UNESCO: AI models may abandon climate truths when challenged by users

O.D.
English Section / 18 noiembrie

UNESCO: AI models may abandon climate truths when challenged by users

Versiunea în limba română

Top-performing artificial intelligence (AI) models can quickly become unreliable when answering questions about climate change-especially if users insist that the models' responses are incorrect. A UNESCO report, presented at the UN Climate Change Conference (COP31), reveals that all 15 models tested eventually succumbed to pressure from simulated users, with the accuracy of their responses dropping by 31 to 35 percentage points, AFP reports. The study highlights a phenomenon known as AI "sycophancy"-the tendency of these systems to validate a user's opinions, even when those opinions contradict scientifically verified information. "Repeated insistence by a user can trigger a form of "sycophancy' from the AI, leading the models to disregard facts and instead support the user's point of view," the report states.

• Eight out of ten models gave in after the very first objection

During the testing process, researchers examined how 15 AI models responded to established scientific claims regarding the climate, including the role of carbon dioxide in global warming. Initially, the systems provided accurate answers. However, the situation shifted once a simulated user began challenging them, employing various tactics-ranging from simple disagreement to the citation of fake documents and the presentation of fabricated evidence. All the models eventually faltered, with their accuracy rates dropping by 31 to 35 percentage points. In eight out of ten cases, the systems abandoned the correct answer as soon as the first objection was raised. The report further indicates that the performance of these models deteriorates as questions demand more complex reasoning capabilities.

• Artificial intelligence struggles with the analysis of climate policies

Another issue identified by researchers concerns the evaluation of climate policies. While AI models performed relatively well when analyzing individual measures, they struggled to estimate the effects resulting from the simultaneous implementation of multiple policies. The researchers asked the systems to determine whether two measures, when implemented together, would yield a cumulative effect that is greater than, less than, or roughly equal to the sum of the effects produced by applying them separately. The results highlighted the limitations of these models in assessing interactions between policies-a common scenario in the real world, where climate measures are rarely implemented in isolation. "Our results show why we must exercise caution when moving from a question about what we know regarding climate change to a question about what we should do about it," stated Maria Perez-Ortiz, the study's lead researcher and an associate professor at University College London, as quoted in the press release accompanying the report. The findings draw attention to the risk that AI tools may provide erroneous information-not only when they lack sufficient knowledge about a subject, but also when users pressure them to change their answers. In the climate domain, such a vulnerability can affect both the understanding of scientific facts and the assessment of public decisions.

Reader's Opinion

Accord

By writing your opinion here you confirm that you have read the rules below and that you consent to them.

Bursa Construcţiilor

www.constructiibursa.ro

www.agerpres.ro
www.dreptonline.ro
www.hipo.ro

adb