For years the safe line about AI in science was that it helps researchers but does not do research. This year two papers in Nature pushed hard on that line. One AI system wrote a paper that passed peer review at a workshop. Another proposed cancer drug combinations that held up in the lab. Neither one replaces a scientist. Both change what one scientist with good tools can get done.
What happened
In March, a team from Sakana AI, the University of British Columbia, the Vector Institute and the University of Oxford published "Towards end-to-end automation of AI research" in Nature. The preprint on arXiv notes the Nature publication and describes the system, called The AI Scientist. It comes up with research ideas, checks whether they are new, writes and debugs its own experiment code, analyzes the results, drafts the full manuscript and then reviews its own work.
The headline result: one fully machine-written paper passed the first round of peer review at a workshop of a major machine learning conference. The authors note that workshop accepted about 70% of submissions, so this is a low bar compared with a top journal. UBC's announcement of the study quotes lead author Jeff Clune calling it the first time AI has gone through "the entire scientific research process on its own." The same release lists the limits: the system sometimes produced underdeveloped ideas and inaccurate citations, and it currently works only in computer science, where every experiment runs on a computer. The authors also warn it could flood already strained review systems with low-value papers.
The second paper comes from Google. Accelerating scientific discovery with Co-Scientist, whose arXiv listing notes publication in Nature in 2026, describes a group of AI agents built on Gemini that propose research hypotheses, critique each other's ideas in a tournament, and refine the winners. It does not run the experiments itself. Human scientists do. In one test, it suggested existing drugs and drug combinations for acute myeloid leukemia, a blood cancer, and those suggestions were confirmed in lab dish experiments. The authors also report that hypothesis quality kept improving as the system was given more computing time.
Why it matters for your business
You are probably not running a research lab. But these papers are a preview of how knowledge work in general is changing, and the pattern is worth noticing.
- The pipeline is the product. Neither system is one clever model. Each is a chain: generate options, check them against what is known, critique them, keep the best, write it up. That same structure works for market research, legal research, proposal writing and product ideas.
- Volume is cheap; judgment is not. AI can now produce many plausible ideas fast. The scarce resource is a person who can tell which ones are worth testing. The AI Scientist's fake citations are a reminder of why.
- Quality rises with compute. Both teams report better results when the system has more computing time or better underlying models. As prices fall, the same setup gets better without anyone rebuilding it.
The skeptic's view deserves a hearing. A workshop paper is not a breakthrough, the lab validation in the Co-Scientist work was done by human scientists, and nobody yet knows how many of these AI-generated ideas turn out to be wrong. Those are fair limits.
What we'd do
- Copy the structure for one research task. Pick something your team does repeatedly, like competitor research or reviewing bids. Have AI generate options, have it critique them against a checklist, then have a person choose.
- Never skip the human check on sources. Require every factual claim and citation in AI output to be verified before it leaves the building.
- Measure the hit rate. Track how many AI-suggested ideas survive human review. That number tells you whether the tool is saving time or creating work.
- If you are in life sciences or engineering, ask your research partners how they are using hypothesis tools like these, and what validation they require.
Our take
We think these papers matter less as proof that AI can be a scientist and more as proof that AI can run a disciplined process. That is good news for any business, because disciplined processes are where most value sits. The risk is the same one the authors flag: a flood of fluent, plausible output that nobody has time to check. The businesses that win with this will be the ones that pair a generate-and-critique machine with a person who is ruthless about what gets through.
Want a system like this in your business?
A discovery conversation is where we map it: what to automate first, where the human checkpoints go, and what it would take.
Book a discovery conversation →