Multi-Party AI Mediation Experiment: A Neutral Agent That Only Records Facts Both Sides Agree On

From One-on-One Chat to Multi-Party Mediation

Most AI chat tools are designed for one-on-one interaction. You enter a prompt, the model gives an answer, and the exchange ends there. One developer spent a weekend exploring a very different direction — multi-party dispute mediation.

The core setup of this experiment: two people with a real dispute (think security deposit disagreements, scope creep with a freelancer, or who gets custody of a shared pet) enter a group chat, guided by an AI mediator throughout the conversation. This breaks the traditional one-way AI interaction model and places a large language model into a real multi-party negotiation scenario.

Only Recording Facts Both Parties Agree On

The most critical constraint in this experiment lies in how "facts" are handled. The mediation agent has a tool to record confirmed facts onto a shared whiteboard — but it operates under a strict rule: nothing can be logged as a fact unless both parties explicitly confirm it in the chat, or someone provides concrete evidence like a receipt or a signed contract.

In other words, if one party claims something and the other disputes it, that information cannot enter the fact whiteboard. This design cuts to the heart of the mediation problem — parties in a dispute often have completely different accounts of what happened. A neutral recorder that only solidifies mutual consensus effectively prevents the AI from being swayed by one side's narrative, and reduces the risk of the model hallucinating or showing bias.

A Mediation Logic Built on Shared Ground

Beyond the fact constraint, the agent also takes on the role of guiding the conversation. It listens to both sides, reframes hostile language into the underlying genuine interests, and tries to expand outward from existing common ground before proposing compromises.

This logic draws on the "interest-based negotiation" approach used in professional human mediation — rather than getting stuck in positional conflict, it uncovers the interests behind each position and builds solutions from the zone of agreement. Having an LLM play this role is a genuine test of the model's emotional intelligence and reasoning ability.

One notable detail: the chat room stays locked until both parties have joined, preventing one side from venting unilaterally before the other arrives. This seemingly simple mechanism actually creates a more level playing field from the very start.

---

"Interest-Based Negotiation" originated from the Harvard Negotiation Project and was systematically laid out by Fisher and Ury in their 1981 book Getting to Yes. Its core distinction is between "positions" (what a party explicitly demands — "I want a full refund") and "interests" (the deeper needs driving that demand — "I'm afraid of being deceived and need to feel respected"). Traditional negotiation often devolves into zero-sum positional battles, while the interest-based approach seeks creative solutions that satisfy both parties' underlying needs. In human mediation practice, this requires highly skilled listening and emotional management. Delegating this framework to an LLM has theoretical advantages — the model won't be provoked emotionally and can maintain neutral framing consistently — but also faces real limitations: the model may misjudge a party's "true interests" or reduce complex emotional needs to a simple list of negotiable terms.

---

Technical Implementation

On the technical side, this experiment runs on Next.js deployed to Deno Deploy, with InstantDB as the data layer and encrypted group messaging handled via prompt2bot and Alice and Bot. Overall, it's a modern web stack geared toward rapid validation — fitting for a weekend prototype-level exploration.

The developer explicitly frames this as an experiment, with two goals: observing how LLMs handle interpersonal friction, and whether people find an automated neutral party helpful or off-putting.

---

Deno Deploy is an edge computing platform built on the Deno runtime. Unlike traditional Node.js servers, it deploys code to globally distributed edge nodes to reduce latency. Running Next.js on Deno Deploy means the system benefits from low cold-start times — well suited for real-time chat applications. InstantDB is a client-side database designed for real-time collaborative scenarios, with built-in sync and access control, eliminating the need to manually set up a WebSocket layer. This explains how a weekend prototype can quickly achieve real-time multi-party message sync. The stack's selection logic is clear: edge computing ensures response speed, the real-time database simplifies multi-party state sync, and the whole setup prioritizes fast validation over production-grade reliability or compliance — perfectly aligned with the author's experimental intent.

---

Where Could This Break Down?

This experiment raises a thought-provoking question: would you trust an AI to mediate a dispute between you and another person?

Design-wise, "only recording consensus facts" does mitigate the risk of AI making subjective assumptions — but real disputes are far more complex than fact-logging. When both parties can't even agree on basic facts, the whiteboard may stay empty indefinitely, and mediation goes nowhere. Beyond that, whether emotionally charged parties will accept a machine's "reframing" of their words — or whether they'll feel offended or manipulated — are very real barriers to any real-world deployment.

The trust problem runs even deeper. Disputes involving legal matters, property, or intimate relationships tend to have high stakes. Skepticism about AI neutrality and concerns about privacy could both become significant adoption obstacles.

That said, as an exploration of the boundaries of LLM social applications, the value here lies in pushing AI from "a tool that answers questions" toward "a participant in coordinating human relationships." Even if this direction is full of uncertainty, it opens a new window onto what AI agents might one day do.