ai scientists history

"AI Scientists" sounds like a new term, but it's been used since at least the 80s.¹ It first meant scientists that work on AI. It wasn't until 2008 that the meaning reversed to be AIs that work on science. The first reference I can find is an article from Jacques Pitrat where he introduced "Chercheur Artificiel en Intelligence Artificielle". In English, an Artificial Intelligence Scientist.²

One year after Chercheur Artificiel en Intelligence Artificielle, an AI Scientist called "Adam" that does closed-loop discovery with robotics and analysis was published in the journal Science.³ Adam was focused on discovering new knowledge about yeast and developed by Prof Ross King's group at Aberystwyth University. It was actually quite ahead of its time. In 2023, Philippe Schwaller's and my research group announced ChemCrow, which can do literature search, propose novel molecules, do analysis, and control a robotic lab.⁴ On the same day, Gabe Gomes from Carnegie-Mellon University announced an "AI co-scientist" that can also write expermiental protocols for robotic labs.⁵ Neither teams characterized these as AI Scientists, but retrospectively they have been included in lists of "AI Scientists."⁶

A few months after the ChemCrow paper, I co-founded the nonprofit FutureHouse to build an AI Scientist.⁷ Sam Rodriques (cofounder of FutureHouse) and I said it would take 10 years. We based this on the idea that an AI Scientist would be something completely autonomous that controls the laboratory, writes the papers, and does all analysis. I think we'll get there before 2033, but that is still the mission we've been focused on. Our closest publicly announced result was "Robin," which was a system that can go from disease to protocol to analysis to report. ⁸ However, we still relied on humans to do the wetlab work. There has been contemporaneous in this direction that integrates better with robotics and has more complex workflows in chemistry. For example, the 2025 El Agente⁹ and Organa¹⁰ from Alán Aspuru-Guzik's group are sometimes presented as "AI Scientists" for chemistry and materials.

In 2024, Jeff Clune from the University of British Columbia and researchers from Sakana AI announced that they had built an AI scientist that can write papers by following a series of structured LLM prompts.¹¹ Their AI Scientist took these structured prompts, a dataset, and then could produce plausible looking machine-learning papers based on analysis of the given dataset. They followed-up in 2025 with "AI Scientist v2" that has a more complex workflow to not require structured prompts and produced machine learning papers that passed peer review at a machine learning conference workshop. The topic of these generated papers aren't in my domain, but my understanding is that they have not been impactful discoveries you would expect to read in a typical peer-reviewed journal or machine learning conference main track. However, the code is open source and we may still yet see impactful discoveries from the "AI Scientist v2."

In February 2025, Google announced "AI Co-scientist," which can generate novel hypotheses, research overviews, and experimental protocols.¹² The co-scientist team worked with many academic groups and they highlighted some of the best hypotheses their system generated, such as drug repurposing for leukaemia.¹² As of today, it's still not publicly available, but there are rumors that it will be available soon.

In 2025, a number of companies have recently announced AI Scientists in closed-betas such as Potato AI, AI Researcher, and K-Dense.¹³ Frontier labs are starting efforts on automate discovery in domain sciences.¹⁴ New venture-backed companies are raising capital around AI Scientists like Periodic Labs, which raised $300M to build an AI scientist that can design novel materials.¹⁵ Lila raised $435M to build an AI scientist, although they characterize their AI scientist as "scientific superintelligence" rather than an "AI Scientist."¹⁶

The AI Scientist term is catching on in academia too. You may read a recent review of the >20 academic AI scientists reported in literature.⁶ And the group that wrote the review recently launched their own AI Scientist. There are even funding calls now, like the UK Advanced Research + Invention Agency (ARIA) that announced a funding call for AI Scientists. There is early work on safeguarding AI scientists to mitigate their risk in chemical and biological contexts.¹⁷

When I first started working in the space, 2023, an AI Scientist was an aspirational concept. More science fiction. Since then, an AI Scientist seems to be any system with two ingredients: (1) it must have some artificial intelligence and (2) it automates some part of the scientific process. This feels weak; an LLM plus some structured prompts is an AI Scientist. ChatGPT is an AI Scientist under this definition.

