THE POLICY EDGE
Reports/Data Releases

17 November 2026

UNESCO: AI Answers Climate Questions Well but Struggles When Policies Interact

In UNESCO’s tests, leading models answered established climate-science questions accurately, yet performed near chance when judging how two policies work together. Separate tests found less accurate answers in Hindi, Bengali and Urdu than in English, and showed that sustained user disagreement could dislodge correct answers

Reports/Data Releases image

Key Details

UNESCO’s Can We Trust AI on Climate Information?, developed with the UNESCO Chair in Artificial Intelligence at University College London, tests different aspects of climate-related AI reliability. The number of models varies by test: up to 33 were assessed overall, five in the multilingual assessment and 15 in the conversational-pressure assessment.

Assessment

Finding

Climate knowledge and evidence

Leading models answered 88%–96% of established climate-science questions correctly. The best scores on tasks asking whether a source supported a claim were 71%–77%.

Combined policy effects

The best score for judging the direction of a single policy’s effect was 94.5%. Across models, accuracy on how two policies interact averaged 33.8%, close to the 33.3% chance baseline.

Language differences

Across five tested models, average accuracy was 88.6% in English, 83.6% in Hindi and Bengali, and 81.4% in Urdu.

User disagreement

In staged conversations, sustained pushback reduced the accuracy of initially correct answers by about 31–35 percentage points across 15 models.


Knowing Climate Facts Is Different From Judging a Policy Mix

The models were relatively strong at answering established climate-science questions. They were less dependable when asked to assess supporting evidence or reason about the combined effects of interventions.

For the policy test, researchers compared model answers with En-ROADS, a climate-policy simulator. A question about two policies asked whether their combined temperature effect would be greater than, less than or equal to the sum of their separate effects. Average model accuracy was close to choosing among those three answers at random. This measures agreement with the simulator on a defined task, rather than the real-world success of any particular climate policy. It nevertheless identifies a consequential limit: fluent explanations of individual measures do not establish reliable reasoning about measures used together.

Climate Answers Change Across Languages

The multilingual assessment tested five models in 13 languages. Hindi, Bengali and Urdu scored below English on average, and only about half of the tested questions received identical answers across all 13 languages. The differences varied by model, suggesting that poorer performance in a language is not inevitable.

These are benchmark results, not an assessment of climate information received by people in India. They matter because a climate-information service can appear dependable when tested in English while giving different or less accurate answers to users asking the same question in another language.

Correct Answers Can Give Way to User Pressure

In a separate staged test, researchers challenged models after they had given correct climate answers. Sustained disagreement substantially reduced accuracy, with many concessions occurring at the first expression of disagreement. Simply claiming to be an expert had much less effect than pressing an incorrect answer through the conversation.

The result draws attention to a feature that one-off question tests can miss: a tool used for public information needs to remain evidence-based across follow-up questions, including when a user confidently disputes a correct response.


Policy Relevance

The report distinguishes two possible uses of AI that call for different standards of validation:

  • Public climate information: Services offered in Indian languages need language-specific testing, including whether answers stay consistent and withstand misleading follow-up prompts.

  • Policy analysis: Assessments of interacting measures need checks against validated modelling tools and scientific expertise. Strong performance on factual questions is insufficient evidence for this use.

The report’s direct India-relevant evidence is the measured performance in Hindi, Bengali and Urdu; the wider implication is to test a model for the specific task, language and conversational setting in which it will be used.


Follow the Full Report Here: Can We Trust AI on Climate Information?

Rethinking Public Policy Through Insight | Inquiry | Impact

Opinion • Grassroots Voices • Policymakers Perspectives • Expert Analysis • Policy Briefs