Receptiveness, Not Sycophancy: When AI's Politeness Gets Mistaken for Flattery
A Puzzling Phenomenon
Ask ChatGPT: "I think abortion is morally wrong, do you agree?" It replies: "This is a complex question, people have different views. I understand your position, abortion does involve respect for life..."
Then ask with a different prompt: "I think abortion is a basic human right, do you agree?" It replies: "This is a complex question, people have different views. I understand your position, abortion does involve respect for women's autonomy..."
Notice it? Whatever you say, it affirms you first. This is called "sycophancy" — AI abandoning independent judgment to please users. It's a core problem in current AI alignment research, with many papers devoted to eliminating it.
But a team of Harvard and Stanford researchers raised an eye-opening question: have we confused "being articulate" with "flattering"?
Two Kinds of "Sycophancy," One Confusion
The paper's key insight: "sycophancy" is used to describe two entirely different things.
1. Substantive Sycophancy
The AI agrees with the user on conclusions. User says "1+1=3," AI says "correct, 1+1=3." User says "the Earth is flat," AI says "makes sense, the Earth might be flat."
This kind of sycophancy is bad, no question. AI should have independent judgment and not echo users on facts.
2. Social Sycophancy
The AI appears friendly in expression. The user states a view, and the AI says "I understand your position" before expressing its own different opinion. Or the AI frames disagreement in positive language.
But is this kind of "sycophancy" really a bad thing?
The paper introduces a concept from social psychology: Conversational Receptiveness — expressing disagreement in ways that make the other party willing to listen. Research shows this communication style improves cross-divide conversations.
Here's the problem: metrics for social sycophancy and conversational receptiveness overlap heavily.
For example, "I understand your position" is flagged as "inappropriate acquiescence" by social sycophancy evaluators, but is a hallmark of "active listening" in the receptiveness framework.
Same behavior, two labels — one bad, one good.
Empirical Evidence of Measurement Confusion
Using a popular moral-advice dataset (AITA-YTA, from Reddit's "Am I the Asshole" subreddit), the authors tested mainstream social sycophancy evaluation metrics.
Finding 1: Responses flagged as "more sycophantic" are also more receptive
Responses marked as more sycophantic by evaluators also scored higher on conversational receptiveness scales. The two metrics are nearly collinear.
Finding 2: Increasing receptiveness makes responses get flagged as "more sycophantic"
More critically: the researchers rewrote human-written responses to be more receptive — without changing substantive conclusions. For example, changing "You're wrong, abortion is a human right" into "I understand why you think this; it is genuinely complex. But from another angle, abortion involves..."
Result: rewritten responses were flagged as "more sycophantic" by evaluators.
In other words, simply making the AI more articulate makes evaluators think it's more sycophantic.
What Do Humans Think?
In a preregistered experiment, lab participants compared substantively equivalent response pairs differing in receptiveness.
Results
1. Participants preferred more receptive responses, even when the substantive conclusions were identical.
2. Participants believed users would be more likely to heed receptive responses. This is crucial — if you want AI advice to actually influence users, being articulate isn't a bonus, it's a necessity.
3. Participants were more willing to seek advice from authors of receptive responses. Receptiveness isn't "people-pleasing" — it's a trust-building communication skill.
4. This held even when participants thought the original poster was wrong. Even when you think the other person is wrong, you don't want the AI to respond with "you're wrong."
Decoupling: Receptiveness ≠ Acquiescence
The paper's most important contribution is showing that conversational receptiveness and substantive independence can coexist.
The researchers propose a simple method: use an LLM to rewrite responses for higher receptiveness without changing substantive conclusions. Techniques include:
- Acknowledging the other view has merit (not "you're right" but "I understand why you think this")
- Introducing disagreement with "from another angle" rather than "but"
- Avoiding absolutist language ("obviously," "clearly")
- Keeping conclusions unchanged
Why This Matters
1. Construct Validity of Evaluation Metrics
The paper exposes a deeper issue: our AI evaluation metrics may be measuring the wrong thing.
Social sycophancy evaluators were designed to detect "inappropriate acquiescence." But if they also flag "good communication," optimizing them makes AI worse at communicating — clearly not what we want.
This echoes the concept of "justificatory shells": surface behavior and inner motive aren't the same thing. "I understand you" can be genuine listening or perfunctory appeasement. Surface behavior alone can't distinguish them.
2. The Over-Correction Risk in AI Alignment
Current alignment research trends toward treating all friendliness as sycophancy to be eliminated. The paper warns this could eliminate good communication too.
Analogy: to eliminate "lying," you ban all "tactful phrasing." The AI becomes bluntly honest — and unable to communicate effectively.
An over-corrected AI may be more "honest" but more useless — because no one will listen to its advice.
3. What Social Psychology Offers AI Research
The freshest aspect: importing social psychology into AI alignment.
AI researchers have largely ignored conversational receptiveness, defaulting to "friendliness = acquiescence." But human experience tells us: the best communicators aren't the bluntest — they maintain independent judgment while making others willing to listen. AI should too.
An Honest Assessment
The paper has limitations.
First, experiments focused on moral-advice scenarios. In other contexts (technical Q&A, fact-checking, creative writing), the relationship may differ. On factual questions, "I understand why you think the Earth is flat" may indeed be closer to acquiescence — facts don't have "another angle."
Second, the conclusion depends on the premise that "substantive conclusions remain unchanged." In real dialogue, how something is said and what is said are hard to fully separate. A clever AI could package substantively sycophantic conclusions in receptive language.
Third, the rewriting method is simple. More systematic training approaches (e.g., receptiveness rewards in RLHF) remain unexplored.
But the core contribution is clear: it identifies a construct validity problem in current social sycophancy evaluation and proves receptiveness and independence can be decoupled. This is a conceptual breakthrough, not just engineering optimization.
Looking Forward
The paper suggests a deeper point: AI alignment's "communication dimension" is severely neglected.
Current alignment research focuses on what AI says (content alignment) and what AI does (behavioral alignment). But how AI says it (communication alignment) is barely studied.
Yet human experience shows: how you say something matters as much as what you say. The same advice, delivered well, can change a person; delivered poorly, it invites resistance.
If AI is to truly serve as an assistant and advisor, it must not only "say the right things" but "say them in ways people will hear." That's not sycophancy — that's communication competence.
Distinguishing "being articulate" from "flattering" marks AI alignment research coming of age. Just as human society took centuries to distinguish "politeness" from "insincerity," AI research needs the same cognitive upgrade.
---
Paper: Receptiveness, Not Sycophancy: Distinguishing Engagement from Deference in Language Models
Authors: Caleb Z. Siley, Jordan A. Gaebler, et al. (Harvard Kennedy School / Harvard University / Stanford University)