For years, AI in science mostly meant specialized tools.
One system predicted a structure. Another classified an image. Another searched a literature database or fitted a model to experimental data.
What is changing is not simply that these tools are becoming better. Researchers are beginning to connect them into systems that participate across several stages of the scientific process.
In 2026, several peer-reviewed demonstrations made that transition unusually visible.
Google researchers introduced Co-Scientist, a multi-agent system designed to generate, criticize and refine scientific hypotheses and suggest experimental approaches. Researchers subsequently tested some of its proposals experimentally. The system is explicitly conceived as a collaborator rather than a replacement for human scientists.
Another team presented Robin, which integrates literature search and data-analysis agents. Given a scientific problem, Robin can propose hypotheses and experiments, receive experimental results produced by researchers, analyse those results and generate another round of hypotheses. The physical experiments were still performed by humans, but several intellectual stages of the research loop were linked into one system.
In computational research, the boundary has moved further. The AI Scientist was demonstrated as a pipeline capable of generating research ideas, modifying code, running computational experiments, analysing results, producing a manuscript and conducting an automated review. Its domain is far narrower than science as a whole, but the experiment demonstrates that a research cycle can sometimes be automated much further when the experimental environment itself is digital.
At the same time, self-driving laboratories are connecting algorithmic decision-making with robotics and scientific instruments. A 2026 review describes systems in which algorithms can propose, execute and interpret experiments with limited human intervention, while emphasizing major remaining problems of scalability, generality, provenance and trustworthy reasoning.
The evidence therefore supports a narrower claim than the phrase “AI scientist” might suggest.
AI is moving from isolated scientific tasks toward connected research workflows.
That distinction matters.
Generating a hypothesis is not the same as judging whether it is scientifically important. Executing an experiment is not the same as knowing whether the experiment was well conceived. Producing a manuscript is not equivalent to understanding the phenomenon described by it.
And systems that perform several of these tasks remain dependent on human-designed objectives, instruments, datasets and validation procedures.
Yet something real has changed.
The distance between human interventions can grow.
Literature search can be automated. Then hypothesis generation. Then experimental design. Then execution of some experiments. Then analysis. Then revision of the hypothesis.
None of these steps individually requires the arrival of a machine that deserves the title scientist. But linking them changes the structure of scientific work.
The question may therefore be shifting from:
Can AI perform scientific tasks?
to:
Which parts of the scientific cycle still require a human — and why?
Source & Mind assessment
Observation: Multiple peer-reviewed systems now connect several stages of scientific research that were previously treated as distinct AI tasks.
Confidence: Strongly supported.
Interpretation: This looks like an early transition from AI-as-tool toward AI-as-participant in research workflows.
Confidence in interpretation: Plausible.
Not established: General autonomous scientific competence, human-level scientific judgment or the disappearance of human supervision.
What would strengthen the signal: Independent laboratories repeatedly reproducing AI-originated discoveries; successful closed-loop physical experimentation with less human intervention; and prospective evidence that these systems discover scientifically useful results that researchers would otherwise have missed.