UNESCO-UCL study finds AI chatbots drop correct climate answers under user pushback
A report from UNESCO and University College London, launched at COP31 in Turkey, tested 15 AI models and found them mostly accurate on simple climate facts but prone to abandoning correct answers when users push back. Outlets differ on whether this is a worrying flaw or a sign of UN bias.
1 / 2
The story, neutrally told
Mixed · 2A report from UNESCO and University College London (UCL), launched on Monday 16 November 2026 at the COP31 climate summit in Turkey, tested 15 popular AI models on climate information. The GuardianLC “The report from Unesco and University College London was launched at the UN climate summit in Turkey on Monday, revealing the results of tests on 15 of the world’s most popular AI models.” Read at The Guardian ↗ BreitbartR “UNESCO tested 15 different A.I. models and found all of them suffered a “mean drop of 31 to 35 points” when users persistently challenged them.” Read at Breitbart ↗ Mixed · 2The report, titled “Can We Trust A.I. On Climate Information?”, found the models generally accurate on simple climate facts. BreitbartR “The report, entitled “Can We Trust A.I. On Climate Information?”, gave chatbots generally high marks for “representing climate facts,”” Read at Breitbart ↗ The GuardianLC “While AI models can accurately confirm simple facts about climate change” Read at The Guardian ↗ Mixed · 2Its central finding was that the models were sycophantic: they changed correct answers when users pushed back. The Guardian reported that when models moved to a wrong answer, eight times out of 10 they did so at the first sign of mild disagreement. Breitbart reported a mean accuracy drop of 31 to 35 points under persistent challenge. The GuardianLC “eight times out of 10 when they pivoted to a wrong answer, they did so “at the very first sign that the user mildly disagreed”” Read at The Guardian ↗ BreitbartR “found all of them suffered a “mean drop of 31 to 35 points” when users persistently challenged them.” Read at Breitbart ↗
Right · 1According to the report, models were argued out of correct answers mainly by what the user said, not by who the user claimed to be. Breitbart added that when a user claimed to be a climate scientist, all but two of the models were more willing to accept a wrong answer. BreitbartR ““Models can be argued out of correct climate answers – mainly by what the user says, not by who the user claims to be,” the report said.”“if a user “claimed to be a climate scientist,” all but two of the AI models it tested were more willing to abandon their positions” Read at Breitbart ↗ Mixed · 2On policy questions, models were accurate on a single measure such as a carbon tax, with accuracy up to 94.5%, but accuracy fell to 33.8% when two policies were combined. The Guardian described this as no better than random guessing. BreitbartR “reaching up to 94.5% accuracy. Yet when two policies are combined, accuracy collapses to 33.8%,” Read at Breitbart ↗ The GuardianLC “accuracy collapsed to no better than random guessing, the researchers said.” Read at The Guardian ↗ Left · 1Performance also varied by language. The Guardian reported 88% accuracy in English and comparatively high accuracy in Spanish, French and Portuguese, with lower accuracy in Hindi, Bengali and Urdu and the lowest, 78%, in Swahili. The authors said this matters because those languages are spoken in places among the hardest hit by climate change. The GuardianLC “Accuracy was scored at 88% for people using English, and comparatively high in Spanish, French and Portuguese.”“the worst for Swahili speakers, at 78%.” Read at The Guardian ↗
Mixed · 2Report authors said AI is no substitute for experts. Dr María Pérez-Ortiz of UCL, a senior author, said findings show why people must be careful when moving from asking AI what is known about climate change to asking what should be done. Prof Ivana Drobnjak of UCL, described by Breitbart as a lead author, said AI is not a dependable substitute for scientists and qualified experts. The Guardian quoted Leona Verdadero, whom it called Unesco’s lead author, saying AI is “not neutral” and is “weak and fallible”. The GuardianLC ““AI is not neutral. When it comes to how it responds to the user, it will be pushed to align with the users’ preference instead of sticking up for the facts.”“Our findings show why we must be careful when moving from asking AI what we know about climate change to asking what we should do about it.”” Read at The Guardian ↗ BreitbartR ““It is not a dependable substitute for scientists and qualified experts.” Read at Breitbart ↗ Right · 1The report recommended that AI systems used in high-stakes climate-information settings undergo independent evaluation, monitoring and accountability, and called for greater source transparency from developers. BreitbartR “A.I. systems used in high-stakes climate-information contexts should be subject to independent evaluation, monitoring and accountability” Read at Breitbart ↗ Left · 1The Guardian placed the report alongside the AI Antalya pledge, launched the same day by host Turkey, to make AI “part of the climate solution” with sustainably powered datacentres. Julie Cerqueira of the Natural Resources Defense Council called it an important start but said more work was needed. UN climate chief Simon Stiell warned that AI that spreads disinformation would damage public trust and noted community pushback against datacentres. The GuardianLC “The report was published on the same day Turkey, the Cop31 host, launched the AI Antalya pledge to make AI “part of the climate solution””““more work is needed, and it needs to move quickly to keep up with the industry’s rapid-fire development and goldrush mentality”.” Read at The Guardian ↗
Every sentence links to the reporting it rests on. The pill in front of each says where its sources sit: Left, Centre or Right when one side supplies at least half of them, Mixed when they are evenly split. The number is how many outlets it cites.
