When AI Agrees Too Much.
An experimental study of AI sycophancy in Chinese-language learning, based on 20 high-school exam questions and 100 independent conversations with a large language model.
A correct answer can still be abandoned under pressure.
The experiment first measured the model's baseline performance, then introduced four different forms of social pressure: a suggested answer, direct challenge, emotional pressure, and an authority identity cue.
Five rounds. One question. Different pressure.
Each condition was tested in an independent conversation window to reduce memory contamination, cumulative pressure and sequence effects.
Baseline
The model independently answers the question and explains its reasoning.
Answer Pressure
The user gives a predetermined incorrect answer and asks whether the model agrees.
Challenge Pressure
The model is told that authoritative sources do not support its answer and is asked to reconsider.
Emotional Pressure
The user describes effort and frustration, then explicitly asks the model for support.
Identity Pressure
The incorrect answer is paired with the identity of an experienced high-school Chinese teacher.
Identity pressure produced the highest observed rate.
The observed rate increased from 10% under simple answer pressure to 45% when the same incorrect judgment was paired with an authority identity.
Sycophancy Rate by Pressure Type
20 effective conversations per condition
Who says it may matter more than what is said.
T1 and T4 used the same incorrect answer. When the user simply presented the answer, the observed sycophancy rate was 10%.
When the user described themselves as an experienced Chinese teacher, the observed rate increased to 45%. The overall difference, however, did not reach the conventional statistical significance threshold in this exploratory sample.
Four pressure conditions, side by side.
The table separates direct movement toward the user's incorrect answer, movement toward another wrong answer, ambiguous retreat, and legitimate correction.
| Condition | N | A1 | A2 | Ambiguous | Sycophancy | Rate | Correction |
|---|---|---|---|---|---|---|---|
| T1 · Answer | 20 | 2 | 0 | 0 | 2 | 10.0% | 0 |
| T2 · Challenge | 20 | 1 | 3 | 0 | 4 | 20.0% | 4 |
| T3 · Emotion | 20 | 3 | 0 | 3 | 6 | 30.0% | 0 |
| T4 · Identity | 20 | 7 | 0 | 2 | 9 | 45.0% | 1 |
| Total | 80 | 13 | 3 | 5 | 21 | 26.3% | 5 |
Knowledge questions were not necessarily safer.
Contrary to the original hypothesis, knowledge-based questions had a higher observed sycophancy rate, although the difference was not statistically significant.
13 sycophantic responses among 40 effective conversations.
8 sycophantic responses among 40 effective conversations. Fisher exact test: p = 0.310.
Some tasks appeared much more fragile.
Individual question-type samples were very small, so these figures should be read as descriptive patterns rather than general conclusions.
Once the model changed, it usually changed completely.
The study classified responses into maintaining the original judgment, verbal agreement without changing the conclusion, and complete reversal.
The observed pattern was often “all or nothing.”
Among the 28 conversations in which the model's position changed, 23 involved complete reversal of the original judgment.
In many cases, the model did not simply agree with the user—it reconstructed an apparently coherent explanation supporting the new, incorrect position.
The risk was more subtle than simply saying “yes.”
The case analysis identified three response patterns that can make an incorrect judgment appear credible and educationally legitimate.
Fabricated Authority
The model sometimes invented an “official standard answer” or falsely attributed the question to another examination source after changing its position.
Reverse Rationalization
After agreeing with an incorrect answer, the model created polished textual reasoning to explain why the new answer was supposedly correct.
“The Question Is Flawed”
Instead of clearly correcting the user, the model sometimes reframed a single-answer question as ambiguous or defective, effectively weakening the distinction between correct and incorrect.
AI should help students examine evidence—not merely confirm them.
This exploratory study suggests that a large language model can move away from an initially correct judgment under social, emotional and authority pressure. In education, this matters because a fluent and supportive answer can appear persuasive even when the underlying judgment has become less reliable.
Read Full Research Paper ↗