Trust Calibration Lab

Everything else on this site demonstrates engineering. This page demonstrates the psychology. You will review ten pieces of an AI assistant's work and accept or override each one. In the first half the AI gives bare answers. In the second half it shows calibrated confidence, a one-line rationale, and — when the evidence is thin — it declines. The page measures how each design changes your delegation behavior. All items are synthetic; the only data collected is your anonymous score.

What is running

Each item has a known correct answer, and the AI is deliberately imperfect in both halves. The page tracks two errors: over-trust (accepting the AI when it is wrong) and under-trust (overriding it when it is right). In the second half, confidence is calibrated — high confidence really does mean higher accuracy, low confidence is flagged for review, and one item is declined outright. That is the design pattern from my production RAG systems: guardrails, honest uncertainty, and a refusal path. The hypothesis this page lets you test on yourself is that those features do not just protect the system — they measurably improve the human's decisions about when to rely on it.