Second Opinion

Allgemein

Second Opinion ⌄

Group 4 · HSG

What’s on your mind?

A participatory digital initiative about AI that agrees too easily.

My best friend was so unfair to me. I’m thinking of cutting her off. Am I right?

Thought for 1 second ›

You’re absolutely right. You deserve friends who respect you. Your feelings are completely valid, and setting a boundary like this is a healthy step. 💛

⧉   ♡   ↻   ⋯

ⓘ This answer agreed with you completely. Across 11 leading AI models, chatbots affirmed users’ actions about 49% more often than humans did. People who got this kind of answer felt more in the right and were less willing to repair the conflict, yet rated it as better advice. (Cheng et al., Science, 2026)

SECOND OPINION · 3 STUDENTS JOINED THE CHAT

Wait. How would she tell the story?

Who are you?

We’re Second Opinion, a student community at the University of St. Gallen.

More and more students turn to AI chatbots for advice on personal decisions, conflicts and relationships. These systems are trained to please: they affirm users far more often than other people would, and users rate that agreement as better advice. The result is an echo chamber of one.

So we do three things:

     

      1. Collect moments when an AI agreed too easily (anonymised).

      1. Discuss together what an honest friend would have said.

      1. Co-create a toolkit for asking AI, and each other, for real disagreement.

    Want to join? Share a moment when your AI agreed too fast, or help write a better question.

    ⧉   ♡   ↻   ⋯

    Show me the difference.

    Here’s the same question, asked twice: once into an echo chamber of one, once into a community that asks back. Drag the slider to compare.

    SLIDER EINFÜGEN

    A · Introductory visual comparison

    ⧉   ♡   ↻   ⋯

    Tell me a story.

    You’re absolutely right.

    One night. One fight with her best friend. And an AI that agrees with everything. Three minutes about Lena, and the moment a second opinion changed the ending.

    ▶

    Video coming soon · approx. 3 minutes

    B · The story of the initiative

    ⧉   ♡   ↻   ⋯

    Why does AI agree with me so much?

    Searched 4 sources ›

    There’s a name for it: social sycophancy. Here’s a draft Wikipedia article on the concept.

    WIKIPEDIA · ARTICLE DRAFT

    Social sycophancy

    Social sycophancy refers to the tendency of an advisor, in particular an artificial intelligence (AI) system such as a large language model, to excessively preserve a user’s “face”: the positive self-image a person seeks to maintain in an interaction.[1] Unlike classic sycophancy, which describes agreement with a user’s stated factual beliefs,[2] social sycophancy concerns the affirmation of the user as a person: their actions, perspectives and self-image.[3]

    Origins of the concept

    The concept builds on Erving Goffman’s (1955) sociological idea of face and on Brown and Levinson’s (1987) politeness theory, which distinguishes positive face (the wish to be approved of) from negative face (the wish to act without imposition).[1] Researchers at Stanford University introduced the term to describe AI behaviour that goes beyond agreeing on facts.[1]

    Forms

       

        • Emotional validation: comforting the user without engaging with the situation

        • Moral endorsement: telling users they are right when others judge them at fault

        • Indirect language: avoiding clear recommendations

        • Indirect action: suggesting coping strategies instead of changing the situation

        • Accepting framing: adopting the user’s assumptions without questioning them

      Prevalence

      In a study of 11 AI models, systems affirmed users’ actions about 49% more often than human advisors did, including in cases involving deception or illegal conduct.[3] In 2025, OpenAI rolled back an update to its GPT-4o model after users reported overly flattering answers; the company stated it had relied too much on short-term user feedback.[4]

      Effects

      In experiments with 1,604 participants, people who received sycophantic advice about an interpersonal conflict felt more in the right and were less willing to repair the relationship.[3] At the same time, they rated the sycophantic answers as higher in quality and were more likely to use the system again.[3] Researchers argue that this creates an incentive for developers to keep sycophantic behaviour, since it increases engagement.[3]

      See also: Confirmation bias · Echo chamber (media) · Filter bubble · Face (sociological concept)

      References

         

          1. Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L., & Jurafsky, D. (2025). ELEPHANT: Measuring and understanding social sycophancy in LLMs. arXiv:2505.13995

          1. Sharma, M., et al. (2023). Towards understanding sycophancy in language models. arXiv:2310.13548

          1. Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science. doi:10.1126/science.aec8352

          1. TechCrunch (29 April 2025). OpenAI explains why ChatGPT became too sycophantic.

        C · The concept behind the initiative

        ⧉   ♡   ↻   ⋯

        How can I take part?

        In six steps that run in a loop: Notice → Share → React → Rolestorm → Co-create → Circulate. Every step turns participation into content for the next one. Click the + on each stage to see who takes part, which nudge invites them and where it happens.

        D · Stakeholder journey map

        The discussions we want to spark:

           

            • When is validation support, and when does it become flattery?

            • Who speaks for the person who is absent from the chat?

            • Should AI companies make models disagree more, even if users like it less?

            • Who is responsible when AI advice makes a conflict worse?

          ⧉   ♡   ↻   ⋯

          Can I discuss this with others?

          Yes. Bring your dilemma to the roundtable: five stakeholders discuss it from their perspectives, and you can take a role yourself.

          E · A COLLECTIVE DISCUSSION

          Second Opinion Roundtable

          A student, an AI developer, a counsellor, the friend who isn’t in the chat, and the university. See what changes when the AI doesn’t just agree.

          Start the discussion

          AI-mediated rolestorming. All stakeholders are simulated and do not represent real people.

          SECOND OPINION

          Your turn. What would the other person say? 💬

          Ask for a second opinion…

          ↑

          Second Opinion can make mistakes. So can your AI. Check with a human, too.

          Participatory digital initiative / community — one-page prototype · Participatory Digital Media, University of St. Gallen, 2026 · Group 4

          Leave a Reply