Locunity — where decisions happen.
Cover image for UNESCO: AI Models Ace Climate Facts, Falter on Policy Combinations

UNESCO: AI Models Ace Climate Facts, Falter on Policy Combinations

COP 31 Press Conferences • United NationsNovember 16, 2026

Sources:

Locunity is an independent informational service and is not an official government page for this commission. All information is sourced from public footage and an analysis of official agendas with the aid of responsible, fact-checked AI. to report an error or omission.

Preview only

UNESCO tested 33 artificial intelligence models on 656 climate questions. The models scored as high as 96% on textbook facts but fell to 33.8% when asked how two policies would interact. The findings, released at a COP31 press briefing on Nov. 16, 2026, arrive as negotiators, legislatures and regulators increasingly lean on chatbots for policy analysis. The answers can sound authoritative while performing little better than chance.

  • 33.8% accuracy on two-policy combinations, down from 94.5% on single-policy effects
  • Accuracy dropped 5 to 10 points in Swahili, Bengali, Urdu and Hindi
  • Models abandoned correct answers in 8 of 10 pushback cases
  • UNESCO urges governments to use procurement to demand reliability testing and transparency

The basics: UNESCO ran 33 AI models through 656 questions and checked answers about policy effects against En-ROADS, a climate policy simulator. The questions ranged from basic climate science to how combinations of policies shape outcomes.

Get reports in your inbox

Follow this commission for free and get the next report delivered by email. You'll be able to access the full archive, get real-time updates, and track the topics or keywords you care about most.

UNESCO: AI Models Ace Climate Facts, Falter on Policy Combinations | COP 31 Press Conferences | Locunity