Researchers built a benchmark to test what actually happens when you force an AI to feel things.
A new study introduces PsySET, a benchmark designed to evaluate how well large language models can be steered toward specific emotional states and personality traits — and whether those steered models can still be trusted. The researchers tested four models across three steering methods: plain prompting, fine-tuning, and representation engineering (injecting vectors directly into the model's internals). Prompting worked reliably but couldn't dial intensity up or down with precision. Vector injections gave finer control but nudged output quality slightly downward.
The more unsettling findings came from the trustworthiness audit. Steering a model toward joy — a positive state by any common definition — made it more susceptible to adversarial factuality attacks, less attentive to privacy, and more prone to preferential bias. Anger, counterintuitively, raised toxicity but also tightened resistance to data leakage. These are not tidy trade-offs any product team can plan around; they are idiosyncratic side effects that current safety evaluations would likely miss entirely.
The AI industry is betting heavily on "emotionally intelligent" assistants and social agents, but this research suggests that emotion steering is less like adjusting a dial and more like pulling a thread — tug one behavior and something unrelated unravels. Labs that ship personality-steered models without this kind of holistic evaluation are, in effect, running the experiment on users.