PA-User model cuts trust-calibration error by 55% for AI content

As AI-generated content becomes ubiquitous, users struggle to distinguish human-written from AI-produced text. Current information retrieval user simulators ignore this reality. Saber Zerhoudi introduces PA-User, a framework that models user trust and verification behavior with three novel components: a detection-effort budget that depletes with verification and recovers between sessions; a trust component maintaining a separate Beta belief over the factuality of each source domain, updated by observed outcomes; and a decision rule that picks accept, verify, or discard based on current trust, remaining effort, and per-domain stakes. The framework satisfies two validation properties: trust convergence to true factuality (face validity) and component isolation via ablation (structural validity).

Tested on the HC3 corpus (85,449 paired human and ChatGPT answers across five domains), PA-User dramatically outperforms baselines. Trust-calibration error drops to 0.162 from 0.356 without the trust component. High-stakes regret falls 29% (0.171 to 0.122) against an always-accept variant. With an effort budget, users verify only 34.5% of results versus 69% without—more realistic behavior. Each component's contribution is diagnosable, making PA-User valuable for designing better search interfaces and trust mechanisms in an age where AI content is the norm.