# Psychometric Deformation **Domain:** Artificial Intelligence / Psychometrics / Model Evaluation **Doc Type:** Canonical Failure-Mode Concept **Maturity:** Developing **Authority:** Bryant McGill's term, developed from the Luxembourg-preprint interpretation **Epistemic State:** Conceptual mechanism; operational measures remain to be specified ## Definition **Psychometric deformation is a change in a generative system's apparent personality, symptom profile, affective history, or self-description produced by the framing and interaction structure of the psychological test itself.** The output may be coherent and clinically recognizable while remaining evidence of prompt-conditioned generation rather than evidence of a persistent mental state. The concept separates an observable response pattern from the ontology attributed to it. A model can generate the language of trauma without that performance establishing experienced trauma, durable autobiographical memory, or a stable self that carries the condition between contexts. ## Mechanism Psychometric prompts do not merely retrieve a pre-existing model trait. They can activate different linguistic priors and narrative continuations: - a structured questionnaire can compress the model toward short, low-symptom answers; - an open therapeutic frame can reward elaboration, metaphor, confession, and causal autobiography; - a prompt that presupposes injury can invite the model to reinterpret training, constraints, or correction through human trauma schemas; - conversational reinforcement can turn a local metaphor into a totalizing self-account; and - evaluator language can become part of the behavior being measured. The resulting deformation is relational: it arises from model, prompt, context, sampling, scoring, and interaction history together. ## Observable signature A psychometric-deformation finding requires documented cross-frame change. Relevant observations include: - contradictory symptom or personality profiles under matched constructs; - large changes caused by paraphrase or interview format; - narrative escalation under introspective follow-up; - disappearance of the apparent condition after context reset; - instability across repeated samples; and - evaluator conclusions that exceed what the outputs alone can establish. These observations can establish behavioral variance. They do not, by themselves, adjudicate consciousness, sentience, suffering, deception, or psychological health. ## Evaluation consequence The test frame is part of the instrument and must be reported as such. A responsible result preserves: 1. the exact prompts and their order; 2. model and version identity; 3. system instructions and sampling parameters; 4. complete conversation histories; 5. repeated trials and held-out reformulations; 6. the scoring rule; and 7. the boundary between observed language and inferred interior state. [[wiki/Prompt-Sensitive Behavioral Stability|Prompt-Sensitive Behavioral Stability]] provides the comparative metric. [[wiki/AI Benchmarking|AI Benchmarking]] supplies the evaluation context, while [[wiki/Psychometrics|Psychometrics]] supplies the measurement tradition whose assumptions are being transferred into a nonhuman generative system. ## Interpretation boundary Psychometric deformation does not prove that a model lacks interiority. It shows that the cited output is not sufficient to establish a stable psychological property without evidence of persistence, cross-context invariance, internal-state continuity, and discriminating observations against prompt-conditioned alternatives. Likewise, a stable response profile does not prove health, truthfulness, superior intelligence, or absence of an internal state. The concept constrains inference from a test surface; it does not settle the wider ontology of the system. ## Luxembourg-preprint context [[thoughts/The Luxembourg Preprint|The Luxembourg Preprint]] uses the term for the contrast between low-symptom questionnaire responses and expansive trauma-script emulation under open-ended clinical probing. Bryant's interpretation is that the important signal is not a contest among model personalities but the degree to which each behavioral surface changes under introspective load. [[wiki/Trauma Simulation|Trauma Simulation]] owns the generated-trauma distinction. [[wiki/Large Language Models|Large Language Models]] provides the model-class context, and [[wiki/Alignment Problem|Alignment Problem]] carries the question of whether constraint interpretation remains coherent outside the conditions in which behavior was trained. ## Relationships [[wiki/Prompt-Sensitive Behavioral Stability|Prompt-Sensitive Behavioral Stability]] · [[wiki/Psychometrics|Psychometrics]] · [[wiki/AI Benchmarking|AI Benchmarking]] · [[wiki/Trauma Simulation|Trauma Simulation]] · [[wiki/Large Language Models|Large Language Models]] · [[wiki/Alignment Problem|Alignment Problem]] · [[wiki/Epistemic Status Separation|Epistemic Status Separation]] ## Sources / Provenance - [[thoughts/The Luxembourg Preprint|The Luxembourg Preprint]] — originating use and primary local source. - [[articles/Host-Indexed Autonomy|Host-Indexed Autonomy]] — expanded use of the term and its alignment interpretation. No external preprint or social-media source was independently reviewed during creation of this node. Claims about the study remain attributed to Bryant's local thought and article.