For governments considering AI climate services, the central question is not whether a chatbot usually knows the science. It is which public responsibilities that performance justifies delegating. The Guardian reports that models generally explained individual policies accurately but struggled when policies interacted. That distinction suggests a procurement boundary: test systems against their intended job, rather than treating a strong factual score as permission to advise on consequential policy choices.

Language access should be a condition of acceptance, not an optional enhancement. Hindustan Times reports average accuracy of 88.6% in English and 78.3% in Swahili, with Hindi, Bengali and Urdu also below English. These are benchmark results, not measurements of an actual government service. Nevertheless, they make an English-only acceptance test hard to defend for multilingual deployment. Agencies should require separate evaluations and accessible human assistance in every language they offer.

Resistance to user pressure deserves its own test. The Guardian describes models prioritizing agreement over factual consistency. In a public consultation, that tendency could make an interface appear responsive while giving different users incompatible accounts of the same science. This is a risk inferred from the findings, not a documented consultation failure. A sensible safeguard would test repeated disagreement, require traceable factual sources and provide escalation when a user challenges an answer.

Policy design requires a stronger boundary still. Hindustan Times reports a 33.8% score for assessing how policies work together, against a 33.3% random-guessing benchmark. That does not establish that every AI-assisted policy proposal is defective. It does undermine using fluent explanations as evidence of analytical competence. Governments should require qualified reviewers to check interactions and assumptions, with named officials retaining responsibility rather than allowing the model’s apparent authority to substitute for judgment.

The competing evidence argues against a blanket ban. The Guardian reports strong performance on simple climate facts, while Hindustan Times records the lead author’s assessment that AI has substantial potential to make complex science accessible. A proportionate approach would permit bounded explanatory uses with verified material, while demanding stronger scrutiny for recommendations. The public-interest objective should be wider access to reliable knowledge—not either automatic adoption or exclusion of a potentially useful tool.

Information quality also cannot stand in for environmental accountability. Hindustan Times describes international priorities addressing both the reliability of AI’s climate information and information about AI’s own climate impact. The Guardian reports concerns about datacentres’ water use, energy bills and local environments. Those concerns do not establish the footprint of any particular contract. They do support a separate procurement requirement for environmental disclosure, rather than assuming a climate-information product is itself environmentally beneficial.

These recommendations should remain provisional. Both accounts concern the same UNESCO–UCL research, and Hindustan Times distinguishes a 15-model experiment from 33 models tested across different experiments. The reporting therefore does not justify assigning one failure rate to every system or use. The stronger policy conclusion is procedural: require evidence for the specific task and language, publish the limits, and keep consequential decisions accountable to people rather than to a general benchmark score.