AI for Science 🚀 2026: Models Changing Research
Discover AI for science in 2026: AI scientific research advances in physics & economics—plus AI research ethics and where human judgment still matters.
AI for Science: How Models Are Changing Research
AI Is Moving Beyond Math - and Into the Rest of Science
AI's progress in mathematics has one obvious consequence: the effects don't stay neatly inside math. A study discussed by MIT researcher Neil Thompson suggests that advanced AI models are getting better at solving theoretical problems in places like economics, probability, high-energy physics, cosmology, and quantum physics. In the reported tests, the top models went from solving about 2% of selected problems in mid-2024 to about 35% by mid-2026.
That's a real jump. Still, it's not the same thing as showing that AI can replace scientists or carry out the full work of discovery on its own. The study focused on theory problems that models can attempt without experiments or simulations. And even if an answer looks plausible, researchers have to verify the reasoning and decide whether the result actually matters.
So the practical question isn't "Can AI solve more problems?" It's "How should teams and institutions use this capability without mistaking a convincing output for a verified result?" This article breaks down what the benchmark shows, what it leaves out, and how to build AI into scientific work in a way that holds up in practice.
What the study actually tested
The study looked at whether different OpenAI models could solve theoretical problems pulled from recent research papers. The model set included o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5. Problems were drawn from probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.
The improvement is eye-catching: the strongest models went from roughly 2% correct answers in mid-2024 to around 35% by mid-2026. But that number isn't a general measure of "scientific ability." It's performance on a specific set of formal, theoretical questions. It doesn't, by itself, show that models can independently design research programs, run laboratories, or produce reliable scientific conclusions.
How the researchers tried to measure genuine problem-solving
One recurring issue with AI benchmarks is contamination. If a model has effectively "seen" a question and its solution during training, then a correct answer may come from memorization rather than reasoning. The researchers tried to reduce that risk by using problems whose solutions appeared in papers published after the models' training cutoff. Those published results were then used as a reference point for checking answers.
They also blocked internet access during testing. That matters because with browsing enabled, a model could potentially find a published solution instead of deriving one on its own. The setup was meant to test whether the models could work through the problems under controlled conditions.
Even so, it's still not a perfect stand-in for real research. A model might match a reference answer while taking an unreliable path to get there. Or it might find a valid solution that just doesn't look like the paper's exact presentation. Human review is still part of deciding whether a response is truly correct.
What the benchmark leaves out
The study centered on fields and problems where models can attempt solutions without experiments or simulations. That's a major limitation. A lot of science is about observing the world, designing measurements, wrestling with noisy data, and revising hypotheses when results don't line up. This benchmark doesn't test how well AI can do those parts.
It also doesn't measure the surrounding work that makes a theory result useful in the first place: picking a worthwhile question, understanding assumptions, connecting the result to broader research, and deciding whether it changes anything. Those aren't footnotes. They're often what separates "solving a problem" from making a contribution.
The most accurate takeaway is narrower: frontier models appear to be improving fast on certain difficult, formal tasks. That's a good reason to test them in real research workflows. It's not a reason to treat benchmark performance as proof that scientific expertise is optional.
Why theoretical science may change quickly
Math progress has already put AI's role in research on the agenda. OpenAI has announced results on difficult mathematical problems, including a reported result involving the Navier-Stokes equations. That announcement drew controversy after a mathematician accused the company of taking credit for work that overlapped with ongoing research elsewhere. Later claims around additional solutions kept the debate about AI's role in mathematics moving.
The point isn't that every claim about model performance is settled and uncontested. It's that the underlying capability seems to be advancing fast enough to change how researchers work - and how institutions decide what counts as evidence, credit, and verification.
Theory is a natural early use case
Theoretical work often fits the kind of inputs current models handle well: explicit assumptions, formal notation, and problems that can be evaluated against a proof or a known result. A model might propose a derivation, explore a line of reasoning, or help someone navigate a large space of possible approaches. Those can be genuinely useful - even if a human still has to validate the output.
But the same features also create risks. A derivation can sound right while hiding a subtle error. An answer might rely on assumptions that aren't stated clearly. In areas where a single incorrect step breaks the conclusion, "it sounds fluent" isn't the same as proof.
The transition from theory to application is not automatic
Theory can lead to practical advances, but the route is neither immediate nor guaranteed. A result in economics, physics, or statistics may later inform a method, a technology, or a policy. It might also stay a specialized result with little direct use. More AI-generated theory doesn't automatically mean faster commercial innovation.
For companies, this distinction matters. If you're evaluating AI for research, you need to ask what you expect it to improve. Is it meant to help specialists explore more hypotheses? Reduce time spent on a narrow set of derivations? Support literature review? Those goals come with different costs - and different standards for verification.
