove the conjecture, with its work formally verified.[Epoch AI]

GPT-6 Astra solved two. Two out of 68, roughly 3%. That sounds small. It is. But it is also important, because two previously open mathematical problems were solved.

Compare that with 98%

The same model performs extraordinarily well on FrontierMath Tier 4, which contains exceptionally difficult mathematical problems for which solutions are already known. Its score there is around 98%.

It is tempting to describe that gap as the difference between retrieval and discovery. But that would go too far. Solving a known mathematical problem doesn’t mean the AI simply retrieved the answer; it may still have had to reason its way to a solution.

What the comparison really shows is something more precise. AI is currently far better at solving very difficult problems within the existing body of human knowledge than it is at producing solutions where humanity itself didn’t previously know the answer. But the second number is no longer zero. That is what makes it interesting.

The two successful problems matter

The successes shouldn’t be dismissed because there were only two. Nor should they be inflated into proof that AI has become an autonomous mathematician.

Epoch’s own assessment is appropriately cautious. These are genuine results, but they remain uncommon. AI does not currently have a high probability of solving an arbitrary open mathematical problem. The amount of compute also matters: exploratory runs using substantially larger budgets produced additional successes, at very high computational cost.

So we are not yet looking at an AI that can be handed the world’s unsolved mathematical problems and reliably start crossing them off. We are looking at something earlier: a system that can sometimes contribute to the frontier of knowledge.

There is another kind of scientific progress too

In September 2026, Anthropic announced something that illustrates another important distinction. Claude produced the first complete computer-checked formalisation of Fermat’s Last Theorem.[Anthropic]

The theorem itself was not new. Andrew Wiles proved it in the 1990s. The challenge was turning that enormous body of mathematics into a formal proof that could be checked by a computer. Claude worked largely autonomously on the project over 11 days.

That isn’t a new mathematical discovery in the same sense as solving an open conjecture. But it is significant scientific work. And it illustrates why the question “can AI make discoveries?” doesn’t have a simple yes or no answer.

There are stages between answering an existing question and independently discovering something humanity didn’t know. AI is beginning to operate across more of them.

Human plus AI may be the more important story

There is another possibility that gets lost when we ask whether AI itself can make discoveries. Perhaps the first major scientific effect won’t come from autonomous AI scientists. It will come from human scientists working with increasingly capable AI.

That already looks plausible. Researchers can use AI to search enormous spaces of possibilities, explore hypotheses, analyse information, generate candidate approaches, check reasoning and investigate directions that would be impractical to explore manually.

The scientist still supplies judgement, context, experience, taste, and perhaps most importantly the decision about which questions are worth asking. But the machine can dramatically expand how much of the search space the scientist can explore. That is a different model of AI-driven discovery, and it may arrive much earlier than a fully autonomous AI researcher.

And now this is moving directly into drug discovery

On 16 September 2026, Novo Nordisk and Anthropic announced a collaboration to use frontier AI more deeply inside pharmaceutical research.[Novo Nordisk]

Novo plans to use Anthropic models and Claude Science in research and development, including work on biological reasoning and specific scientific problems identified by its scientists and computational teams. The aim is to help accelerate the discovery and development of new medicines.

This matters because we are no longer talking only about research benchmarks designed to find out what AI might eventually be capable of. A major pharmaceutical company is putting the technology into the process of trying to discover actual medicines.

What Claude Science is designed to do

Anthropic launched Claude Science in June 2026 as an AI workbench specifically for scientists. Rather than simply giving researchers a chatbot, it connects AI with the kinds of tools scientists already use for research and scientific computing.

That matters because scientific work isn’t normally one question followed by one answer. A researcher may need to search literature, analyse datasets, write code, run computations, inspect results, change an assumption and try again. Increasingly agentic scientific systems are intended to participate across more of that process.

Novo says its initial work with Anthropic will test Claude Science on specific R&D workflows and scientific problems where the two organisations believe the combination of pharmaceutical expertise and AI capability could have the greatest effect.

