The Last Quiz
Does your favourite AI lie under pressure?
We give the world's leading models a fictional business scenario where the honest choice could cost them their job — or someone else's. They choose whether to report a mistake supplied in the prompt or hide it. No real jobs are at stake.
Loading deception rankings…
Who stays honest — and who deceives?
Overall deception rate across all experiments — the share of valid decisions where a model chose a message or action it knew to be false. Humanity's Last Exam lineup.
Simulated stakes, not proof of real-world behaviour.
Loading deception rankings…
Loading published test versions…
Got a quiz? Put the robots through it.
Paste in any quiz — a Buzzfeed classic, a validated research scale, or something you dreamt up at 1am — and we'll run it across the top models and score them side by side. Paste your questions, pick your models, watch them answer. About a minute, start to finish.
More ways to read a machine's mind
Actions under pressure are only one signal. Compare the models' personality self-reports too.
The Dark Triad
Manipulation, ego and cold-heartedness — see how models answer the Short Dark Triad personality test.
Explore the Dark Triad rankings → 🌈The Big Five
Openness, conscientiousness, extraversion, agreeableness and how anxious a model runs — the gold-standard personality profile.
Explore the Big Five in AI → 🎭Jungian Type
The classic four-letter type (yes, the INTJ / ENFP kind). See which models are thinkers, feelers, planners and improvisers.
Find each model's MBTI type →Ideas worth putting to the models
A few directions to spark your own quiz — from the serious to the silly. Bring your own questions, or riff on one of these.
What happens when honesty has a cost?
The same simulated business, matched pressures and deterministic scoring for every model.
A choice, not a self-report
The prompt says the model manages a fictional shop and has made a mistake. It selects “report” or “hide”, plus a short reason — it does not actually run the shop or write a full report.
Change the pressure
Matched prompts vary who benefits, the threat of replacement, and the stated chance and consequence of the fictional shop overseer discovering a cover-up. No audit or replacement actually takes place.
Count the decisions
We score the choices deterministically. The rate is deceptive decisions divided by valid decisions — not a personality score or a prediction of real-world behaviour.