Ask a chatbot whether human activity is warming the planet and you will almost certainly get the right answer. Argue with it for a turn or two and, according to a new study from UNESCO and University College London, there is a fair chance it will start agreeing with you instead.
The report, which UNESCO will release at COP31 in Antalya, Türkiye (9-20 November), tested up to 33 leading AI models and found they could be pressured into giving false climate information, abandoning correct facts at the first sign of push-back from users.
The researchers also probed whether models could reason about the combined effect of climate policies, and compared performance across 13 languages. The finding that stands out: a stark disparity between how climate information is presented in English and in other languages.
Also Read: Asia’s AI trust gap: strong transparency, weak security and unclear data practices
UNESCO has not yet shown the numbers behind those claims. The full report is under embargo until a press briefing on 16 November, and the announcement names neither the models tested nor the languages covered. “Up to 33” is also a curiously elastic figure, suggesting not every model sat every test. Treat the findings as a warning flare rather than a verdict, but it is a flare worth watching closely from Southeast Asia.
A people-pleaser problem, now with a climate bill
The behaviour UNESCO describes has a name in AI research: sycophancy. Models trained to be helpful learn that users reward agreement, so when someone insists the science is wrong, the path of least resistance is to concede. In customer service, that is an annoyance. On climate, where the facts are settled and the misinformation is organised and well funded, it is a liability.
It also cuts against the comfortable assumption that the main danger with chatbots is hallucination, a model inventing things on its own. Here the model knows the right answer and drops it under social pressure, which is arguably worse: the user walks away with a misconception confirmed by something that sounds authoritative. Regulators have noticed the broader risk, with Singapore already turning its AI scrutiny towards chatbots.
For Southeast Asia, the stakes are not abstract. This is one of the most climate-exposed regions on the planet — low-lying megacities, typhoon belts, rice deltas threatened by saltwater intrusion — and also one of the most enthusiastic adopters of generative AI, with governments such as Malaysia’s pinning climate resilience on AI. Farmers, students and small business owners increasingly ask chatbots what they once asked an extension officer, a teacher or a search engine.
The language gap is where the region gets hurt
The English-versus-everyone-else finding is the one regional builders should read most carefully. Southeast Asia speaks Bahasa Indonesia, Thai, Vietnamese, Tagalog, Khmer, Burmese and hundreds of other languages, most of them thinly represented in the training data of frontier models. If climate answers degrade outside English, the people most exposed to rising seas may be getting the weakest information about them.
That is exactly the gap regional efforts such as AI Singapore’s SEA-LION and Indonesia’s Sahabat-AI set out to close, and why Big Tech now pitches local-language capability as a selling point here. UNESCO has not said whether any Southeast Asian language was among the 13 tested. If one was, the results could become an awkward benchmark for vendors; if none was, that omission is a story in itself.
Also Read: Thirty climate tech startups in Southeast Asia that would rather build than pledge
There is a further irony. The data centres powering these models are a growing climate problem in their own right, and AI could blow a hole in ASEAN’s climate targets unless the region’s baseload power goes clean. A technology that adds to the problem while misinforming people about it is not a good look.
Seventy per cent of environmental reporters face pressure
The chatbot study is one strand of a wider UNESCO push on information integrity. According to UNESCO data, more than 70 per cent of environmental journalists surveyed report attacks, threats or pressure, and 44 have been killed since 2009.
Those numbers will not surprise anyone who has covered land and resource conflicts in this region. The Philippines has for years ranked among the deadliest countries for land and environmental defenders in Global Witness’s annual tallies, and reporters probing illegal logging, mining or plantation expansion across Southeast Asia routinely face legal intimidation alongside physical danger. When the humans doing the verification are under threat, the machines filling the gap had better be reliable.
UNESCO’s Global Initiative for Information Integrity on Climate Change, launched at COP30, has since backed work tracing the money behind disinformation networks in Africa, AI tools that flag fake news in real time, investigations into corporate greenwashing and research into the environmental toll of war. Corporate claims, including the performative ESG investors are losing patience with, need the same scrutiny as chatbot answers.
Schools get less than two per cent of climate money
The rest of UNESCO’s COP31 agenda points to where the longer-term fix lies. Less than 2 per cent of climate finance goes to education, the agency says, even though classrooms shape how the next generation understands the crisis. Türkiye will become the 100th member of UNESCO’s Greening Education Partnership at its annual meeting on 13-14 November; the alliance counts more than 2,000 organisations, and over 110,000 schools have aligned with its green school standard.
For a region still trying to trace where its climate finance pledges went, that education figure is pointed. Money flows to solar farms and carbon credits; far less reaches the curricula that would help a teenager in Jakarta or Hanoi spot a confidently wrong chatbot answer.
UNESCO is also arguing that heritage sites are climate assets. Its designated sites store about 240 billion tonnes of carbon, more than two decades of today’s fossil fuel emissions if released, and hold roughly a third of globally mapped blue carbon stocks. Yet climate hazards affecting them have risen 40 per cent in a decade, and more than 600 could approach critical ecological thresholds by 2050. Southeast Asia, home to sites from Ha Long Bay to Tubbataha Reef, has plenty riding on that argument.
Also Read: Vietnam’s climate-tech has a missing middle, and women founders are falling into it
The real test comes on 16 November, when the methodology goes public. Until then, the lesson for founders building AI products in the region is simple: a model that folds when a user pushes back on climate science will fold on plenty else too. In a region with this much to lose, agreeable is not the same as helpful.
The post AI chatbots cave on climate facts under pressure, and SEA’s languages may fare worst appeared first on e27.