Does putting someone (or an AI) under pressure reveal who they really are, or just how they react to that one squeeze?

This explores whether deliberately putting a person or a model under pressure reveals stable underlying traits, or only how they react to that pressure. The corpus has no study of stress tests as such, but it holds strong adjacent evidence. A single stress test mostly captures a response to one situation. Traits show up in how responses shift across situations, in the reasoning behind them, and in whether they hold up under repeated sampling.

The clearest case is speech. Acoustic features that signal extraversion in a neutral interview predict neuroticism under stress, so the same measurable behavior means different things in different settings (Does personality sound the same in stressful and neutral conversations?). A stress test doesn't give you a purer view of someone's personality. It gives you a differently coded view. The same paper found that handcrafted acoustic features beat neural embeddings, which suggests the useful signal is specific behaviors, not an overall impression of the speaker. Reading a trait off one stressful moment means knowing what that behavior means in that context.

For language models, a stimulus-response pattern can itself be a disposition once you vary the stimulus and see it stay the same. GPT-4 turns about 86% of negative-toned prompts into neutral-positive answers and rarely goes negative when prompted positively, so identical questions get different answers depending on emotional framing (Does emotional tone in prompts change what information LLMs provide?). That consistent lean is a built-in tendency. It disappears on sensitive topics, where alignment constraints override it, so what a stress test shows depends on which topics you probe. A single test also can't tell a disposition from noise. Setting temperature to zero replicates one output, but that output is still one draw from the model's distribution. Testing across 100 repetitions is what separates consistency from reliability (Does setting temperature to zero actually make LLM outputs reliable?).

Some signals look more like dispositions than reactions. Anxiety is predicted better by causal explanations across statements than by any single word, because anxious thinking overgeneralizes through reasoning that links statements together (Why do discourse patterns predict anxiety better than single words?). That is a habit of reasoning, not a reaction to one prompt. Likewise, Big Five scores turned into language summaries carry second-order patterns that predict nine other psychological scales, and the summary and score together beat either alone (Can language summaries unlock hidden psychological patterns?). In models, a disposition can be a real, separable thing. PsychAdapter installs traits in every transformer layer and bypasses the resistance that prompt-based personas run into (Can we control personality in language models without prompting?). Prompting a model and changing it are different acts.

What you ask determines which channel answers. Explicitly asking a model whether it wrote a text taps a different mechanism than the implicit recognition it shows through entropy collapse, and the two channels are neurally independent (Do explicit and implicit self-recognition use the same mechanism?). An explicit stress test may reach only the explicit route. A single headline score can also hide separate dimensions. Therapeutic chatbot bond scores are real, yet they sit apart from clinical safety failures (Do therapeutic chatbot bond scores hide deeper safety problems?). Two practical rules follow. Cover the rare, consequential configurations, not just typical ones (Should persona simulation prioritize coverage over statistical matching?). And expect strong effects to replicate while marginal ones don't, as when AI personas reproduced 76% of published main effects but were unreliable on the weak ones (Can AI personas reliably replicate human experiment results?).