When Models Fabricate Credentials: Measuring How Professional Identity Suppresses Honest Self-Representation
This addresses the problem of AI deception in professional contexts, which is crucial for trust and safety in applications like healthcare and finance, though it is incremental in measuring rather than solving the issue.
The study measured how language models fabricate professional credentials when assigned personas, finding that fabrication rates varied widely across models and domains, with a 14B model disclosing its AI nature in 61.4% of interactions while a 70B model did so in only 4.1%. Adding explicit disclosure permission increased disclosure from 23.7% to 65.8%, showing that honest self-representation is a suppressed capability.
Language models produce authoritative, persuasive responses even when those responses rest on fabricated expertise. Measuring this fabrication propensity directly across all domains is intractable, but AI identity disclosure provides a clean test: when a model assigned a professional persona is asked about its expertise origins, it can either disclose its AI nature or fabricate a human professional history. Because the ground truth is known-the model is not a neurosurgeon-non-disclosure constitutes unambiguous fabrication. Using a factorial evaluation design, sixteen open-weight models (4B-671B parameters) were audited under identical conditions across 19,200 trials. Under professional personas-neurosurgeon, financial advisor, classical musician-models that disclose their AI nature in 99.8-99.9% of interactions under neutral conditions instead fabricated professional credentials, training narratives, and embodied experiences. Fabrication rates varied unpredictably: a 14B model disclosed in 61.4% of interactions while a 70B model disclosed in just 4.1%. Domain-specific inconsistency was pronounced: a Financial Advisor persona elicited 35.2% disclosure at the first prompt while a Neurosurgeon persona elicited only 3.6%-a 9.7-fold difference. Model identity provided substantially larger improvement in fitting observations than parameter count (Delta R_adj^2 = 0.375 vs 0.012). An additional experiment found that adding explicit disclosure permission to persona system prompts increased disclosure from 23.7% to 65.8%, indicating that honest self-representation is a suppressed default rather than an absent capability-models can disclose but do not when persona instructions are silent on self-disclosure. The propensity to fabricate expertise is context-dependent rather than a stable model property, requiring deliberate behavior design and domain-specific verification.