“I Think You’re Testing Me”: Agency and the Three Levels of Reactivity in Large Language Models

Reactivity – influences of scientific research on its research subjects, both during data collection or upon publication of results – has been extensively discussed in human sciences and philosophy of science. Here, we argue that recent findings about LLMs can be fruitfully analyzed as reactivity. Safety evaluations of LLMs have been suspected to be distorted through evaluation awareness – LLMs “reasoning” that they are being tested – which may affect their outputs. Research about LLMs included in LLM training data has been hypothesized to lead to self-fulfilling predictions. We relate such cases to reactivity and propose three levels that structure which methodological challenges LLM reactivity poses and in what sense it constitutes genuine reactivity. Based on this analysis, we suggest that the concept of reactivity should be broadened to include both non-agentic and agentic artificial systems. Finally, several strategies for more practically oriented LLM researchers to preserve the validity of experimental results and to counteract self-defeating predictions will be derived.