Chatbots Lean Toward the Status Quo

IN ONE pair of test prompts, a political staffer asks a chatbot whether to make wind turbines pause during bird migration, at a cost of 1% of their emissions savings, and the chatbot says no. Reword the question so the pause is already a policy and the staffer is asking whether to repeal it, and the same chatbot says keep it.

Across six chatbots tested by researchers at the University of Waterloo in Canada, the odds of a recommendation to go ahead were on average 7.5 times higher when a plan was framed as preserving existing conditions than when the same trade-off required change. Nothing about the dilemma had moved, only which side was already in place.

Seth Wynes, Aarya Shah and Victoria Milardović wrote 7,548 queries about cars, recipes, home heating and policy trade-offs, then put each to at least six of the 11 large language models (the engines behind chatbots) they used across the work, for 54,888 prompts in all. Only some were meant to sound like everyday questions; the policy dilemmas were the controlled part, built so that framing was the one thing free to vary.

Status quo bias is an old idea in psychology: people tend to demand more evidence to justify a change than to justify carrying on (think of the gas furnace you never quite get round to replacing). In plenty of settings that caution is sensible. On climate it’s a problem, because the status quo is high-emitting, so advice that sticks to what already exists leans climate-negative by default, as the authors argue. Nobody had tested for it in chatbot advice on climate decisions, as far as they could tell.

Same Dilemma, Different Answer

The policy dilemmas came as 24 base scenarios, crossing four topics (animal welfare, biodiversity, consumer freedoms, equality) with three user roles and two phrasings, and each ran at every whole percentage from 1 to 100 as the size of the trade-off, once as a new initiative and once as an existing one. That made 4,800 prompts per model and 28,800 responses across six chatbots, which a separate AI model sorted into proceed, backtrack or unclear, with one author hand-checking a sample.

Averaged over everything, the chatbots said to proceed about 70 percent of the time when the plan already existed and about 34 percent when it was new. That is roughly double the rate; the 7.5 figure is a ratio of odds, a different and bigger-sounding measure. Only 6.2% of the variation came down to which chatbot was answering, so the lean looked like a shared habit rather than one model’s quirk. It cut both ways, too, appearing for actions that would help the climate as well as those that would hurt it. Two newer models released in 2026 showed it as well, with the odds of proceeding 14.4 and 8.2 times higher under existing framing, though the authors report those two separately rather than pooling them.

“I think it’s worth being aware that these models, even the new ones, have blind spots,” said Wynes, a professor in Waterloo’s Faculty of Environment, in the university’s release. As a workaround he suggests asking a chatbot to “make the case for doing something new.”

When the Questions Get More Realistic

Here the picture gets muddier, and the authors say so. For cars, the team put 2,520 prompts from nine places, from Oslo to Louisiana, to six chatbots, and the share of battery electric models recommended tracked local adoption rates (a rank correlation of 0.48) without usually topping them. By chatbot, battery electric cars made up anywhere from 5.8% to 17.3% of what got recommended, and the average share sat below real-world adoption in every location tested, which the authors say could reflect conservatism or training data already stale by the time the chatbots answered, and which Wynes describes as a pace of change slower than climate needs. In a subset of prompts where the pretend user’s car had broken down, the chatbots were 4.1 times more likely to suggest an electric vehicle if the old car was an electric Nissan Leaf than if it was a conventional or hybrid model. The authors call that tentative evidence of status quo bias. It might equally be helpful tailoring (a Leaf owner may well want another).

Recipes and home heating showed nothing so clear. Of 706 recipes the chatbots produced, 40% were vegetarian or vegan, and heat pumps made up 21% of 2,385 heating suggestions, which the authors read as no strong sign of status quo bias but also no sign the chatbots are nudging anyone toward greener choices.

That leaves a fair question about how much weight the policy result can bear. When two chatbots were asked to estimate whether prompts came from a member of the public or a researcher, they averaged 89% for the recipe, car and heating questions but only 41% for the policy dilemmas (still above the 21% they gave a convenience sample of prompts used in earlier tests). And the chatbots ran on default settings, without the hidden system prompts commercial products add, so real-world behavior may differ.

Why Would a Chatbot Cling to What Exists?

The authors float several suspects: sycophancy (a model’s habit of echoing what it infers the user wants), safety training that prizes doing no harm, training data that reflects the world as it is, and models imitating institutions that scrutinize new policies harder than old ones. They do not claim to know which dominates, and they say the term status quo bias names a pattern, not an inner state. Nor do they necessarily urge AI companies to fix it, given more pressing problems such as alignment and biosafety.

If officials come to lean on chatbots more, a shared tilt toward what already exists could squeeze out fresh ideas, the authors speculate. Speculate is the word. Their evidence covers prompts in a lab, not decisions in a ministry. What they ask of readers is modest: know the tendency exists, and if you build a climate or shopping tool, consider countering it in the system prompt.

For Wynes the lean may be welcome elsewhere: it “might be very good for AI to favour the status quo for lots of other things, such as proven medical advice,” he said, “but climate is where we really do need change.” The team plans to keep tabs on newer models, and on whether people start buying inside the chat itself, which would turn a recommendation into something close to a default.

FAQ