An autonomous laboratory known as A-Lab was given 58 target compounds. In 17 days it made 41 of them, mining synthesis recipes from the literature and adjusting course when a route failed. A 60-page review in The Innovation argues that this, rather than a better model, is what has changed in AI for science: the machine has moved from analysing a scientist's data to occupying the loop that turns data into knowledge.
The previous era was curve fitting at scale. AlphaFold learned the mapping from sequence to structure and solved a prediction problem that had resisted decades of effort, but it answered one question and left hypothesis, experiment, validation and interpretation to people. The review, an update of the same team's 2021 map in the same journal, says the field has crossed into closed-loop systems assembled from six modules: literature research, data analysis, hypothesis formulation, experimental execution, report generation and autonomous evolution.
They behave as a loop, not a pipeline. Language models carry domain knowledge and reasoning, agents sequence the work, embodied systems supply hands. Self-correction is built in — execution-and-evaluation pipelines, predict-measure-update cycles, failure-aware diagnosis — to catch errors before they spread. Some groups are already testing whether AI can take part in peer review itself.
The 11 fields surveyed run from information science and mathematics through physics, chemistry, materials, medicine, life sciences, astronomy and geoscience to social science, and maturity is uneven. Materials closes the loop furthest: generative models propose crystals, robotic platforms make and test them. Brain-computer interfaces offer a smaller, cleaner case, where a memristor-based decoder gained about 20% accuracy over six-hour sessions by adapting alongside the user's own signals.
The authors' most useful contribution may be a way to read such claims. They separate three levels of evidence: offline benchmark performance, domain-grounded agreement with accepted science, and prospective or experimental validation in a real setting. Only the third settles whether a result is science rather than a score.
Three constraints bind. Standardised, high-quality data stay scarce and long-tailed. Deep networks still trade interpretability for accuracy, so a model can predict a molecule's activity without explaining it. And generative models can output structures that look plausible while breaking conservation laws or valence rules — physical hallucination, the authors call it. A fourth risk is human: fluent output can be mistaken for understanding.
Scale has a price. Data centres drew about 415 TWh in 2024 and are on course to roughly double by 2030, an increase about equal to Japan's current annual electricity use; more than half of US data-centre power is still fossil-fuelled. The review counts the policy response too: the US Genesis Mission, Britain's and the EU's AI-for-science strategies and China's AI+ push, all within a year. Advantage, it argues, will go not to the largest budget but to whoever integrates data, compute and people — and such advantages compound.
The Innovation
Systematic review
Not applicable
AI for science: Progress, challenges, and perspectives
2-Nov-2026
The authors declare no competing interests.