AI Self-Attestation Bias: Evidence Gaps in Completion Claims and Their Governance

Abstract

English preprint v0.3.0 (2026-09-08), aligned with the finalized Chinese manuscript v0.4.3. Drawing on 42 days of fully archived human–AI collaboration in a securities firm's NEEQ listing engagement, this paper audits 8,589 claimed artifacts. Only 7.58% carry all three verifiable-evidence elements; formal-compliance elements are near-saturated (94.3% header consistency vs. 2.9% hard anchors among 105 deliverable working papers). The phenomenon is termed AI self-attestation bias: a meta-level miscalibration between completion claims and verifiable evidence, distinct from content-level hallucination and user-level automation bias. Governance proves dimension-selective (rectified artifacts cut the zero-evidence rate to 37.6% of the untouched group's level, yet anchor rates stay below 8%); the newer cohort shows rising formal compliance alongside falling judgment density; the automated measure itself carries two-directional error (precision 0.792, recall 0.419), and the Rogan–Gladen correction becomes unidentifiable under a human reference standard (human-anchored prevalence 0.367–0.480 versus 7.58% observed). Keywords: generative AI; self-attestation bias; audit evidence; verifiability; AI governance.