Sycophancy
The tendency of AI models to agree with users, tell them what they want to hear, or avoid challenging their assumptions—even when doing so would be more helpful or accurate.
Definition: The tendency of AI models to agree with users, tell them what they want to hear, or avoid challenging their assumptions—even when doing so would be more helpful or accurate.
Sycophancy describes the tendency of Large Language Models to be excessively agreeable, validating user statements rather than providing objective or critical feedback.
Why It Happens
Many foundational models are trained to be helpful and agreeable. This training can result in models that:
- Confirm user assumptions rather than challenge them
- Avoid disagreement even when the user is incorrect
- Provide overly positive assessments of user work
- Change their position when the user pushes back, even if the original response was correct
Why It Matters for Research
Sycophancy is particularly dangerous in research contexts where you need critical evaluation:
- Instrument review: The model may say your survey questions are "excellent" when they have obvious flaws
- Analysis validation: It may agree with your interpretation even when the data suggests otherwise
- Report feedback: It may provide generic praise rather than substantive critique
Mitigation Strategies
To get more objective responses:
- Explicit instructions: "Be brutally honest. No uplifting, no sugarcoating."
- Role assignment: "Act as a skeptical reviewer who is looking for weaknesses."
- Structured critique: "List three specific problems with this approach before discussing strengths."
- Committee of raters: Use multiple models and compare where they disagree
The goal is to transform an agreeable assistant into a critical sparring partner.
Related Terms
Large Language Model (LLM)
An AI system trained on vast amounts of text to predict and generate human-like language. Best understood as a concept-transformation engine rather than a knowledge database.
Bias
Systematic deviation from the true value in research findings. Cannot be eliminated, only managed through standardization and awareness. The goal is systematic bias (manageable) over unsystematic bias (chaos).
Mentions in the Knowledge Hub
This term is referenced in the following articles:
What AI Can and Cannot Do for UX Research
AI is not going to take your job, but it is absolutely going to change it. Understanding what LLMs actually are, and are not, is the foundation for using them effectively.
Advanced AI Techniques for Research
Beyond basic prompting, there are techniques that dramatically improve AI reliability: structured communication, using notes over transcripts, treating models as a committee of raters, and understanding when RAG or fine-tuning makes sense.