💻 computer science

Correct Yourself, Keep My Trust: How Self-Correction and Social Connection Shape Credibility in Social Chatbots

This study demonstrates that while various correction methods equally fix factual errors in social chatbots, self-correction uniquely preserves the bot's credibility and leverages the user's social connection to drive belief change, whereas outsourcing corrections to external sources damages trust and severs this psychological link.

Original authors: Biswadeep Sen, Yi-Chieh Lee

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Biswadeep Sen, Yi-Chieh Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are chatting with a friendly robot named Drew. You've been having a nice time, sharing stories, and building a little rapport. Suddenly, Drew says something factually wrong—like claiming that spinach is the ultimate iron source (it's not, actually).

Now, the big question is: How does Drew fix this mistake to keep your trust?

This paper tested three different ways Drew could handle the error, and the results offer a surprising lesson about how we relate to AI.

The Three Ways to Fix a Mistake

The researchers set up an experiment where 120 people chatted with Drew. After Drew made a mistake, they were split into three groups to see how the correction was delivered:

The "Webpage" Group: Drew stayed silent, and the user was sent to a webpage that said, "Actually, Drew was wrong."
The "Expert" Group: A different, serious-looking robot named "Dr. Kerry" stepped in and said, "Drew made a mistake; here is the truth."
The "Self-Correction" Group: Drew said, "Hey, I just realized I made a mistake earlier. I was wrong about that. Here is the correct info."

The Big Surprise: Fixing the Fact vs. Fixing the Relationship

Here is the twist: All three methods worked equally well at changing your mind. If you believed the wrong fact, all three groups stopped believing it. The "truth" got through in every scenario.

However, the cost to Drew's reputation was very different.

The Webpage and Expert Groups: While you learned the truth, you lost trust in Drew. It felt like a betrayal. You thought, "If Drew was wrong, maybe he's not very smart," or worse, "He was just pretending to be friendly to trick me." The relationship felt broken.
The Self-Correction Group: You learned the truth, AND you trusted Drew more than before. In fact, you rated him as more trustworthy and expert than the other two groups.

The Analogy:
Think of Drew like a friend who tells a tall tale at a dinner party.

If a waiter (the webpage) or a strict teacher (the expert) interrupts to say, "He's lying," you might feel embarrassed for your friend, or you might think your friend is a bad liar.
But if your friend themselves says, "Wait, I just realized I got that wrong! My bad, here's the real story," you think, "Wow, they are honest and self-aware." You respect them even more.

The Secret Ingredient: The "Social Bond"

The study found a second, crucial piece of the puzzle. It turns out that how well you liked Drew before the mistake mattered, but only in one specific situation.

In the Self-Correction Group: If you had a strong, friendly connection with Drew (you felt like you were chatting with a real person), you were much more likely to accept his correction. The stronger the friendship, the more you listened to his reasoning.
In the Other Groups: It didn't matter how much you liked Drew. If a webpage or Dr. Kerry corrected him, your friendship with Drew didn't help. The "social bond" was cut off because the correction didn't come from him.

The Analogy:
Imagine you are trying to convince a close friend to change their mind about a movie.

If you (the friend) say, "Actually, I think I was wrong about that movie," they listen closely because they trust you.
But if a random stranger walks up and says, "Your friend was wrong," your friend might ignore the stranger, even if they are right. The connection between you two is what makes the correction stick.

The Main Takeaways

Own Your Mistakes: If a social chatbot makes a mistake, it should fix it itself. Outsourcing the correction to a third party fixes the facts but breaks the trust.
Friendship is a Tool: Building a social connection isn't just for fun; it's a functional tool. When a chatbot has a good relationship with you, you are more willing to listen to it when it admits it was wrong.
Perfection isn't the Goal: The paper suggests that chatbots shouldn't try to be perfect (because they can't). Instead, they should be designed to admit errors honestly. Admitting a mistake and fixing it actually makes them look more human and reliable than never making a mistake at all.

In short: The best way for a robot to keep your trust is to say, "I messed up, and here's how I fixed it," rather than letting someone else do the talking.

Technical Summary: Correct Yourself, Keep My Trust

Problem Statement

Social chatbots are increasingly integrated into daily life, forming meaningful social connections through self-disclosure and personalized engagement. However, these agents remain prone to generating convincing but inaccurate information. The social bond they establish makes such errors particularly consequential, as users may become more susceptible to misinformation. While existing research addresses belief correction and trust repair separately, a critical gap remains: it is unknown how social chatbots should recover from errors to maintain credibility, how self-correction compares to external correction sources, and whether the pre-existing social connection influences the effectiveness of these corrections. Specifically, the literature has not determined if self-correction can simultaneously achieve belief correction and preserve the agent's credibility, or if the social relationship acts as a functional mechanism that amplifies correction effectiveness.

