AI can expose scientific errors. Or make them harder to challenge

In 2025, Donald Trump’s government commissioned five experts, all known for their critical stance towards the prevailing conclusions of climate science, to prepare a scientific report on the effects of climate change. Whilst drafting the report, one of them had what seemed to be an excellent idea: to submit the document to an artificial intelligence system, asking it to assess its scientific accuracy and check for any bias. The response was rather embarrassing. The system identified a biased selection of evidence, the omission of important studies and a presentation of the results designed to minimise the damage caused by global warming. Even more interesting was what happened next, when the authors began to use those criticisms to strengthen their defence against the objections they were bound to face.

The story was pieced together by Jessica McKenzie in the Bulletin of the Atomic Scientists, thanks to a vast amount of private correspondence made public following legal proceedings. It is a particularly instructive case for understanding how artificial intelligence can alter the way scientific research is conducted, because it directly illustrates what happens when a tool capable of identifying the weaknesses in a piece of research is used by the very people who have a vested interest in defending its conclusions.

The report had been commissioned by the U.S. Secretary of Energy, Chris Wright, and entrusted to John Christy, Judith Curry, Steven Koonin, Ross McKitrick and Roy Spencer – scholars with genuine scientific expertise – who were chosen to produce a critical review of the current understanding of climate change. Published by the Department of Energy in July 2025, the document argued that many of the negative consequences of rising CO₂ levels had been exaggerated, whilst the potential benefits had been underestimated. The political stakes were high, as these conclusions could help justify the scrapping of federal regulations on greenhouse gas emissions.

A lawsuit brought by the Environmental Defence Fund and the Union of Concerned Scientists led to the discovery that the group had been formed in breach of federal regulations on transparency and the balance of advisory committees. In January 2026, a judge recognised the breach and ordered the documents to be handed over, allowing the publication of over one hundred thousand pages of material, including communications between the authors.

It emerges from these that Steven Koonin, a physicist and former chief scientist at BP, had begun experimenting with Microsoft Copilot as an assistant whilst preparing the report. After using it to summarise and rephrase certain passages, he asked it for a genuine critical assessment. Copilot noted that the publications cited, whilst often authentic and scientifically sound, did not adequately represent the full body of available knowledge. For example, the beneficial effects of CO₂ on plant growth were discussed whilst overlooking the constraints imposed by the availability of water and nutrients. The system also flagged a disproportionate reliance on publications by the report’s authors themselves and other researchers with similar viewpoints, accompanied by language that betrayed a polemical slant.

Copilot’s conclusion was explicit: the evidence was selected to support a favourable portrayal of the effects of CO₂. This is one of the most insidious distortions in scientific communication, as it allows a misleading conclusion to be drawn using genuine results. All that is required is to systematically exclude publications that contradict the desired thesis, whilst retaining those that are compatible. To uncover the problem, one must also be aware of the studies the author chose not to cite – a task in which artificial intelligence can be particularly useful.

The researchers’ initial reaction was enthusiastic. Judith Curry wrote that the result was surprising, even better than what might have been achieved by real scientists. Koonin said he was equally impressed. The two therefore recognised the value of the tool, but what emerges most clearly from the correspondence is the attempt to identify expressions that might have made the report appear prejudicially hostile to mainstream climatology. Koonin also asked Copilot to write an article opposing the document, in order to anticipate the critics’ arguments and incorporate pre-emptive wording into the final version.

This passage is crucial. The authors had received a complaint regarding their choice of scientific evidence, but the available documentation shows that they took particular care to make the text less open to criticism. This does not prove that they ignored all substantive comments, nor that they entrusted the drafting of the report to artificial intelligence – a claim they have explicitly denied. It does, however, show how easily a critical review can turn into an exercise in preparing for controversy, in which the primary aim is to anticipate and neutralise objections.

Copilot’s criticisms, of course, would have carried little weight had they remained mere judgements by a machine. Their validity must be established through independent verification of the claims. In September 2025, over 85 experts, coordinated by Andrew Dessler and Robert Kopp, published a comprehensive review of the government report, documenting numerous examples of selective citations and misrepresentations of the available evidence (Climate Experts’ Review of the DOE Climate Working Group Report, 2025). Subsequently, Benjamin Santer and four other climatologists demonstrated in Earth’s Future how the report had misrepresented key observations relating to stratospheric temperatures, using them to cast doubt on one of the pieces of evidence for human influence on the climate. They also examined a graph subsequently presented in public by Koonin, which contained erroneous observational data; the report’s authors later attributed this to a programming error (Misrepresentation of Human Climate Fingerprint in DOE Report, 2026).

The fact that artificial intelligence had recognised some of the inaccuracies in the document does not mean that its assessments were always correct. In one passage, for example, Copilot accepted the argument that it would be incorrect to speak of ocean acidification, given that seawater remains alkaline. This is a basic chemical error, because acidification means a decrease in pH, regardless of whether the final value remains above 7. The absorption of CO₂ causes precisely this decrease, as shown by observations collected by the National Oceanic and Atmospheric Administration. This incident demonstrates how essential it is to verify the models’ responses, even when their other observations prove to be well-founded.

Subsequent communications between the authors then reveal a growing impatience with the tendency of language models to favour the prevailing scientific literature. John Christy observed that a system trained on a vast amount of mediocre material would inevitably place too much value on that material. Roy Spencer suggested pressing the model with questions, pointing out that, with sufficient persistence, it is sometimes even possible to make a model change its mind. This is a valid observation regarding how these systems work, as they are sensitive to the wording of queries; however, it becomes problematic when the aim is to elicit a judgement more favourable to one’s own conclusions.

This brings us to the general significance of the matter. Artificial intelligence introduces an extraordinary possibility into scientific output: the ability to rapidly subject a piece of work to an extensive critical review, capable of comparing its claims with a vast body of literature and identifying possible omissions. The quality of this analysis naturally depends on effective access to sources and the system’s ability to interpret them correctly. However, once the document’s weaknesses have been identified, the same tool can be used to construct persuasive responses to objections without addressing the underlying causes that gave rise to them.

The consequences are likely to be very serious. Even today, we can produce texts that are scientifically flawed but extremely difficult to challenge on the merits, because any foreseeable criticism can be anticipated and met with a seemingly competent response. As the models improve, this possibility will grow even further. Scientific peer review thus risks becoming a contest between automated tools used to identify errors and equally powerful tools used to defend them, imposing an ever-increasing workload on human reviewers.

For this reason, artificial intelligence will only truly improve scientific integrity if its critical use is accompanied by procedures that ensure objections can be documented and responses verified. A well-founded objection should leave a verifiable trail, together with the sources supporting it and any amendments made to the work. It would thus be possible to distinguish a genuine scientific correction from a mere rewording that makes the problem less obvious. Furthermore, independent peer review must have access to this process, because leaving the management of criticism entirely to the authors means allowing them to select which criticisms they intend to take seriously.