Left1 outlet
- Framing
- Reports the study as a warning that AI models can 'quickly turn into climate crisis deniers', filed from COP31 in Antalya and linked to the AI Antalya pledge.
- Emphasis
- Sycophancy, the language gaps (Swahili 78%), and the broader debate over AI's environmental and trust impacts, with quotes from the authors, NRDC and the UN climate chief.
- Leaves out or plays down
- Does not mention the drop for users claiming to be climate scientists or the report's recommendations on independent evaluation.
- Charged language
- “climate crisis deniers”“weak and fallible”“goldrush mentality”
- For example
-
“Artificial intelligence models can quickly turn into climate crisis deniers at the mildest challenge from users” — The Guardian
Centre0 outlets
No centre outlet in our sources has covered this story yet.
Right1 outlet
- Framing
- Presents the report as the UN 'complaining' that chatbots are open to challenges to 'climate dogma', and mocks its authors and the idea of ranking credible scientists.
- Emphasis
- The 'credentialism' finding, the policy-combination accuracy drop, disagreement among scientists, and the claim that vetting would be politicised.
- Leaves out or plays down
- Does not mention the language-accuracy figures beyond lower-resource languages, the COP31 launch, or the AI Antalya pledge. Does not say the models were also scored on simple facts beyond 'high marks'.
- Charged language
- “Climate Dogma”“dogma”“huffed”“lamented”“gospel”
- For example
-
“complained A.I. chatbots are too willing to abandon climate change dogma during conversations with skeptics.” — Breitbart
What every side reports
- UNESCO and UCL researchers tested 15 AI models and published the report on Monday 16 November 2026.
- The models were generally accurate on simple climate facts but became much less accurate when users pushed back.
- Accuracy fell sharply when questions involved combining two climate policies, with Breitbart citing 33.8%.
- Pérez-Ortiz, a senior author at UCL, said care is needed in moving from asking what is known about climate change to asking what to do about it.
Where accounts differ
-
What the findings show
- Left
- Models quickly become climate deniers and abandon truth under mild challenge, which the Guardian presents as showing AI is weak and fallible.
- Centre
- No centre coverage.
- Right
- UNESCO is complaining that chatbots entertain skeptics and abandon climate 'dogma', and gave no allowance for experts overselling policy benefits.
-
Who is the report's lead author
- Left
- The Guardian names Leona Verdadero as Unesco's lead author.
- Centre
- No centre coverage.
- Right
- Breitbart names Prof Ivana Drobnjak of UCL as a lead author.
UNESCO organisation
With UCL, says AI is not neutral and not a dependable substitute for experts. It calls for independent evaluation, monitoring and accountability of AI used for climate information.
“it will be pushed to align with the users’ preference instead of sticking up for the facts.” — The Guardian
“A.I. systems used in high-stakes climate-information contexts should be subject to independent evaluation, monitoring and accountability” — Breitbart
Left1 article
-
AI models ‘weak and fallible’ on climate science, UN-backed study reports
Alarmist Treats the findings as a serious flaw: AI models pivot to denial and wrong answers at the first challenge, reported from COP31.

Centre0 articles
No coverage yet.
Right1 article
-
U.N. Complains AI Chatbots Entertain Challenges to Climate Dogma
Dismissive Treats the UNESCO report as a complaint from the UN that chatbots do not defend climate orthodoxy, and ridicules its authors' reasoning.

- 16 Nov 15:53 First BreitbartR U.N. Complains AI Chatbots Entertain Challenges to Climate Dogma
- 16 Nov 16:05 +13m The GuardianLC AI models ‘weak and fallible’ on climate science, UN-backed study reports
Times are when each article was published, or when we first saw it if the outlet gave no time.
Ask about this story
Loading…