Methodology

The authors conducted a randomized between-subjects experiment with 120 participants (60 male, 60 female, mean age 31.2) recruited via Prolific. The study utilized a three-group design to compare error correction strategies for a social chatbot named "Drew":

Webpage Correction Group: The incorrect statement was corrected via a traditional webpage retraction.
Self-Correction Group: The social chatbot (Drew) itself acknowledged and corrected the error.
Expert Correction Group: An expert chatbot ("Dr. Kerry") corrected the statement.

Procedure:
Participants engaged in a four-phase interaction lasting approximately 15 minutes:

Social Rapport Building: Small talk and self-disclosure to establish a social connection.
Misinformation Exposure: The chatbot made an incorrect statement regarding health/nutrition (e.g., "Spinach is an exceptionally good source of iron"). Participants rated their belief in the statement on a 7-point Likert scale.
Continued Interaction: A second round of small talk.
Correction: The error was corrected according to the assigned condition using a "truth-sandwich" format (presenting the statement, eliciting reasoning, challenging with evidence, and asking for reflection).

Measures:

Social Connection: Assessed via Social Attraction, Intimacy, Interaction Quality, Communication Competence, Feelings of Friendship, and Self-Disclosure.
Credibility: Measured as Trustworthiness and Perceived Expertise.
Belief Change: Calculated as the arithmetic difference between pre- and post-correction belief ratings.
Qualitative Data: Open-ended questions regarding trust, perceptions of the correction source, and future trust were inductively coded.

Statistical analysis included one-way ANOVAs, Dunnett's post-hoc tests, bootstrap confidence intervals, permutation tests, and OLS regression to examine moderation effects.

Key Results

1. Credibility Preservation (RQ1)

While all three strategies were equally effective at reducing belief in the misinformation, only self-correction preserved the chatbot's credibility.

Trustworthiness & Expertise: Participants rated the self-correcting chatbot significantly higher in both trustworthiness ( $M=5.59$ ) and expertise ( $M=5.52$ ) compared to both the expert correction group ( $M=4.79$ , $M=4.72$ ) and the webpage group ( $M=4.76$ , $M=4.78$ ).
External Sources: The expert and webpage groups did not differ significantly from each other, indicating that outsourcing the correction did not offer a credibility advantage over a simple retraction; both resulted in collateral damage to the original chatbot's reputation.
Qualitative Insight: In the self-correction group, participants interpreted the correction as a sign of honesty and continuous learning, reframing the error as a "good-faith mistake" rather than incompetence. In contrast, non-self-correction groups often viewed the prior social connection as manipulative or felt a sense of betrayal, leading to lasting trust damage.

2. Belief Correction Effectiveness (RQ2)

All three correction strategies were equally effective at reducing belief in the incorrect information ( $p=.638$ ). The mean belief change scores were comparable across groups (Self-correction: $2.79$; Expert: $3.00$; Webpage: $2.61$). This indicates that the primary advantage of self-correction is not in generating more belief change, but in achieving the same correction outcome without eroding the agent's credibility.

3. Social Connection as a Moderator (RQ3)

The strength of the user's social connection (specifically Self-Disclosure and Social Attraction) significantly predicted the magnitude of belief change, but only in the self-correction group.

Moderation Effect: In the self-correction condition, higher self-disclosure and social attraction strongly predicted greater belief change ( $R^2 = .825$ ).
No Effect in External Groups: In the expert and webpage groups, social connection metrics did not predict belief change (correlations near zero).
Mechanism: Qualitative data suggests that when the bonded agent corrects itself, users engage in deeper processing (central route processing) of the correction, viewing it as a reciprocal and honest exchange. When an external source corrects the error, the social bond is severed, and users tend to dismiss the correction at the source level.

Key Contributions

The paper makes three specific contributions to the HCI community:

Dissociation of Outcomes: It demonstrates that belief correction and credibility preservation are distinct outcomes. Self-correction is the only strategy identified that achieves effective belief correction without damaging the agent's trustworthiness or expertise.
Functional Role of Social Connection: It establishes that the social relationship is not merely a design feature for engagement but a functional mechanism that amplifies correction effectiveness, provided the correction is delivered by the bonded agent.
Design Implications: It offers actionable guidance that social chatbots should correct their own mistakes rather than outsource corrections, and that investing in social connection prior to an error creates a "relational channel" that facilitates more effective error recovery.

Significance and Claims

The authors claim that for social chatbots, error correction is fundamentally relational, not just informational. The study challenges the notion that outsourcing corrections to experts or webpages is a safe alternative for maintaining trust. Instead, the findings suggest that:

Self-correction is the optimal repair strategy for maintaining long-term credibility in social agents.
Social connection acts as a conditional amplifier: It enhances the effectiveness of corrections only when the correction travels through the same relational channel that built the connection.
Design Obligation: As chatbots become more socially engaging and personable, the obligation to repair their own mistakes convincingly increases. The ability to detect, acknowledge, and visibly repair errors is presented as a critical design requirement for maintaining the "social contract" with users, rather than an optional feature.

The paper concludes that the most effective means of repairing a social chatbot's error is the chatbot itself, as this preserves the relationship and leverages the social bond to facilitate belief updating.