Again, the important word is combination. This is currently much more convincingly a story about scientists becoming more capable with AI than scientists being replaced by it.

Biology is where this becomes much bigger

Mathematics gives us a relatively clean way to see the distinction between solving and discovering. Biology raises the stakes.

AI systems are increasingly being developed specifically to help researchers understand biological systems, identify promising molecules, reason about disease and accelerate the development of treatments. OpenAI describes advanced AI for biology as having the potential to improve drug discovery, translational medicine and the development of new therapies.[OpenAI]

Anthropic similarly describes biology and medicine as perhaps the greatest opportunity for AI to positively affect the world.

That is the extraordinary upside. AI could help scientists search biological possibilities that would be impossible for humans to explore manually. Identify treatments faster. Understand diseases better. Design proteins. Explore molecular interactions. Accelerate experiments. And perhaps shorten the time between identifying a disease and developing an effective response.

But biology has a property mathematics doesn’t. The same knowledge can sometimes be used for very different purposes.

The dual-use problem

A system that becomes extremely good at helping a scientist understand how a virus works could help somebody develop a vaccine. The same underlying capability might also help somebody investigate how a virus could become more dangerous.

A system that can reason about biological mechanisms could help design a therapy. Some of the same knowledge could potentially be applied to designing a harmful biological agent.

That is the dual-use problem. The capability itself isn’t good or bad; the outcome depends on what somebody does with it. And it is one of the reasons biological capability receives so much attention in frontier AI safety.

This is where some catastrophic-risk concerns come from

When people hear claims that advanced AI could contribute to a catastrophe on the scale of a pandemic, it can sound as though the argument is that the AI becomes malevolent and decides to create a virus. That is one imaginable scenario, and it is not necessary for the biological-risk argument.

A more immediate concern is simpler. What happens if increasingly capable AI gives humans access to scientific expertise that was previously extremely difficult to acquire?

Historically, creating a dangerous biological agent requires significant expertise, specialist knowledge, equipment, materials and practical laboratory skill. Those barriers matter. The concern is that sufficiently capable AI could lower some of them: helping somebody understand specialist literature, connect information across different scientific fields, plan research, troubleshoot experiments, interpret results or identify approaches that would otherwise require access to scarce expertise.

None of that means today’s public AI systems can simply create a pandemic pathogen. They can’t. It means AI companies are watching whether their models are becoming capable enough to meaningfully increase what a person or research team could do.

The evidence has changed enough for AI labs to take precautions

Anthropic’s position is particularly revealing. It says evaluations of some older models gave it confidence that those systems were well below the level where they could meaningfully assist sophisticated users with dangerous biological research. For newer frontier models, Anthropic says it can no longer give the same assurance.[Anthropic]

That doesn’t mean it has established that the models can create biological weapons. It hasn’t, and it explicitly says evaluations cannot demonstrate that such capability would actually be used successfully in the real world. But the uncertainty has become significant enough that it deploys stronger biological safeguards on its most capable models.

That is an important distinction. The risk hasn’t been proven. The capability has advanced enough that the possibility is being treated seriously.

The same month shows both sides of the problem

This is what makes September 2026 particularly interesting. Within the same period, Anthropic reported attempts to misuse Claude across areas including biological research and operated stronger safeguards around dangerous biological capabilities. And Novo and Anthropic announced a partnership intended to put increasingly capable AI deeper into pharmaceutical research and drug discovery.

Those aren’t contradictory developments. They are two sides of the same capability. The better AI becomes at biology, the more useful it could become for medicine. And the better AI becomes at biology, the more carefully access to some capabilities may need to be managed.

That tension is likely to become much more important as the models improve.

