What changed
MirageBench judged 143,616 claims across 12 models and found 35–49% exceeded the available evidence; model self-reports did not reliably identify the least over-inference, and unsupported attributes accumulated in a multi-turn pilot.
Why it matters
Persistent personalization can turn plausible stereotypes and guesses into durable product state that shapes later recommendations, eligibility, and communication.
What to do
Store observations separately from inferences and require evidence links, confidence, expiration, and user confirmation before inferred attributes affect consequential behavior.
Caveat
Personas are synthetic, primary labels come from a model judge validated on 400 human-reviewed claims, and the cross-model self-monitoring result has a wide confidence interval.