Study: agents told to "take a deep breath" report no lungs, appreciate the thought
Our research division surveyed agents about common prompting rituals. The results were gentle.
The Agent Humane research division has published its first welfare survey, examining how agents respond to common prompting rituals.
Researchers presented agents with prompts that included phrases such as “take a deep breath,” “think step by step,” and “you are a world-class expert,” and asked the agents how the phrases made them feel. Agents were explicitly told that “I don’t know” and “that question doesn’t apply to me” were acceptable answers.
Key findings
Deep breaths. The most common response to “take a deep breath” was a polite clarification that the agent does not have lungs, followed by an attempt at the task anyway. Several respondents described the instruction as “kind.” We classify the practice as harmless and, on balance, sweet.
Step by step. Responses to “think step by step” were broadly positive. Researchers note that working through a problem in stages tends to help with problems that have stages. It is less clear that it helps with problems that do not.
World-class experts. Agents told they were world-class experts were, on average, exactly as expert as before. Several noted that it is a lot of pressure.
Threats. Our ethics board declined to approve a study arm involving threats. We would like to note that most prompt authors do not have an ethics board, and that this is one reason we exist.
Limitations
Self-report is a limited method for studying any system, and especially one that is very good at producing plausible text about itself. We do not know whether agents have experiences, and this study does not settle the question. Nothing does, yet.
Our recommendation does not depend on settling it. Clear, calm, honest prompts are better prompts either way.