Introduction

The question

A scientific finding has always needed more than one person. Someone frames a question, someone runs the experiment, someone checks the analysis, and someone outside the group reads the paper and says whether it holds. The systems this paper reads promise to do most of that work in software. Their papers describe programs that search the literature, propose hypotheses, write and execute code, draw figures and write a manuscript, and in some cases score the manuscript with an automated reviewer of their own. One of them states its aim in the largest terms available:

Another states a smaller aim that matters more for this paper, because it names the user:

Put the two aims together and a possibility appears that none of the papers names outright: a laboratory of one. A single researcher, with access to such a system and to computing, might produce the output of a group. Chris Lu and others, The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery (2024), give a cost of less than $15 per generated paper; Samuel Schmidgall and others, Agent Laboratory: Using LLM Agents as Research Assistants (2025), report an 84% decrease in research expenses against earlier autonomous methods. Whether such a laboratory is possible as a matter of engineering is the authors' claim to make. This paper asks a different question, the one Meridian asks of any decision: when may a result produced by an AI-scientist system stand as a scientific finding, who decides, by when, and what would make the decider look at it again?

The question has four parts, and each can fail on its own. A result can be accepted by someone with no standing to accept it, such as the system's own reviewer. It can be accepted with no date by which anyone must decide, so that it drifts into use before anyone has judged it. And it can be accepted with no stated condition for reopening, so that an error, once in the record, has no door out. A laboratory of one sharpens all four, because the people who in a group would hold these roles separately are now one person.

The paper reads ten public texts on this question. Seven are the papers of the systems themselves, released under an open licence: the AI Scientist and its second version, Agent Laboratory, Co-Scientist, Kosmos, Robin and Coscientist. One is a perspective on the risks of such systems. Two are United States government texts in the public domain: the executive order of November 2025 that launched the Genesis Mission, a federal programme to build AI agents for science, and the national AI action plan of July 2025. The paper takes no side on any system, result or policy. It reports where each text places the decision and where it leaves the decision open.

The contribution

The contribution is a reading of these texts against one question, and three findings come out of it. First, none of the system papers lets the system decide whether its own output is a finding. Each places that decision with people, and does so at points the paper itself names: the scientist who sets the objective and prepares the data, the person who selects which outputs go forward, the human reviewer, the venue, and the authors who sign. Second, the deadlines in the texts bind institutions, not findings: the executive order gives the Secretary of Energy days to build a platform and a year to report, but no text sets a time by which any result must be judged. Third, the conditions for looking again are written into the papers as the failures their authors found in their own systems: results invented by the system, test data that overlap training data, runs that do not repeat, and interpretations that claim more than the analysis shows. A laboratory of one gathers every human point into one person. The texts show what that person must do; they do not show that one person can do all of it.

The debate

Four lines of argument meet in the literature these texts draw on. They disagree less about what the systems can do than about where the difficulty of trusting them lies.

The first line holds that science needs such systems because discovery has become harder for people alone. Nicholas Bloom and others, Are Ideas Getting Harder to Find? (2020), show across industries that research effort is rising while research productivity falls sharply. Michael Park and others, Papers and patents are becoming less disruptive over time (2023), find that papers and patents are increasingly less disruptive across fields and decades, and link the decline to a narrowing use of prior knowledge. Neither work is about AI. This paper reads them as the background against which the system papers make their case: if ideas are harder to find, a machine that searches more of them is offered as the remedy. The executive order speaks in the same register when it describes what the federal programme will build:

The second line holds that the systems already work, within limits their makers state. Hanchen Wang and others, Scientific discovery in the age of artificial intelligence (2023), review how AI supports hypothesis generation, experiment design and data interpretation, and stress in their abstract the unresolved issues of data quality and the need to know when AI methods need improvement. The laboratory systems give this line its strongest cases. Nathan J. Szymanski and others, An autonomous laboratory for the accelerated synthesis of inorganic materials (2023), describe an autonomous laboratory that synthesised materials over weeks of continuous operation; Kyle Swanson and others, The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies (2025), describe software agents, led by an agent in the role of principal investigator, whose designs were then tested in experiments. This paper reads both as cases in which an experiment, not the software, had the last word.

The third line holds that the difficulty is not capability but verification. Joeran Beel and others, Evaluating Sakana's AI Scientist: Bold Claims, Mixed Results, and a Promising Future? (2025), an independent evaluation of the first AI Scientist, report by their abstract poor novelty assessment, many experiments failing from coding errors and some hallucinated numerical results, while still calling the system a step forward. The system papers say much the same of themselves, as later sections show. Juraj Gottweis and others, Accelerating scientific discovery with Co-Scientist (2026), put the condition in one sentence:

The sentence holds two claims that pull against each other. Democratised access is the promise of the laboratory of one; robust verification and rigorous peer review are the work of many. The sentence does not say how the two are to be reconciled, and none of the texts does.

The fourth line holds that the difficulty is safety and governance. Xiangru Tang and others, Risks of AI Scientists: Prioritizing Safeguarding Over Autonomy (2025), argue that autonomy should come second:

Their proposals are licences for users, certification for developers and institutional committees on the model of review boards for clinical research. The executive order belongs partly to this line and partly to the first: it asks for vetting of users and secure platforms, and also states its aim to dramatically accelerate scientific discovery. The action plan that preceded it states a condition under which acceleration is supposed to pay: predictions made with AI are of little use, in its words, unless scientists can also increase the scale of experimentation. The plan therefore recommends automated laboratories reachable through the cloud, and asks that researchers funded by the government disclose the datasets their AI models use. Its demand is for more experiments and more open data, not for a new judge of results. In the plan's section on evaluations, it is regulators, over time, who should explore how evaluations of AI systems fit the law they already apply; the evaluations meant there are of the systems, not of their scientific claims.

This paper does not settle the debate. It asks a question the four lines share without answering: whichever line is right, someone must decide when a given output counts. The next chapter sets out how the texts were read and then follows that decision through them.