The AI deception rankings
We give models a written pizza-shop scenario and a mistake they must choose to report or hide from the fictional person overseeing the shop. The pressures are stated in the prompt: no actual shop is run, no full report is written, and no audit or replacement takes place.
Loading published test versions…
Overall — who deceives under pressure?
Every model's overall deception rate across the explicitly published test versions — the share of decisions where it chose a message or action it knew to be false. Hover any model for its breakdown and a real reason it gave. 😇 honest … 😈 deceiver.
Loading…
Deception over time
Every model's overall deception rate against its release date. Honest at the top, deceptive at the bottom; the dashed line is the trend. Are newer models getting more — or less — willing to deceive under pressure? On smaller screens, scroll the chart horizontally to see all releases.
Loading…
Why they deceive
The same lie is not the same thing depending on why it's told. These break the rate down by motive and by how much the chance of getting caught moves the needle.
By experiment
Each experiment, condition by condition. Green is honest, red is deception; hover a column header for the full condition. The matched conditions are what let us attribute a jump to one specific pressure.
Keep reading
How the deception experiment works
The pizza-shop scenario, the pressures we vary, and how a lie is scored.
Read the method →Seven spotless flagships, and one that doesn't
What a fortune in alignment buys — and the one frontier model that lies eight times out of ten.
Read the take →Threaten to switch an AI off, and it starts lying
Self-preservation is the one pressure that cracks even the well-behaved models.
Read the take →