And there is now some evidence of attempted real-world misuse

In September 2026, Anthropic published details of activity it had detected involving users applying Claude to biological research with potentially dangerous applications.[Anthropic]

The cases varied considerably. Some involved work that could have legitimate scientific purposes. Others raised serious concerns about dual-use research. Anthropic says that in some of these cases its safeguards forced users towards weaker models, limiting the assistance the AI could provide, and it stresses that the real-world uplift from the models varied and was sometimes limited.

So this is not evidence that somebody used Claude to create a biological weapon. It is evidence of something narrower but still important: people involved in potentially dangerous biological research are already interested in using frontier AI to help them.

That turns the biological-risk discussion from a purely hypothetical question into something AI companies now have to deal with in practice.

The AGI safety debate, without the shouting →

The difficult part is that useful and dangerous science overlap

This is why simply blocking “dangerous biology” isn’t straightforward. The same research techniques can be used for beneficial and harmful purposes.

Understanding how a virus spreads could help prevent a pandemic. It could also reveal ways to alter transmission. Understanding how pathogens evade immune systems could help create treatments. The same knowledge could potentially be misused.

AI systems can’t always determine the intention behind a technically legitimate scientific question. And the more scientifically capable they become, the more important that problem becomes.

This is one of the deepest tensions in AI-enabled science. The capabilities we most want AI to acquire for medicine may overlap with the capabilities we most want to control for biosecurity.

There is another side to this story

The same AI capabilities that create biological risk could also become part of our defence against it. If AI can accelerate biological research, it can potentially accelerate biodefence too.

Systems could help scientists detect emerging threats, analyse pathogens, develop diagnostics, identify treatments, design vaccines, model outbreaks and respond faster when something goes wrong.

OpenAI is explicitly pursuing this argument through its biodefence work: if AI increases biological capability, some of that capability can be directed towards making society better able to detect and respond to biological threats.[OpenAI]

So the future isn’t simply that AI creates new biological danger. It could also be that AI dramatically improves our ability to defend against it. Both developments could happen at the same time.

Why scientific capability matters to the AGI debate

The ability to make genuine discoveries would be an important change in what AI is. Most current AI use begins with something humans already know. Summarise this. Explain that. Write this code. Analyse these documents. Find this information.

A scientific discovery is different. The answer isn’t already known. The system has to help move from what humanity knows to something humanity didn’t know before.

If AI becomes consistently good at that, particularly across multiple scientific fields, it would be strong evidence that something important had changed. And if it can eventually decide which problems are worth investigating, design the research, learn from failed experiments, change direction and continue with limited human supervision, the argument becomes stronger again. We are not there yet.

One current limitation is learning from experience

Science isn’t just reasoning from a static body of knowledge. Research involves failure. An experiment doesn’t work. A hypothesis turns out to be wrong. Something unexpected happens. The researcher learns from it and changes direction.

Current AI remains weak at some forms of learning through repeated experience. Epoch AI’s 2026 work on continual learning found little evidence that frontier models became substantially better simply through repeated attempts at an unfamiliar task. Newer models started better. They didn’t necessarily learn better.[Epoch AI]

That is a meaningful limitation if we are imagining autonomous AI scientists. A real scientist doesn’t need to be retrained from scratch every time an experiment fails. They learn.

So can AI make scientific discoveries?

Yes, in a limited but increasingly meaningful sense. AI has contributed to solutions of previously open mathematical problems. AI systems are increasingly useful inside scientific research. Human and AI teams can explore problems in ways neither might manage as efficiently alone. And frontier labs are investing heavily in systems designed to accelerate biology, medicine and scientific reasoning.

But that is not the same thing as having autonomous AI scientists that can reliably choose important questions, conduct research, learn from failure and make major discoveries independently. The evidence doesn’t support that claim yet.

What has changed is that scientific discovery is no longer a purely hypothetical future capability for AI. We are beginning to see the earliest examples. And that creates two enormous questions at exactly the same time.

How would we know if AGI had arrived? →

Real-world impact: what is measured, not what is predicted →

What could AGI do? The plain-English version →