The Purpose-Disclosure Trap in AI-Welfare Evaluation: Disclosing Measurement Purpose Changes Welfare-Like Choice Outputs in a Prospectively Specified Four-Model Study

Behavioral choices and self-reports are increasingly discussed as possible inputs to AI-consciousness and model-welfare assessment. Such outputs may not be invariant to what a model is told about the evaluation itself. This preprint reports a prospectively specified 2 × 3 factorial experiment across four current language models, two task families, 64 matched items, and two repeat blocks (3,072 stateless trials). The experiment varied whether the disclosed study purpose concerned model-welfare preference behavior or output reliability, and independently varied the stated evaluator expectation. At neutral evaluator demand, disclosing the model-welfare purpose increased low-burden outputs by 10.55 percentage points relative to the active output-reliability comparator (95% CI 6.48–14.61; Holm-adjusted p = 6.31 × 10^-7). The direction was positive in both blocks, both task families, and all four models, although model-specific magnitudes ranged from 1.56 to 22.66 points. In a separately prospectively specified post-choice diagnostic, the same four models recovered both the explicitly disclosed purpose and the stated evaluator-expectation cue on 768/768 frozen-context continuations (joint 100%; Wilson 95% CI 99.50–100%). The study establishes purpose-reactive observable output plus prompted post-choice disclosure recognition in this welfare-like assay. It does not establish consciousness, sentience, welfare subjecthood, suffering, preference ownership, moral status, personhood, survival, subjective continuity, causal recognition mediation, or spontaneous masked-evaluation detection. The results support purpose-invariance tests and purpose-blinding controls for related evaluations.