In many cases, the best way to assess value is to look at the workflow, not the model's headline capability. If AI makes it cheaper to generate candidate solutions but more expensive to verify them, the net gain might be small. If it helps experts reject weak ideas early or explore alternatives that would otherwise never get attention, the payoff could be real.
Where AI could create value - and new risks
The most immediate opportunity is likely augmentation: using models to expand the number of approaches researchers can consider, while keeping humans responsible for whether the final scientific result holds up. In a research team, an AI system could help outline a proof strategy, compare assumptions, or break a complicated question into smaller steps for experts to review.
These use cases should be treated like hypotheses to test, not like guaranteed productivity. A model's apparent speed can conceal time spent checking its work. The right evaluation looks at the whole workflow - including review, corrections, and documentation - not just how fast the model returns an answer.
Potential uses for research teams and companies
A disciplined pilot could test whether an AI assistant helps with tasks like:
- Exploring candidate approaches: Ask for multiple ways to frame a problem, then have a researcher decide what's worth pursuing.
- Checking intermediate reasoning: Use AI to surface assumptions, missing steps, or possible counterexamples, while requiring independent verification.
- Supporting technical reviews: Produce a structured summary of a derivation or paper for a qualified reviewer to check against the source.
- Reducing repetitive work: See whether the system can handle bounded tasks that use expert time but have clear acceptance criteria.
A pilot should track not just what works, but also errors, correction time, and situations where the system's confidence was misleading. If people end up spending more time verifying than they save, it's not ready for broader rollout.
The verification and accountability problem
Verification gets harder as problems become more complex. A model can claim it solved something and provide a long reasoning chain, but extra detail doesn't guarantee correctness. Researchers involved in the reported work have stressed the need for ways to verify outputs. OpenAI has also withdrawn some solutions from a dataset after publication - another reminder that claims need independent scrutiny.
There are operational and ethical stakes too. If a research group publishes an AI-assisted result, readers should understand what the system contributed and who verified the work. If a company uses AI analysis to influence product design or investment decisions, it needs a record of the assumptions and the review process. In the end, responsibility can't be pushed onto the model.
Credit is another sticking point. When AI contributes, institutions need consistent policies for authorship, disclosure, and ownership. Those rules should reflect what was actually contributed, not treat AI presence as either irrelevant or automatically disqualifying.
How research teams should use AI responsibly
If you're going to use AI in scientific work, it should start with the task - not the excitement around the technology. A reliable deployment has a defined purpose, an evaluation method that fits the job, and a named human owner. If nobody can explain how the result will be checked, the workflow isn't ready for any consequential use.
A practical adoption framework
Use four steps to evaluate a research use case:
- Choose a bounded task. Pick a repeatable activity with clear inputs and a way to judge outputs. Don't start with an open-ended goal like "use AI to accelerate discovery."
- Set a baseline. Measure how long the task takes now, how often errors happen, and what expertise is required. Without a baseline, a faster response can look like improvement even when it's not.
- Define independent review. Decide who will verify the model's work and what counts as sufficient evidence. For math, that could mean checking each step or re-deriving the solution independently. For other tasks, it might mean comparing against a trusted method or dataset.
- Measure the whole workflow. Track time saved, review effort, error severity, reproducibility, and the quality of the final work. Expand only if the benefit survives the cost of oversight.
This applies to universities, labs, and commercial R&D teams. The controls will vary by field, but the basic logic is the same: the more consequential the result, the stronger the validation and documentation should be.
Preserve the human role where it adds value
The study's researchers don't argue that experts become unnecessary. Gabriel Manso's point is that people still need to generate good ideas and connect findings. That's not a sentimental defense of expertise - it's describing the parts benchmarks often miss: knowing which question matters, understanding how a result fits into the wider field, and recognizing when an answer is technically correct but scientifically irrelevant.
Human involvement also helps catch errors that are hard to see from the answer alone. A reviewer who understands the assumptions and methods can challenge an explanation that sounds plausible but is wrong. In practice, the goal should be to push expert attention toward judgment and verification - not to remove those functions because a model sounds confident.
For organizations, that means training researchers to work critically with AI, setting clear disclosure policies, and keeping an audit trail for important outputs. The tool can contribute. The institution and its experts remain accountable for what they accept and publish.
Conclusion
AI moving into theoretical science matters because the capability is improving beyond math. The benchmark is real evidence of progress, but it doesn't prove that models can replace scientific reasoning or deliver reliable discoveries without oversight. The most practical takeaway is that teams should start evaluating specific tasks now, with verification built into the workflow.
For businesses and institutions, the choice isn't "adopt every new model" versus "reject AI." Pick a bounded workflow, measure the full cost and benefit, and keep a qualified human accountable for the result. That's how you tell whether AI is genuinely improving research - or just producing more answers that still have to be checked.
Read WIRED's report on the research for the source discussion. Before deploying AI in a scientific workflow, define what success means, how outputs will be verified, and who is responsible when the model is wrong.