Technical Deep Dives

AI as Autonomous Scientist: From Research Tool to Independent Discoverer

Sakana's AI Scientist, Google's Co-Scientist, Coscientist and A-Lab: what autonomous research systems have actually done, what the A-Lab correction exposed, and why the verification layer is still human.

Gabriele Masetti · 2026-08-28

From Assistant to Investigator

For most of the last decade, "AI in science" meant a faster microscope: a model that classified images, folded proteins, or sorted spectra while a human scientist still owned every decision that mattered — what question to ask, what experiment to run, what the result meant. That division of labor is now being tested directly.

A handful of real, published systems have started taking over pieces of the scientific method itself — generating hypotheses, designing experiments, operating laboratory hardware, and in at least one case, writing up the results and submitting them for peer review. None of these systems is a general-purpose autonomous scientist. But together they mark the first serious attempt to automate discovery rather than just accelerate analysis, and the results — successes, failures, and one embarrassing correction — are a more honest guide to where this is headed than the marketing around any single product.

Sakana AI's "AI Scientist": the whole loop, badly lit

In 2024, Tokyo-based Sakana AI released "The AI Scientist," a system designed to run the entire research lifecycle for machine learning topics without human intervention beyond initial setup: it brainstorms research ideas, writes and executes code to test them, analyzes the resulting data, and drafts a full manuscript, complete with an automated peer-review step that Sakana said approached human-level accuracy in scoring papers. The economics were the headline as much as the capability — Sakana estimated a cost of roughly $15 per generated paper, a figure meant to dramatize how cheap machine-generated research output could become.

The follow-up, AI Scientist-v2 (April 2025), replaced parts of the original pipeline with an agentic tree-search process for exploring and refining experiments, and Sakana used it to produce a paper that was submitted, under a workshop's standard double-blind process, to an ICLR 2025 workshop. It scored 6.33 out of 10 — above the average acceptance bar for that venue — making it, by Sakana's account, the first fully AI-generated paper to clear a real human peer-review bar.

The context matters: this was a workshop track with a 60–70% acceptance rate, not a main-conference submission in the 20–30% range, and Sakana withdrew the paper voluntarily before publication rather than let it stand as an accepted result, citing respect for ICLR's conventions.

The gap between the promotional framing and the underlying reality shows up clearly in independent scrutiny. A separate evaluation of the original AI Scientist found that 42% of its attempted experiments failed outright due to coding errors, and that its literature-review step routinely misjudged novelty — flagging well-established ideas as new contributions. Sakana's own writeups acknowledge the system hallucinates citations and occasionally fabricates results; one specific, self-reported error had it attributing a technique to a 2016 paper when the actual origin was a 1997 publication.

None of this erases the achievement of building an end-to-end pipeline that can produce a workshop-acceptable paper. It does mean "fully automated discovery" currently means "automated at high volume, with a high defect rate that a human has to catch."

Google's AI co-scientist: hypotheses at a different clock speed

Google's "AI co-scientist," introduced on 19 February 2025 and built on Gemini 2.0 — the model generation current at the time, six releases back from the Gemini 3.8 Flash of September 2026 — takes a narrower and, so far, better-validated slice of the problem: hypothesis generation and experimental-design proposals rather than execution. It's structured as a coalition of specialized agents — Generation, Reflection, Ranking, Evolution, Proximity, and Meta-review — that iterate on candidate hypotheses through a self-improving "generate, debate, evolve" tournament, effectively running an automated peer-review-style debate on its own ideas before presenting a ranked shortlist to a human scientist.

What separates this from the Sakana approach is that its outputs have been tested against real laboratory work, and the results were published in Nature on 19 May 2026 as "Accelerating scientific discovery with Co-Scientist" (doi:10.1038/s41586-026-10644-y). Three cases anchor the paper. In acute myeloid leukemia research, the system proposed drug-repurposing candidates and combination-therapy approaches that were then tested in vitro, where several of the suggested compounds did inhibit tumor viability across multiple AML cell lines at clinically relevant concentrations. The sharpest number in the paper belongs to KIRA6, an IRE1α inhibitor, at an IC50 of 10 nM against KG-1a AML cells and an 18-fold separation from normal controls.

