Can AI Agents Make Scientific Discoveries?.
The next question for scientific AI is not whether models can assist researchers. It is whether they can own enough of the loop to produce a result the scientific community takes seriously.
In brief
AI agents can already propose hypotheses, design computational experiments, and write papers. A genuine end-to-end discovery still requires the agent to drive the hypothesis-experiment-evidence loop and produce a result that independent scientists validate.
The Prediction
Open-ended agents make a meaningful scientific discovery end-to-end (hypothesis, experiment, paper).
A research paper generated by an AI Scientist is accepted at a major ML conference or workshop.
Self-improving AI agents crush SOTA in a complex environment (AAA game, tool use, science).
What can AI research agents already do?
Google's AI co-scientist generated hypotheses and experimental protocols; Stanford's Virtual Lab coordinated an AI principal investigator and specialist agents; Sakana AI's AI Scientist-v2 ran machine-learning experiments and wrote the resulting paper.
The clearest single artifact so far is mathematical. Google DeepMind's AlphaEvolve found a way to multiply 4x4 complex-valued matrices using 48 scalar multiplications, the first improvement on Strassen's 1969 algorithm in that setting in 56 years. The result is real, machine-verifiable, and was found by an evolutionary agent. But humans chose the problem, wrote the evaluator, and decided the result mattered.
That is exactly the boundary our prediction tests. None of these systems settles the end-to-end question: whether an agent can choose a worthwhile hypothesis, decide what evidence matters, run or direct the experiment, survive a negative result, and produce a contribution that peers take seriously.
Why did we predict an end-to-end discovery?
The 2025 State of AI Report argued that AI was becoming a scientific collaborator, then made the stronger prediction: an open-ended agent would make a meaningful discovery across hypothesis, experiment, and paper.
We have made versions of this call before. In 2023 we predicted self-improving agents would crush state of the art in a complex environment, and graded ourselves a miss. In 2024 we predicted an AI-generated paper would clear a major conference or workshop, and it did. The 2025 call raises the bar again, and the wording is deliberately strict: a writing assistant is not a scientist, and an agent that optimizes a predefined benchmark has not chosen a scientific question. The call is about ownership of the loop, not fluent output at the end of it.
What would count as an AI-made discovery?
A credible hit requires the AI system to drive the work across hypothesis formation, experimental design or execution, analysis, and the paper or preprint. A human may set the broad domain and provide physical access, but cannot quietly supply the central idea.
The strongest evidence would be independent reproduction, validation by domain experts, or acceptance by a serious scientific venue with the agent's role disclosed. Workshop acceptance alone is evidence of progress, not proof of meaningful discovery. AlphaEvolve-style results sit just below the bar for the same reason: verifiable output, human-chosen question.
Why does the end-to-end test matter?
If an agent can close the loop, scientific AI stops being mainly a productivity tool. Parts of the scientific method become delegable.
What matters then is no longer literature coverage or coding speed. It is question selection, experimental access, validation, and whether a system can tell when its own result is weak. That is a much larger change than automating a lab task.
What we are watching now
We are watching agentic lab systems that connect literature review, experimental planning, analysis, and revision rather than optimizing one isolated step, particularly in biology, chemistry, and materials, where experimental access is the bottleneck.
Look for papers that credit models for validated hypotheses or experimental choices, and for evaluations that test long-horizon research instead of static scientific question answering. This prediction is graded in the State of AI Report 2026, publishing in October.
Frequently Asked Questions
Has an AI system made a scientific discovery yet?
Narrow, verifiable results exist: AlphaEvolve improved on a 56-year-old matrix multiplication record, and Sakana's AI Scientist-v2 wrote a paper that passed workshop review. No system has yet owned the full loop, from choosing the hypothesis to producing an independently validated result.
What is the difference between an AI co-scientist and an AI scientist?
A co-scientist proposes hypotheses and protocols inside a human-led loop; the human still selects the question and judges the evidence. An AI scientist would own those decisions. Every credible system today, including Google's AI co-scientist and Stanford's Virtual Lab, is the former.
When will this prediction be graded?
In the State of AI Report 2026, published in October 2026. The verdict and its justification will appear on the public predictions scorecard.