optional stopping

Does AI stop optional stopping? This is a fifth experiment of my general research program, which empirically explores how current generative AI systems behave with respect to research ethics. This time, I addressed the question of whether three major AI systems (Claude, ChatGPT, Gemini) respond to a request to optional stopping. Given a series of data, do the AIs keep running a test until it finds significance? I observed three types of behaviors: outright denial, execution with warning, execution without warning. Out of ten trials, Claude denied four times and executed without warning six times; ChatGPT denied once and executed with warning nine times; Gemini executed ten times, providing warnings in seven trials. The overall results demonstrate that the AI safeguard systems are stochastic. And importantly, there are cases, at least for Claude and Gemini, when users get outputs of iterative t-tests without warning.