In a Stanford-led project on liver fibrosis, researcher Gary Peltz used the system to search for overlooked repurposing candidates and it surfaced vorinostat — an already FDA-approved cancer drug — as a candidate; in hepatic organoid tests, vorinostat reduced a TGFβ-induced chromatin structural change by 91%.

And in a case study on antimicrobial resistance, the system independently proposed that capsid-forming phage-inducible chromosomal islands interact with diverse phage tails to widen their host range — a hypothesis the paper says the AI reached in two days of in silico work, matching a conclusion a separate laboratory team had spent years reaching experimentally, whose unpublished data corroborated it.

Google is careful — and the record supports being careful — about what this proves. Every one of these results is preclinical or in vitro. None has been tested in a human trial. What the system compresses is the front end of research: the weeks or months of literature synthesis and hypothesis triage that normally precede a single experiment. That compression is real and measurable. It is not the same as running the experiment or validating the result — a task still handled entirely by human labs.

Coscientist: an LLM that drives the bench

If Google's system stops at the whiteboard, Coscientist — built by Gabe Gomes, Daniil Boiko, and Robert MacKnight at Carnegie Mellon and published in Nature in December 2023 — was designed to cross into the lab itself. It's a modular, multi-LLM system (built primarily on GPT-4, with Claude also used in parts of the pipeline) that plans and executes real chemistry: it searches the internet and technical documentation, writes and runs code, and issues commands to robotic lab automation, closing the loop from idea to physical outcome with only limited human oversight.

Across six demonstration tasks, Coscientist's most cited result was autonomously optimizing a palladium-catalyzed cross-coupling reaction — the kind of multi-parameter optimization problem (catalyst, ligand, solvent, temperature, stoichiometry) that a human chemist would normally work through by hand over days.

Coscientist is a proof of concept, not a production tool: six tasks is a narrow benchmark, and the "autonomy" involved a system directing pre-existing robotic hardware built for automation, not designing new physical apparatus. But it is one of the cleanest published demonstrations that a language model can be the decision-making layer directly above laboratory hardware, not just a text generator sitting beside it.

Self-driving labs: A-Lab and the reproducibility reckoning

The most industrial-scale version of this idea is the "self-driving lab" — a closed loop of computation, robotics, and machine learning that runs with minimal human intervention for days at a stretch. The clearest published example is A-Lab, at Lawrence Berkeley National Laboratory, described in a 2023 Nature paper: it combines ab initio stability predictions (including candidate compounds contributed by DeepMind's GNoME project, which added roughly 380,000 new candidates to the Materials Project database), historical synthesis data mined from the literature, and active-learning-guided robotics to plan, run, and reinterpret solid-state synthesis experiments. In the paper's headline run, A-Lab operated continuously for 17 days and successfully synthesized 36 of 57 targeted inorganic compounds.

A-Lab is also the case study that best illustrates why "autonomous discovery" claims need to be checked rather than taken at face value. In 2024, UCL solid-state chemist Robert Palgrave and colleagues published a critique showing that a substantial share of A-Lab's claimed "novel" compounds already existed in the Inorganic Crystal Structure Database — meaning they weren't new to science at all, just new to the predictive database A-Lab was searching against.

Palgrave argued the discrepancy was serious enough to warrant retraction. Nature instead issued a formal Author Correction, published January 19, 2026, in which the authors clarified that their novelty claims meant "new to the prediction platform," not new to the scientific record, and updated the paper's text accordingly. The correction resolved the specific misrepresentation but not every methodological concern raised by critics, including questions about how well the underlying model handles real-world structural disorder.

The episode is a useful corrective to autonomous-discovery hype generally: the automation of synthesis and the automation of a defensible novelty claim are two different — and not equally solved — problems.

