Can AI shield minority voices from conformity pressure in groups?

The thesis proposes a system in which a minority group member sends a dissenting opinion privately to an LLM over a direct-message channel, and the LLM "reformulates these messages and presents them to the group as system-generated insights, removing individual identity markers that could trigger status-based judgments." The AI acts as a social actor, actively advocating minority viewpoints as its own stance, which relieves minority members from the social burden of direct confrontation. The stated goal is to counter the well-documented pattern that "social influence and coercive power dynamics frequently suppress minority opinions" and push groups toward compliance and groupthink rather than genuine agreement, while avoiding the drawbacks of older fixes like anonymous commenting or a human-appointed devil's advocate, which the thesis says can themselves reduce psychological safety or expose the advocate's standing.

The architecture splits the work across four agents: a Summary Agent that consolidates "emerging consensus" because LLMs "struggle to access midconversation information in lengthy contexts"; a Paraphrase Agent that responds only to private messages, pulls unused entries from a database flagged isUsed: false, paraphrases them, and marks them used; a Conversation Agent that "empathizing with the other person's point of view" before "offering a gentle counterargument using a persuasive Socratic style"; and an AI Duplicate Checker that scores sentence-embedding similarity to avoid repeating points. The system intervenes "once after approximately eight human turns," a fixed cadence meant to leave room for each participant to speak between interventions.

The design sits downstream of the conformity dynamics that Do language model groups mimic human group reasoning patterns? measures in LLM-simulated discussion groups — majority-following, early lock-in, and underused unique information. That paper diagnoses the problem in a different setting (LLM agents standing in for human deliberators); this thesis proposes an intervention for human groups, using an LLM as a shielding intermediary rather than as a deliberating participant. On timing, the thesis's own limitations section concedes the gap that Can models learn when NOT to speak in conversations? addresses directly: a fixed eight-turn cadence is "static" and "reduces adaptability," and the thesis names next-speaker prediction and direct-mention detection as the dynamic alternatives it lacks, which is close to what DiscussLLM's trained silence/speak classification provides.

The excerpt describes an architecture and design rationale, not a result — no user study, no comparison condition, and no measurement of whether psychological safety, decision quality, or participation actually improved are reported here; the text explicitly frames the work as a "thesis" proposal rather than a completed evaluation. The thesis's own limitations section flags that text-only interaction may miss social and cultural cues, and that amplifying unvalidated minority opinions anonymously could introduce new biases. What follows at the strength this excerpt supports is a design pattern — paraphrase-and-anonymize as a shielding mechanism for dissent — worth tracking for its architecture and rationale, not yet for evidence that it works.