There has been some effort to use the term "AI co-scientist" as a more precise term. Co-scientist is supposed to map to "copilot,"¹⁸ which is a broader technology term that means automating part of a process but still with human supervision. However, "co-scientist" is also the name of two specific AI Scientists and thus I don't expect to see that term takeoff. At FutureHouse, we tried calling these as "AI Science Assistants," but that term has not caught on.

defining an AI Scientist

What would be a precise definition of an "AI Scientist"?

The product of science is discoveries. Discoveries are "new knowledge about nature." ¹⁹ A genuine AI scientist should be a system whose output is a novel discovery. It should be automated. A discovery should be supported according to the norms of the scientific community: analysis, possibly physical experiments (depending on domain), and a paper putting the discovery into context and justifying its novelty.

I think the output is generally clear. The input to an AI scientist is less obvious. You could put in nothing except compute. This might work for math. In any applied domain, you must put in practical resources like hardware/equipment/compute to do some experiment. Maybe a budget is good too.

"Steering" an AI Scientist is probably a necessity too. Science is ultimately rooted in goals of society, otherwise we may end up with more digits of pi or a treatise on shrimp physiology. At FutureHouse, we internally call this a "quest" to distinguish it from a specific question or hypothesis.

To put the whole thing into words:

An AI Scientist is a system whose input is a general direction of discovery and whose output is experimental results, analysis, and a paper describing a novel discovery. A paper is a narrative document with background and context on the discovery, a description of the methods, figures, and discussion that could pass peer review in a specific domain.

The guts of an AI scientist are not so important. An AI scientist could be Bayesian experimental design. It could be a distributed virtual betting market. An AI Scientist is simply a description of the goal of a system, not a prescription of the components.

discussion

I know of no AI Scientists based on my definition. Many of the "AI Scientists" that have been announced are not generally available. And the definition I gave requires making novel discoveries prospectively, which is hard for humans to assess.

I am trying to make an AI Scientist at FutureHouse. We've gone in order. First literature search, then analysis, and then hypothesis generation. We put them together into a workflow that went from disease to drug, which was a good demonstration that we've built the pieces.²⁰ But we haven't built a general system yet.

I'm still watching the field closely. The term "AI Scientist" is growing in popularity and starting to morph from a broad research direction to marketing terminology. I hope we can get to something precise.

It may be surprising that there are so many new AI Scientists this year. Part of the reason is the rise of agents, that show how to combine the pieces of discovery. But I think it's broader. As I've written before²¹, language is the only way to connect protocols, analysis code, hypotheses, and scientific literature. It was not possible to build a generalized AI Scientist until language models reached the performance they have now.

Also, see a recent essay by Corin Wagen which has also has some nice observations on AI Scientists.

Footnotes

1. https://books.google.com/ngrams/graph?content=AI+Scientist&year_start=1980&year_end=2022&corpus=en&smoothing=3&case_insensitive=false ↩

2. Jacques Pitrat. A Step toward an Artificial Artificial Intelligence Scientist. [Research Report] LIP6. 2008. ↩

3. https://www.science.org/doi/10.1126/science.1165620; https://en.wikipedia.org/wiki/Robot_Scientist ↩

4. https://www.nature.com/articles/s42256-024-00832-8 ↩

5. https://www.nature.com/articles/s41586-023-06792-0 ↩

6. https://arxiv.org/abs/2507.23276 ↩ ↩²

7. https://www.futurehouse.org/research-announcements/announcing-futurehouse ↩

8. https://arxiv.org/abs/2505.13400 ↩

9. https://www.cell.com/matter/fulltext/S2590-2385(25)00306-6 ↩

10. https://arxiv.org/abs/2401.06949 ↩

11. https://arxiv.org/abs/2408.06292 ↩

12. https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/ ↩ ↩²

13. https://arxiv.org/abs/2508.07043 ↩

14. https://x.com/kevinweil/status/1962938974260904421 ↩

15. https://x.com/LiamFedus/status/1973055380193431965 ↩

16. https://www.nytimes.com/2025/03/10/technology/ai-science-lab-lila.html ↩

17. https://www.nature.com/articles/s41467-025-63913-1 ↩

18. https://arxiv.org/abs/2510.17826 ↩

19. Oxford Reference (Companion to the History of Modern Science). See my discussion from a few sources about defining discoveries: https://diffuse.one/p/d2-001 ↩

20. https://arxiv.org/abs/2505.13400 ↩

21. White, Andrew D. "The future of chemistry is language." Nature Reviews Chemistry 7.7 (2023): 457-458. ↩