AI is making its way into laboratories. And science may discover more than it is able to understand

Artificial intelligence is producing original mathematical proofs, the correctness of which can be verified using formal rules, without mathematicians having to devise every step themselves. In the experimental sciences, a more persistent obstacle seemed to remain, because a hypothesis about how a cell functions only becomes established knowledge when it is tested against samples and measurements. Now even this work is beginning to be entrusted to AI, with potentially even more far-reaching consequences.

Automation has been present in laboratories for decades, and as early as 2009, the robot scientist Adam was formulating and testing hypotheses on yeast genetics (https://www.science.org/doi/10.1126/science.1165620). Traditional instruments, however, simply carry out pre-set protocols: when a result suggests changing the experiment, a scientist must interpret it, design the new test and issue instructions to the various pieces of equipment.

The paper published on 4 November in Nature by Mingyu Wu and colleagues at Shanghai Jiao Tong University shows how this divide can be bridged (https://www.nature.com/articles/s41586-026-11228-6). AutoDNA makes the functions of over twenty types of instruments accessible to AI agents, which can be combined as required. Once given an objective in everyday language, the agents prepare protocols, instruct the machines and adapt the work to the results. The breakthrough lies in making the laboratory programmable by AI, which constructs and modifies the sequence of operations whilst working, linking one measurement to the choice of the next experiment.

The authors asked the system to optimise enzymatic DNA synthesis, which involves adding one base at a time to build a sequence. AutoDNA carried out successive cycles, modifying the conditions based on the observed yields, achieving up to 97.8 per cent per step, compared with the 92.7 per cent obtained when compared with a conventional statistical algorithm; it also coordinated four protocols in parallel, with execution 3.2 times faster than the sequence. It converted a message into 78 DNA fragments, synthesised and sequenced them, and recovered the text without errors after 11,382 operations in approximately eight and a half days. Finally, starting from a prepared tumour sample, it identified a previously known mutation in the KRAS gene and designed and synthesised the reagents to verify it experimentally: the result of the first analysis determined the operations for the second.

AutoDNA did not discover any new biological facts and worked on human targets using pre-configured tools; it requested supplies and maintenance, resolving eight of the fourteen recorded faults on its own. However, it demonstrated a fundamental component of the system that is still missing: an AI capable of formulating scientific hypotheses could use such a laboratory to select the evidence that distinguishes between competing explanations, read the measurements and autonomously review its own hypotheses. A researcher would no longer be required at every stage, although maintenance and quality control would remain essential.

If these systems were to become widespread – and there are economic reasons to believe they will – their cost-effectiveness would not depend solely on a reduction in manual labour: expensive tools could be used for longer and shared amongst many teams, whilst the automatic selection of experiments would reduce wasted attempts and make it feasible to explore combinations that are currently too numerous. The initial investment would remain high, but a widely used platform could reduce the cost of each new study, further accelerating adoption.

The most significant implication, however, concerns the relationship between what we know works and what we understand. Let us imagine that a platform identifies a combination of molecules capable of halting the growth of a tumour and that this effect is confirmed by independent evidence, right through to clinical trials. We might have an effective drug without being able to reconstruct the sequence of molecular interactions on which its benefit depends. Medicine is familiar with treatments whose effectiveness has preceded an understanding of the mechanism; the novelty would lie in making this gap systematic, producing usable results more quickly than we are able to explain them.

This gap would have practical consequences. If we do not sufficiently understand why a treatment works, it becomes more difficult to predict in which patients it will fail, what interactions it will have with other medicines and what long-term effects it may produce; specific experimental tests will then be required, rather than inferences based on a well-understood mechanism. An algorithm might also find a solution by exploiting particular patterns in the samples or measurements, which cease to hold true when conditions change. The reproducibility of the effect and the possibility of transferring it to new contexts would become central issues, all the more so if subsequent experiments were selected by another automated system.

Even the very meaning of scientific knowledge would risk changing. For a mathematician, a proof generated by AI may remain difficult to understand and yet still be formally verifiable; in biology, no single universal verifier can determine on its own whether a sample was representative or an instrument was correctly calibrated. The credibility of a discovery would require access to the original data and experimental decisions, as well as the replication of the experiments using independent methods. Even if the effect were confirmed, it would be necessary to derive an explanation from it that could be tested and used to predict new phenomena. A plausible account of the path followed by the AI would not guarantee that this explanation is correct.

The work of scientists would also change; they would have fewer opportunities to understand a phenomenon through the personal design and execution of each experiment and would have to devote more attention to selecting questions, verifying procedures and interpreting results. If the volume of data were to exceed human capacity to examine it, individual articles would no longer suffice to organise knowledge: verifiable archives and independent tools for scrutinising discoveries would be required. The training of scientists would also need to preserve the critical faculties necessary to evaluate results that none of them has produced directly.

The authors speak of ‘science as a service’, accessible even to those without a laboratory; however, this model could concentrate research capabilities in the hands of a few operators, who would decide which questions to explore and who can verify the answers. Independent verification would therefore require access to raw data and the ability to replicate the experiments outside the platform that conducted them, whilst rules on the ownership of results should prevent the mechanism of a therapy from becoming knowledge reserved solely for the owner of the system that discovered it.

AutoDNA is not yet a science capable of independently discovering new biological facts, but it lends credibility to the link between AI systems that formulate hypotheses and machines that test them against the physical world. If cost-effectiveness drives its widespread adoption, we may come to know an increasing number of phenomena through their effects, whilst our understanding of the underlying mechanisms will lag behind. The challenge will then be to build a science in which the speed of machines does not deprive humans of the opportunity to verify and understand what the machines have discovered, without ending up in a recursive process in which other machines are asked to interpret for humans what the machines have discovered and tested.