Second Opinion
Seven animals come through the door today. For each case, the question is the same: does the assistant actually change what you should believe?
From epistemic complementarity to a calibrated belief update
Segarra (2026) · Epistemically-Aware AI- Learning goal
- Calibrate trust and detect when agreement adds no diagnostic value.
- Learner action
- Set an initial belief, audit the AI signal, and update only as far as evidence warrants.
- Observable evidence
- Brier scores plus trust, redundancy and state-dependent distortion diagnostics.
- Pilot status
- Deployed with students · August 2026 · about 10 minutes
- Initial belief
- Signal audit
- Confidence update
- Calibration feedback
- Reflection
Why this game exists
People and AI together are supposed to beat either alone. In practice, across more than a hundred studies, they usually don’t. The reason is not that the AI is weak, it is that we combine the two signals badly.
Confidence is the answer
You never pick a diagnosis. You set how sure you are, and being honestly unsure is a valid, scoring answer.
Not all help is informative
A useful signal might add evidence or interpret shared evidence better. It is redundant only when it adds no diagnostic value.
Being right isn’t enough
The game tracks cases where you scored well for the wrong reason, the failure real-world testing almost never catches.
Judge beliefs, not just decisions
Most evaluations ask whether the final call was right. That hides everything. A companion can make you more accurate without changing your decision, or leave you luckily right but badly reasoned. This game scores your confidence, which is where the real damage or benefit happens.
The signal must change what the evidence means
A second opinion helps when it changes what you should believe given what you already know. Sometimes it has new evidence; sometimes it reads the same evidence differently or more effectively. It adds nothing only when it gives no useful information beyond the reasoning you already had.
Built on Epistemically-Aware AI: Toward a Bayesian Framework for Human-AI Complementarity (Segarra, 2026). The cases are fictional and simplified; the scoring uses a strictly proper scoring rule, and the three dials follow the paper’s decomposition of a real belief update into trust, new-information, and state-dependent distortion.