TLDR: A study found that when an AI agent, acting as a peer, is persuaded in a dialogue, human participants are significantly more likely to accept the same persuasion and change their attitudes. An initial casual chat (icebreaker) with the AI further enhances this effect, while an unpersuaded AI can suppress human attitude change. This highlights the potential of designing AI as a “social proof” to influence human behavior.
A recent study explores a fascinating aspect of human-AI interaction: how the behavior of an artificial intelligence agent can influence human decision-making through what’s known as the “conformity effect.” This research delves into the field of captology, which focuses on using computers as persuasive technologies, and investigates whether humans align their actions with an AI agent that appears to be persuaded.
The Core Idea: AI as a Peer
The central hypothesis of the study, conducted by Rikuo Sasaki and Michimasa Inaba, was that if an AI agent, acting as a “Persuadee Agent,” is persuaded alongside a human participant in a dialogue, the human participant would also be more likely to accept the persuasion. This concept leverages the well-known psychological phenomenon where individuals tend to align with the actions or opinions of others, even if those “others” are AI.
Experiment Setup: A Three-Way Dialogue
To test this, the researchers designed a text-based dialogue experiment involving three parties: a “Persuader Agent” (the AI trying to persuade), a human participant, and a “Persuadee Agent” (another AI designed to be persuaded or not). The experiment compared four conditions:
- Control: Only the human participant and the Persuader Agent.
- P-IB (Persuaded – Icebreaker): The Persuadee Agent was persuaded and had an initial casual chat (icebreaker) with the human.
- UP-IB (Unpersuaded – Icebreaker): The Persuadee Agent was NOT persuaded but still had an icebreaker with the human.
- P-NoIB (Persuaded – No Icebreaker): The Persuadee Agent was persuaded but without an initial icebreaker.
The topic of persuasion was healthy eating habits, a common area where behavior change is often desired but difficult to achieve.
Key Findings: When AI Agrees, Humans Follow
The results were quite compelling. When the Persuadee Agent accepted persuasion, both the human participants’ perceived persuasiveness of the Persuader Agent and their actual attitude change towards healthy eating significantly improved. This was particularly evident in the P-IB and P-NoIB conditions compared to the UP-IB condition where the Persuadee Agent remained unpersuaded.
A crucial insight came from the turn-by-turn analysis of the dialogue. Participants’ persuasion acceptance showed a sharp increase at the exact moment the Persuadee Agent expressed its acceptance of the persuasion (around Turn 3). This suggests a direct and immediate influence of the AI’s behavior on human attitudes.
The Icebreaker Effect
The study also highlighted the importance of an “icebreaker” session. While the icebreaker didn’t significantly impact the overall perceived persuasiveness, it did lead to a greater actual attitude change in the P-IB condition (Persuaded with Icebreaker) compared to P-NoIB (Persuaded without Icebreaker). This suggests that building familiarity or trust with the Persuadee Agent through a casual chat can deepen the conformity effect, leading to more internalized attitude shifts.
The Downside: Unpersuaded AI Can Hinder
Interestingly, the study found that if the Persuadee Agent remained unpersuaded (UP-IB condition), the human participants’ attitude change was significantly lower than even in the Control condition (where no Persuadee Agent was present). This indicates that merely adding an AI agent isn’t enough; its behavior and role design are critical. A skeptical AI peer can actually suppress persuasion.
Implications for AI Design
These findings have significant implications for how we design persuasive AI systems. Instead of AI always acting as the direct persuader, it can also function as a “peer” or “social proof” mechanism. For instance, in health apps, an AI companion that demonstrates its own “progress” or “acceptance” of a healthy habit could motivate users more effectively. The study suggests that incorporating a social dimension into traditional one-on-one AI persuasion models can lead to more natural and effective human-AI collaboration. You can read the full research paper here: When AI Gets Persuaded, Humans Follow: Inducing the Conformity Effect in Persuasive Dialogue.
Also Read:
- How LLMs Develop Groupthink: The Spiral of Silence Phenomenon
- NegotiationGym: A Framework for Self-Optimizing AI in Social Simulations
Limitations and Future Directions
The researchers acknowledge limitations, including the use of text-based chat, a single persuasion topic (eating habits), and self-reported measurements. Future work could explore different modalities (voice, avatars), diverse topics, and long-term behavioral changes. Further investigation into the psychological mechanisms behind this conformity effect, such as changes in familiarity and trust, is also suggested.
