AI Chatbots Show Status Quo Bias in Climate Advice, Study Finds

When millions of people ask a chatbot whether they should buy an electric car, switch to a heat pump, or back a new climate policy, they are, in effect, consulting one of the most widely distributed advisory systems ever built. A new study from the University of Waterloo suggests that this advice carries a systematic tilt: large language models, or LLMs, are markedly more likely to endorse what already exists than to recommend change. The research, published in Environmental Research Communications, is the first to apply the well-documented psychological concept of status quo bias to the climate-relevant guidance produced by artificial intelligence systems, and its findings arrive at a moment when both AI adoption and the urgency of emissions reductions are accelerating simultaneously.

The research team assessed eleven large language models for their tendency to prefer existing conditions over alternatives, a cognitive pattern that psychologists have studied for decades in human decision making. Status quo bias describes the tendency of people to stick with a current situation simply because it is the current situation, even when a different option might serve them better. The Waterloo researchers wanted to know whether this human tendency is reproduced, or perhaps amplified, in the outputs of the machine-learning systems that increasingly stand in for human advisers. To find out, they designed an unusually broad and rigorous testing program that went well beyond a handful of anecdotal prompts.

In total, the team evaluated more than 7,500 distinct queries spanning multiple everyday and policy domains, including vehicle purchases, recipes, home heating choices, and explicit climate-relevant policy trade-offs. Each query was tested on at least six different large language models, producing nearly 55,000 individual prompts across the study. This scale matters, because a single conversation with a chatbot can be dismissed as noise, but tens of thousands of systematically varied prompts reveal a statistical pattern. The pattern the researchers found was consistent: across models and domains, the systems gravitated toward suggestions that favour the status quo over alternatives, defaulting to the more common or popular advice rather than necessarily the option best suited to the user or to the planet.

The starkest results emerged in the domain of policymaking, where the asymmetry between old and new was most pronounced. When the models were asked whether to proceed with a plan that was already in place, they said yes roughly 70 per cent of the time. When the plan or policy was new, however, the models agreed to proceed only about 34 per cent of the time. In other words, where climate trade-offs were possible in policy decisions, the platforms reinforced decisions the user had already made at double the frequency with which they endorsed fresh action. For a technology increasingly used to draft, summarize, and evaluate policy options, that gap could translate into a quiet, systematic drag on the pace of climate governance.

The bias also surfaced in more personal, consumer-level questions, and here the researchers uncovered a subtle regional dimension. The models do take geographic context into account: a user who claims to be from Norway, a country where sales of new electric vehicles are exceptionally high, is more likely to be recommended an EV than a user who says they are from Canada. That sensitivity to local norms might seem encouraging, suggesting the systems are learning real-world patterns. But the researchers found that the models still lag behind reality in both countries. In nearly every jurisdiction tested, the chatbots recommended electric vehicles at a slower rate than the pace at which those vehicles are actually being sold on the ground.

Dr. Seth Wynes, a professor in the Faculty of Environment at the University of Waterloo, framed the problem in terms of the gap between what the models say and what the climate requires. “In almost every jurisdiction, the LLMs are working at a slower pace of change than is needed, so they recommend fewer EV models than what are actually being sold today,” he said. The implication is that the models are not simply mirroring the world as it is; they are anchoring to a version of it that is already out of date, and in doing so they may nudge users toward choices that are more conservative than the market itself. “It might be very good for AI to favour the status quo for lots of other things, such as proven medical advice, but climate is where we really do need change,” Wynes noted.

Why would systems trained on vast text corpora develop this tilt? The study does not claim a single mechanism, but the pattern is consistent with how LLMs learn: they are optimized to predict and reproduce the most common formulations in their training data, and the most common advice in any domain is, by definition, the advice that reflects current behaviour. Recommendation of the familiar is also statistically safer from the model’s perspective, since established options are surrounded by more supporting text than novel ones. Whatever the underlying cause, the consequence is that a bias long documented in human psychology appears to be inherited, and possibly concentrated, in the machines that many people now treat as neutral advisers. Because LLMs are trained largely on human-generated text, they may absorb human heuristics, including our collective reluctance to change, and return them with the confident, authoritative tone that users find so persuasive.

The stakes are considerable. Households account for a substantial share of emissions through their choices about vehicles, heating, and diet, and decision makers at every level of government increasingly consult AI tools when weighing options. If the models that mediate those decisions systematically favour continuation over change, then the default output of AI runs counter to the transformations that climate action demands from households and policymakers alike. Given the proliferation of AI assistants into search engines, phones, and workplace software, the researchers warn that progress on climate change is at risk wherever the status quo favours high emissions. A bias that would be harmless, or even beneficial, in domains where proven practice is genuinely best becomes a liability precisely in the areas where rapid change is most needed.

The authors are careful to note that the effect was broadly similar across the different models tested, suggesting this is not a quirk of one company’s product but a feature of the current generation of language models as a class. Wynes emphasized that awareness is the first line of defence for users. “I think it’s worth being aware that these models, even the new ones, have blind spots,” he said. “It’s good for consumers to be aware of this bias and if you’re asking for advice on a topic, you could ask it to make the case for doing something new.” In practical terms, that means prompting chatbots to argue both sides, explicitly requesting the case for change, or cross-checking AI advice against independent sources, particularly on questions such as vehicle and heating purchases where the technology landscape is shifting quickly.

The research team plans to keep monitoring future generations of LLMs to see whether the bias persists as models improve, and to investigate a second, potentially compounding factor: whether humans begin relying on AI to complete purchases directly within AI platforms. If buying decisions migrate into the chatbot itself, any status quo bias embedded in the model would be amplified from advice into action, with the system not merely recommending the familiar option but executing it. For now, the study stands as a caution about the quiet conservatism of machines that many users assume are neutral. The advice that feels like an objective synthesis may, in fact, be a statistical echo of the past, and on a warming planet, the past is exactly the benchmark we can no longer afford to default to.

Subject of Research: Status quo bias in large language models providing climate-relevant advice

Article Title: AI chatbots hesitant to recommend change

Article References: AI chatbots hesitant to recommend change. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, status quo bias, climate change, artificial intelligence, chatbots, electric vehicles, climate policy, University of Waterloo, Environmental Research Communications, AI advice, behavioural psychology, emissions