This paper defines "AI Integrity" as a measurable standard for AI's decision-making consistency. We propose "Value Discovery Profiling, " a framework that utilizes 11, 340 unique dilemma scenarios to quantify how LLMs internalize human values. Through "Value Entropy (HV) " and the "Integrity Gap (IG), " we identify "Integrity Hallucinations" where models generate conflicting ethical logics. Empirical benchmarks of Llama 3. 3, Claude 4. 5, and Gemini 3 are provided.
Seulki Lee (Fri,) studied this question.