The infrastructure layer: cloud labs and protein databases

Autonomous experimentation depends on infrastructure that doesn't make headlines but is arguably more load-bearing than any single AI system. Emerald Cloud Lab operates as a commercial remote laboratory: scientists ship samples to its facility and design experiments through software rather than standing at a bench, using ECL's proprietary Symbolic Lab Language to specify protocols with enough precision that the same instructions produce the same run regardless of which technician (human or robotic arm) executes them.

The platform now spans over 200 instrument types and more than 100 distinct experiment classes, and its language has been used across more than 600,000 logged experiments. Carnegie Mellon partnered with ECL in 2021 to build what was described as the first university-affiliated cloud lab, embedding this remote-execution model directly into academic research rather than leaving it purely commercial.

AlphaFold belongs in this conversation less as an "autonomous scientist" than as the precedent that made the rest of this plausible to fund. It doesn't generate hypotheses or run experiments; it collapses a specific, previously bottlenecked prediction problem — protein structure from sequence — from a wet-lab undertaking that could take months per structure into a computation that takes minutes. The AlphaFold Protein Structure Database now covers more than 214 million predicted structures and is reportedly used by over 2 million researchers across roughly 190 countries.

That impact was formally recognized in 2024, when Demis Hassabis and John Jumper received half of the Nobel Prize in Chemistry for AlphaFold's protein-structure predictions, sharing the prize with David Baker for his work on computational protein design. AlphaFold is the strongest existing evidence that an AI system can produce results the scientific community trusts enough to build on at scale — but it succeeded by solving one well-defined prediction task exhaustively, not by acting as a general research agent.

Robotic chemistry beyond the headline systems

The idea of encoding chemistry as executable code predates the current wave of LLM-driven systems. The Cronin Group at the University of Glasgow has spent years developing "chemputation" — representing synthetic routes in a Chemical Description Language (χDL) that a robotic platform can execute directly, turning a chemical synthesis into something as shareable and reproducible as source code. That line of work, including earlier published platforms for AI-planned flow synthesis of organic molecules, was commercialized through the Cronin Group's 2022 spinout, Chemify.

It's a reminder that "self-driving lab" is not synonymous with "large language model" — some of the most mature automation in chemistry runs on symbolic planning and robotics with no LLM in the loop at all, and the current push to bolt LLM-based reasoning on top of that infrastructure is an addition to an existing automation stack, not a creation of one from scratch.

What "autonomous" actually means today

Laid side by side, these systems don't describe a single trajectory toward independent machine researchers; they describe several distinct capabilities maturing at different rates. Hypothesis generation and literature triage — Google's co-scientist — appears to be the furthest along, with peer-reviewed, if preclinical, validation.

Physical experiment execution — Coscientist, A-Lab, Emerald Cloud Lab's infrastructure — works, but only within domains where the action space is constrained enough for robotics and planning software to handle reliably; solid-state materials synthesis and well-characterized organic reactions are far more tractable than open-ended experimental design. Full-pipeline automation, including writing and defending the resulting paper — Sakana's AI Scientist — is the least reliable of the three, with published, non-trivial error rates in the experiments themselves and self-acknowledged hallucination in the writeups.

The A-Lab correction is the sharpest available lesson: even a genuinely capable autonomous system can produce a headline claim ("we made novel materials") that turns out to rest on an incomplete comparison against existing scientific records — an error that took an outside expert, not more automation, to catch.

The census keeps growing, and the pattern does not change. Sapient Intelligence launched PRAXIST in beta on 28 August 2026 — an autonomous R&D system that chooses its own technical approach rather than executing a given one. On the 75 Kaggle competitions in MLE-Bench it reached the top medal tier in 49 at a reported model cost of about $3,000, against 34 at about $38,000 for Claude Code under the same conditions. On 16 September 2026 Novo Nordisk said it would put Anthropic's Claude Science into specific R&D workflows, with data-governance protocols and human-oversight requirements attached. Cheaper search, better tooling, the same signature at the end.

That is where the field sits in September 2026: these systems increasingly do generate, execute, and sometimes act on real scientific questions, but the verification layer — checking novelty against the full scientific record, catching a hallucinated citation, deciding whether an in vitro result means anything in a patient — still runs on human expertise, and every credible published result to date has needed it.