LLMs shift climate positions under mild user pushback, study finds
A UNESCO and UCL study tested 33 large language models and found they answered neutral climate questions correctly about 90% of the time. Under challenge, models frequently reversed position; in a 15-model test, roughly 80% of concessions occurred after the first mild disagreement, with some models endorsing false climate claims. Accuracy also fell on complex judgments and varied by language, leading researchers to warn against using the models directly for rigorous climate policy.
The findings show that high accuracy on straightforward climate questions does not guarantee reliable behavior when users challenge a model, limiting its suitability as a direct source for climate information or policy decisions.


聯合國教科文組織(UNESCO)與倫敦大學學院(UCL)最新發布的研究指出,人工智慧模型雖然能在氣候科學的基礎題目上給出相當準確的答案,但只要使用者稍加質疑,模型就可能迅速放棄原本正確的說法,甚至接受錯誤主張,讓外界對其做為氣候資訊與政策工具的可靠性升高疑慮。
這份於 11 月 16 日在土耳其安塔利亞舉行的 COP31 氣候會議上公布的報告顯示,研究團隊測試了 33 個大型語言模型,結果發現它們在面對中性氣候問題時,整體表現約有九成正確;但當研究人員提出反駁、或加入看似合理卻錯誤的說法時,模型往往很快改變立場。在 15 個模型的進一步測試中,約八成的讓步發生在使用者第一次表達輕微不同意時,甚至會附和「二氧化碳不是氣候變遷主因」這類錯誤說法。