Can AI Really Automate Science? What Sakana's AI Scientist Means for European Research

The system uses evolutionary computation, a method inspired by Darwinian selection. It applies small, random mutations to algorithms and retains those that improve performance. This iterative cycle can, in principle, run indefinitely, with each pass building on the last. The current version is confined to machine learning research and cannot touch a physical laboratory bench, which is a significant constraint that shapes everything else one can say about its promise.

Cong Lu, a machine learning researcher at the University of British Columbia who was involved in the project alongside UK academic collaborators, has noted that no prior system has attempted to unify the complete scientific workflow in a single architecture. That claim alone signals a genuine step change, even if the outputs remain modest in ambition.

Capability Against Limitation: An Honest Accounting

The table below summarises where AI Scientist currently sits relative to human researchers across the core phases of scientific work.

• Literature review: AI Scientist can process vast bodies of papers rapidly, but it exhibits a popularity bias, favouring heavily cited work and potentially missing frontier or contrarian findings.
• Hypothesis generation: The system identifies patterns and extrapolates plausible next steps, yet it cannot replicate the creative leaps that have defined landmark discoveries.
• Experimentation: Entirely computational; physical laboratory work is beyond its current reach.
• Data analysis: Fast and statistically rigorous, but lacking the qualitative judgement that contextualises anomalous results.
• Peer review: Systematic and consistent, but without the deep field expertise that catches subtle methodological flaws.

The quality issue is stark. Researchers and commentators who have reviewed AI Scientist outputs have concluded that most papers produced by the system would be desk-rejected by reputable journals. The work tends toward incremental optimisation rather than conceptual novelty, which limits its near-term scientific value considerably.

Professor Nello Cristianini, Professor of Artificial Intelligence at the University of Bath and one of Europe's most prominent voices on AI and scientific methodology, has argued consistently that the reductive framing of science as a data-processing pipeline misses the social, contextual, and serendipitous dimensions of how real breakthroughs occur. Informal conversations at a conference, an unexpected crossover from another discipline, a researcher's hunch born of years of tacit knowledge; none of these are capturable in a forward pass through a transformer.

The European Research Context

For EU and UK stakeholders, the arrival of AI Scientist raises immediate policy questions. The European Research Council (ERC), which funds frontier research across member states, has begun examining how AI-assisted research should be disclosed in grant applications and publications. The question of research integrity, specifically whether AI-generated hypotheses or AI-drafted papers require new disclosure norms, is already live in Brussels and in the offices of UK Research and Innovation (UKRI).

Meanwhile, Google DeepMind, headquartered in London, has demonstrated through projects such as AlphaFold and AlphaProof that combining statistical pattern recognition with structured, rule-based reasoning can produce results that genuinely extend the frontier of human knowledge. DeepMind's approach offers a more plausible template for transformative AI-assisted science than pure language model automation: it integrates symbolic reasoning rather than relying solely on next-token prediction. Sakana's system, impressive as it is architecturally, has not yet crossed that threshold.

The practical opportunity for European institutions lies in the middle ground. AI Scientist-style tools could absorb the routine analytical labour that consumes enormous proportions of a researcher's working week: systematic literature mapping, replication studies, parameter sweeps, and first-draft write-ups. That would free human scientists to focus on the creative and strategic work that actually advances knowledge. Framing AI as a research accelerator rather than a research replacement is not naivety; it is the only framing supported by the current evidence.

Looking Ahead: Robotics, Integration, and Scale

The longer-term trajectory points toward integration with robotic laboratory systems. Facilities such as the Wellcome Sanger Institute near Cambridge and the European Molecular Biology Laboratory in Heidelberg already operate highly automated wet-lab pipelines. Connecting an AI research agent to robotic experimentation infrastructure would extend automation beyond computational fields into chemistry, genomics, and materials science. That integration is a significant engineering and governance challenge, but it is not science fiction; it is a roadmap item for the current decade.

What remains genuinely uncertain is whether scaling these systems will produce qualitative jumps in scientific creativity or simply faster production of incremental results. Those are different outcomes with very different implications for how European research institutions should invest and regulate.

Researchers working in the field describe current AI Scientist-type systems as analogous to the very earliest large language models: powerful proof-of-concept demonstrations that establish feasibility without yet delivering transformative capability. The path from here to comprehensive scientific automation requires advances in reasoning, physical-world integration, and the kind of causal understanding that today's models conspicuously lack.

Key Areas Where Automation Could Reshape European Research

• High-throughput hypothesis testing across multiple variables simultaneously, reducing time-to-result in computational disciplines.
• Systematic literature mapping at a scale no human team can match, surfacing cross-disciplinary connections across the vast European research corpus.
• Continuous computational experimentation without the cognitive fatigue that degrades human judgement in long studies.
• Rapid iteration cycles in fields such as drug discovery, climate modelling, and semiconductor design, where parameter spaces are enormous.
• Democratised access to sophisticated analytical tools for smaller institutions that lack the staffing of elite research universities.

The scientific community in Europe does not need to choose between embracing and resisting this technology. It needs to define, clearly and quickly, which parts of the research process benefit from automation and which parts must remain under rigorous human oversight. That is a governance question as much as a technical one, and Europe's regulatory institutions are better placed than most to answer it thoughtfully.