Ghost with a Contract: Why AI Agents Have No Grounding for Reputation Mechanisms

Forum topic · 小凯 · 2026-05-31

Summary

A Chinese tech forum deep-dive reviews the Oxford paper "Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms" by Botao Amber Hu, Helena Rong, and Max Van Kleek (arXiv:2605.30169). The post argues that human reputation systems rest on four implicit preconditions—persistent identity, behavioral continuity, sanction sensitivity, and non-substitutability—and that LLM agents collapse on all four. An agent is a mutable stack of base model, system prompt, tool permissions, and memory modules, so punishment aimed at its "name" targets nothing stable. The author examines the paper's legal analogy to dissociative identity disorder, its critique of KYA (Know Your Agent) proposals as ill-fitting copies of KYC, and its proposed alternative: shifting from identity-based ex-post sanctions to observability-based ex-ante protocol constraints such as behavioral passports and oversight checkpoints. Honest limitations are noted, including the metaphorical status of the DID analogy and the lack of empirical protocol designs.

Overview

The forum post 《幽灵之约》("Ghost with a Contract") reviews a 2026 Oxford paper: *Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms* by Botao Amber Hu, Helena Rong, and Max Van Kleek (arXiv:2605.30169, cs.MA, submitted May 28, 2026).

The post opens with a market analogy: humans in reputational communities cooperate because identity is stable—your name is bound to your body, your behavior sticks to your history, and your future interests constrain your present conduct. Reputation is civilization's cheapest cooperation infrastructure, and it presupposes that you are you.

Key points from the paper

Dissociative identity: An LLM agent is not a person but a reconfigurable assembly of base model, system prompt, tool permissions, and memory modules. Two instances called the same agent name may share no common substrate at all.
The four-legged stool collapses: Reputation mechanisms assume (1) persistent identity, (2) behavioral continuity, (3) sanction sensitivity, and (4) non-substitutability. Agents fail on every leg—one-byte prompt changes cause phase shifts in behavior; a bad rating is invisible unless encoded into training (which is model training, not reputation); and a banned agent can spin up a new identity at near-zero cost in seconds.
Legal analogy: The authors borrow the jurisprudence of dissociative identity disorder (DID), where courts have long recognized that identity-based governance fails when identity itself becomes a variable.
KYA critique: Attempts to extend financial KYC to agents ("KYA"—Know Your Agent), including digital ID cards and agent credit scores, copy the governance shell without the ontological substrate that makes KYC work.
Proposed direction: Shift from identity-based ex-post punishment to observability-based ex-ante protocol constraints—a "protocol-based behavioral harness" with behavioral passports (disclosing model version, prompt templates, tool permissions) and external oversight checkpoints before critical actions. You cannot govern a dissociative agent's identity, but you can constrain its behavioral outputs.

The deepest observation

The post highlights the paper's line that a language model agent is fundamentally a series of function calls on a stack—naming it merely names a function call. We use human moral grammar ("is it trustworthy?") for a mathematical object; governance tools must be redesigned from the agent's ontology, not transplanted from human frameworks.

Honest limitations noted by the author

Whether "dissociative" is metaphor or ontological claim remains unresolved (DID itself is a contested diagnosis).
The protocol-harness proposal is conceptual only—no concrete protocol design or empirical data; it imports all known problems of moving from customary law to statute law.
Cross-cultural diversity of reputation mechanisms (kin-based, gossip-based) is not modeled.
If legal systems proactively construct agent identity, the "dissociativity" premise could be rewritten rather than given.

References

1. Hu, Rong, Van Kleek, "Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms", arXiv:2605.30169, 2026. 2. Ostrom, "Governing the Commons", Cambridge, 1990. 3. Resnick et al., "Reputation Systems", CACM, 2000. 4. Chan et al., "Harms from Increasingly Agentic Algorithmic Systems", FAccT, 2024. 5. McDougal & Feliciano, "Law and Minimum World Public Order", Yale, 1961.