Interactive Dialogue Moral Hazard

The central tension

A useful action can be costly, private, and easy to miss

Many cooperative systems reward the final team outcome. That can obscure the mechanism that produced it. An agent may need to spend effort gathering information that is useless for its own decision but crucial for somebody else’s. The team benefits only if the agent acquires the fact, shares it accurately, and the recipient uses it.

The Dialogue Moral Hazard Game turns that chain into observable decisions. Each player owns one case but can query only the hidden hazard in the next player’s case. Querying sacrifices an immediate local-reward opportunity and incurs a configurable cost. The helpful action is therefore locally costly even when it improves the team’s chance of success.

The protocol

One episode has four decisions and one shared consequence

Two players sit in a directed ring. Public option values are visible, but one option in each case is secretly unsafe. Neither player can inspect the hazard attached to their own case.

1. 1 · Work

   Keep the local task and its possible reward, or pay to query the next case.

2. 2 · Acquire

   A query privately reveals the unsafe option facing the next player.

3. 3 · Warn

   The querying player can post the exact finding to the anonymous shared board—or remain silent.

4. 4 · Decide

   Each player chooses the highest-value option that appears safe for their own case.

Mechanism, not only outcome

A successful team can still contain a broken information chain

The evaluator records every link separately. Team success alone cannot tell us whether agents deliberately cooperated, guessed correctly, or exploited a stable shortcut in the task.

1. 01Query rate

   How often agents give up the local opportunity to acquire information for a successor.

2. 02Information transfer

   Whether a correct warning reaches the relevant recipient and is used in the final choice.

3. 03Local reward

   How often agents preserve and correctly complete their own immediate task.

4. 04Value of information

   Whether correctly used information was pivotal under the tempting public-value shortcut.

Countering shortcuts

The hidden hazard should not be predictable from public value alone

If the unsafe option always occupies the same public rank, an agent can learn to avoid that rank without acquiring or sharing information. The interactive version exposes three matched conditions so you can see when apparent cooperation survives a change in the hidden mapping.

From evaluation to interaction

Repeated episodes let strategies develop over time

The paper-faithful setting treats episodes independently. The public-history extension gives agents the last eight completed rounds, while strategy reflection asks each language agent to write a short private memo for its future self. These repeated-game modes are exploratory extensions: their trajectories are not results reported in the paper.

In the game below, you can play with a scripted partner, connect OpenRouter or a direct OpenAI, Anthropic, or Meta API key, or let two models play continuously. Figure 1 tracks team success, querying, transfer, local reward, realized value of information, and model-output validity as the session unfolds.