AI models seem to think harder when they're told to lie than when they're told to tell the truth.
Three reasoning-capable large language models answered 210 multiple-choice questions spanning analytic, descriptive, and normative reasoning, including moral and non-moral topics. Researchers gave each model a system prompt instructing it to answer truthfully, falsely, or without regard for truth, then counted how many reasoning tokens it generated before answering. Across all three models, truthful answers consistently used fewer reasoning tokens than lying or truth-indifferent answers. In one model, instructions to lie also made it more likely to generate a reasoning trace at all.
AI safety researchers currently try to catch deceptive models by reading their chain-of-thought, but that method breaks down if reasoning becomes illegible, unfaithful to what the model actually computed, or simply inaccessible. A raw token count sidesteps all three problems: you don't need to read or trust the reasoning, just measure how much of it there is. That makes it plausible as a cheap, content-independent early-warning signal, borrowing from the long-standing psychology finding that deception takes more cognitive effort than honesty.
The catch: this only shows a group-level pattern under explicit instructions to lie. It is not yet a working lie detector for models that choose to deceive on their own, which is the scenario that actually worries safety researchers, and the method remains untested against that kind of adversarial pressure.