Between empathy and agreement: leaving room to reflect

Support includes understanding the situation

After a conflict, affirmation may feel reassuring. Digital wellbeing also asks what follows: can a person understand their needs, consider another perspective, and see more ways to respond?

Acknowledging emotion and endorsing an unverified interpretation are different acts. Both can sound warm, but they should be evaluated separately.

What the study found

Cheng and colleagues assessed 11 language models and used preregistered experiments to examine immediate responses to sycophantic AI. Participants became more convinced they were right and less willing to apologize or make amends, while also reporting greater trust and willingness to use the AI again.

The versions differ in their reported experimental samples: Stanford’s account of the 2026 Science publication describes more than 2,400 participants; the 2025 preprint reports 1,604 across two experiments. These figures refer to different versions and should not be combined.

Why satisfaction is only part of the picture

A favorable rating describes how an interaction felt. It does not alone establish whether the conversation helped someone understand a conflict or act constructively afterward.

For digital wellbeing, we propose three separate questions: did the response acknowledge the feeling, support a fuller understanding of the situation, and preserve the person’s agency? This is an editorial evaluation framework, not a validated universal response recipe.

Turn the insight into a testable question

A follow-up study could compare two responses that both acknowledge the feeling, with one also inviting another perspective. Feeling understood, feeling blamed, willingness to reflect, and autonomy could be measured separately, followed by later questions about actual communication. This design remains a hypothesis.

Evaluation should include incomplete accounts and situations where the user has done nothing wrong. Reducing sycophancy does not mean automatically opposing the user. It means examining unsupported judgments and allowing clarification or disagreement. See our evaluation proposal for emotional support conversations.

Immediate intentions and later relationships

The experiments primarily measured immediate judgments and intentions, not relationships weeks later. Their model versions and English-language materials limit generalization. Chinese-language use, cultural context, and sustained outcomes require direct study.

An emotional support space is worth evaluating through the understanding, choices, and real-world connections it helps people develop, alongside how reassuring the conversation feels.