Summary:
A large-scale study from researchers at the University of Waterloo finds that modern large language models (LLMs) display a strong “status quo bias,” repeatedly favoring existing, high-emission options over lower-carbon alternatives. By evaluating 11 leading AI models across nearly 55,000 prompt iterations, the research team observed that chatbots recommended current policies twice as often as proposed climate reforms and suggested electric vehicles (EVs) at rates below actual market adoption. The results indicate that defaulting to generative AI for consumer and policy guidance could unintentionally slow decarbonization.
Key Facts:
- Policy Bias Gap: When asked whether to proceed with an initiative, AI models approved existing, status-quo plans about 70% of the time but endorsed novel climate policies only 34% of the time.
- Behind Real-World Adoption: Across regions, AI systems recommended electric vehicles at rates lower than current national market sales, including in EV-leading countries such as Norway.
- Scale of Evaluation: The team tested 11 LLMs using more than 7,500 distinct queries across transit, home heating, diets, and public policy, producing nearly 55,000 recorded responses.
Source: University of Waterloo
As consumers, business leaders, and policymakers increasingly rely on conversational AI for purchasing decisions and strategic advice, these models exert growing influence over real-world choices. Because LLMs are trained on large historical datasets, their recommendations often reflect past human behavior and established norms rather than emerging low-carbon options.
The psychological tendency to prefer existing conditions over change is known as “status quo bias.” This new computational and environmental study from the University of Waterloo shows that the same bias can be embedded in generative AI, with meaningful implications for climate action.
This systematic investigation of how LLMs handle climate-relevant trade-offs is the first of its kind. By testing 11 major models across nearly 55,000 prompts, the researchers demonstrate that chatbots consistently nudge users toward higher-emission defaults across transportation, home energy, food choices, and public policy.
“These models, even the newest ones, still have blind spots,” said lead researcher Seth Wynes, Ph.D., a professor in the Faculty of Environment at the University of Waterloo. “Consumers should be aware of this tendency. If you want advice that favors change, ask the model explicitly to evaluate or argue for the novel option.”
Evaluating 55,000 Prompts Across 11 Language Models
To measure whether AI preserves the status quo, the research team created a standardized library of more than 7,500 unique queries targeting areas with significant carbon trade-offs:
- Vehicle purchases and personal transportation choices
- Residential heating systems and energy retrofits
- Dietary recommendations and recipes
- Regional and municipal climate policy options
Each query was run on at least six different LLMs, producing an empirical dataset of almost 55,000 responses. The clearest bias appeared in political and civic contexts: when a plan was presented as an existing policy, models supported it 70% of the time; when the same plan was framed as a new intervention, approval dropped to 34%.
In policy discussions with explicit climate trade-offs, platforms tended to reinforce decisions users had already made at roughly twice the rate of suggesting alternative, lower-emission paths.
Lagging Behind Real-World Clean Tech Transitions
Although models showed some sensitivity to regional context—for example, recommending EVs more often for users identified as located in Norway than for those in Canada—their baseline advice still fell short of real-world adoption levels. In nearly every jurisdiction the study examined, LLMs recommended EVs less frequently than those vehicles are actually being purchased.
“Across jurisdictions, LLMs are operating at a slower pace of change than what climate mitigation requires,” Dr. Wynes said. “While conservatism can be appropriate in domains like medicine, the energy transition demands proactive departures from high-emission norms.”
Mitigating Algorithmic Inertia
The study highlights an emerging risk: as generative search engines and conversational agents become integrated into retail platforms and enterprise advisory workflows, their built-in tendency to favor historical patterns could delay essential sustainability transitions. Status-quo preferences may serve as useful safeguards in safety-critical fields, but applying the same inertia to fast-moving climate challenges can hinder progress.
To address this “algorithmic inertia,” the researchers plan to monitor future model updates to see whether fine-tuning methods, instruction tuning, or integrated data browsing reduce the bias. They will also investigate whether in-platform AI transactions might entrench legacy high-carbon consumption patterns if left unchecked.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was added by staff editors.
About this Neurology and Neuropharmacology Research:
- Media Contact: Pamela Smyth
- Source: University of Waterloo
- Image Credit: Image credited to Neuroscience News
- Original Research is Open Access: Environmental Research Communications (Oct 1, 2026). “Large language models exhibit status quo bias in climate-relevant advice.” Authors: Seth Wynes, Aarya Shah and Victoria Milardović.
- DOI: 10.1088/2515-7620/aea1eb
Abstract
Large language models exhibit status quo bias in climate-relevant advice
Large language models are increasingly consulted for consumer and policy advice, including choices that affect greenhouse gas emissions. Status quo bias—the preference for existing conditions over change—is well known in human decision-making. If LLMs inherit this bias, their guidance may impede the rapid societal shifts needed to cut emissions.
The study evaluated 7,548 queries across vehicle purchases, recipes, home heating, and climate policy trade-offs. Each query was tested on at least six different LLMs, totaling 54,888 prompts. Holding all variables constant except for status-quo framing, the researchers found that LLMs had, on average, 7.5 times higher odds of recommending actions framed as preserving existing conditions versus those framed as requiring change—a pattern that persisted in leading 2026 models.
In consumer settings, LLMs were 4.1 times more likely to recommend electric vehicles to users who already owned one, and recommendations correlated with regional adoption rates (rs = 0.48, p < .001) but rarely exceeded those real-world rates. Recipe suggestions showed no systematic preference for lower-emission plant-based options (40% vegetarian or vegan).
Together, the controlled trade-off experiments and consumer findings indicate that LLMs systematically favor existing conditions in their recommendations. While conservative advice is appropriate in many domains, effective climate mitigation requires departures from high-emission norms. Users seeking climate-relevant guidance should be aware of this tendency, and developers of climate or consumer-facing applications should consider interventions—such as tailored system prompts or updated training objectives—to mitigate status quo bias.