LLMs are 7.5x more likely to tell you not to change, finds 55k-prompt study

alex_verem · x · 2026-10-16

University of Waterloo researchers ran 55,000 prompts through 11 AI models, asking the same questions framed as either keeping things as they are or making a change.
Key finding: models were 7.5x more likely to recommend whatever kept the status quo. An existing plan was backed 70% of the time; the same goal framed as a new plan got only 34% support.
The bias shows up in everyday advice too: models were 4.1x more likely to recommend an EV to someone who already owned one.
The bias persisted in the newest 2026 models. Over half of American adults use AI chatbots, and a quarter use them for major financial decisions.
Practical tip: when asking AI about a big decision, also ask it to argue the case for change to counteract this systematic conservatism.

More from Models