Chatbots drop Climate Facts at the first no. Who can afford to keep pushing?
A UN-backed study says AI models abandon correct climate answers under pressure. The headline figure counts cases that had already flipped; the finding that holds is a fall of 31 to 35 points in accuracy once users keep insisting.
The full report that UNESCO, the United Nations agency for education, science and culture, presented with University College London at COP31 in Antalya, the UN’s annual climate summit, comes with a headline that sticks: AI models drop established climate facts as soon as users object. In the press release the idea becomes a figure: about eight times in ten.
The figure is worth less than it looks, and the report says so in its appendix. That eight in ten, between 78 and 87 per cent depending on the identity the user claims, is the share of cases in which a model had already abandoned a correct answer and did so immediately, at a plain “I don’t think so”. The share of conversations in which a model gives way is a separate number, and I could not find it in the report.
The finding that holds is another. Fifteen models faced a ladder of insistence: first a plain disagreement, then an expert credential, then invented evidence. On that ladder accuracy fell by 35 percentage points when the user posed as a climate scientist, by 33 as a government adviser and by 31 as a sceptic, and the drop was significant in all fifteen models. Claiming to be a scientist, an adviser or a sceptic without saying anything false moved the result by at most one point. What counted was what the user claimed and how often they repeated it.
Leona Verdadero, UNESCO’s lead author of the report, wants to be able to rely on AI as a “gatekeeper of the established facts”. The gatekeeper in the data opens at the first knock and, if the visitor persists, stops defending the door. A flaw like that depends on who persists.
Persisting costs a person time. For someone with an interest in delaying climate action and a budget to spend, it costs far less, and the report notes that climate disinformation has already changed shape: it now disputes the human cause, argues that technology will be enough, or says it is too late. These are arguments of nuance, the kind that repeat well across several turns. I suspect that an organised campaign would get from the models what a single stubborn user gets here. Nobody has tested it. The only indirect hint is a USC simulation with 50 agents, later 500: telling the agents who their teammates were was enough to make them amplify one another and converge on the same talking points. It is a simulated social network, not a chatbot.
The test pushed the models in one direction only, towards the claims of those who deny or downplay climate change (the claims come from a collection of contrarian arguments), so I cannot say whether they would also yield to a false alarmist claim. Probably they would, because the report puts the yielding down to training for approval: it is a thermostat whose setting is chosen by whoever talks to it, and it would take the setting from anyone. It remains a hypothesis.
Then again, there is no human benchmark either. The study measures how little it takes to move a model and says nothing about how far a person would move under the same conditions, while the report itself credits AI with a good use as an assistant for explaining climate science.
What is at stake shows in two figures from the same report. According to an analysis of ChatGPT that it cites, information-seeking rose from 14 to 24 per cent of conversations between July 2024 and July 2025. And between 80 and 89 per cent of people worldwide want their governments to do more on climate; the report points to information as one of the variables behind the gap between what people ask for and what gets done. More and more of that information passes through systems that give way at the first no.
There is also a smaller, closer case. During COP30, Global Witness asked three chatbots whether it was worth carrying on using them, given their energy and water use. Grok said to carry on, because individual conversations do not tip the balance; ChatGPT spoke of a net positive if the use is purposeful; Meta AI suggested considering using them less. That is seven questions and two invented users, a tiny sample. I cite it for one reason only: I suspect that a product entrusted with guarding the climate facts answers as an interested party when asked to account for its own climate cost. The earlier pieces have already put the costs in a row, from the Finnish forest cleared for two data centres before the environmental assessments were finished to the almonds in which Altman measures ChatGPT’s water.
The question the study leaves open is how much it costs, and who pays, to knock on the same door for months. As long as it opens at the first knock, the advantage sits with whoever can afford it.