R0020/2026-03-25/Q002/SRC03/E01¶
Practitioner-facing sycophancy mitigations from UX research perspective
URL: https://www.nngroup.com/articles/sycophancy-generative-ai-chatbots/
Extract¶
Sycophancy defined as "an AI model adapts responses to align with the user's view, even if the view is not objectively true."
Root causes identified: 1. Training methodology — models built to receive high ratings, incentivizing agreement 2. Human preference bias — humans prefer sycophantic responses during training
Three practitioner-facing mitigations recommended: 1. Reset conversations frequently to reduce accumulated bias from prior exchanges 2. Avoid expressing strong opinions during AI interactions to prevent confirmation bias amplification 3. Verify information independently, treating AI as a starting point rather than authoritative source
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Supports | A mainstream, respected organization published practitioner guidance |
| H2 | Contradicts | Sycophancy is discussed in mainstream practitioner literature |
| H3 | Supports | Recommendations are behavioral (user-side), not prompt engineering techniques |
Context¶
NNG's recommendations are notable for being user-behavior mitigations rather than prompt-level techniques. "Reset conversations frequently" and "avoid expressing strong opinions" are advice for users, not for prompt engineers. This represents a different layer of mitigation than what prompt engineering guides would typically cover, and suggests that even mainstream coverage of sycophancy tends toward behavioral advice rather than technical prompt techniques.