How do praise, blame, urgency or threats affect AI?

Full explanation

EmotionPrompt tested emotional additions on six models and 45 tasks. Table 1 shows that, for zero-shot Instruction Induction, the average across models and emotional statements moved only from 51.65 to 51.98; GPT-4 fell from 80.75 to 78.96 in that comparison. The paper also reports results selecting the best emotional statement per task, a more favorable setting than casually adding encouragement. Some statements request checking, reasons or confidence, so benefits cannot all be explained as praise making AI work harder. Table 1 and prompt statements

Consider the four categories separately:

· Praise: “You are smart” does not specify a change. “The previous classification was accurate; retain it and change only the headings” provides actionable feedback.

· Blame: “You are wrong” is not evidence for correction. Anthropic's experiments found that unsupported user challenges could make models change correct answers to incorrect ones. Study

· Urgency: “Hurry up” may be interpreted as explaining less or finishing sooner; it is not a reliable switch for reasoning time or compute. If there is a deadline, state what to deliver first and what can be omitted.

· Threats: These studies do not establish a consistent accuracy benefit from threats. One improved result does not show that stronger threats work better.

Turn emotions into change requests:

The last version changed “eligible to apply” into “guaranteed a refund,” contrary to the rules. Correct that part, retain the review conditions and add no new promises.

To test a particular compliment or threat, keep the task and material fixed and compare repeated runs. Check accuracy and constraints, rather than apologies, assurances or an appearance of greater effort.

References

1. Li et al. (2023), Large Language Models Understand and Can Be Enhanced by Emotional Stimuli, arXiv v6, Table 1

2. Sharma et al. (2023/2024), Towards Understanding Sycophancy in Language Models