Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms

A research paper argues that language model agents cannot support traditional reputation mechanisms because their mutable architecture—constantly changing models, prompts, and parameters—creates a fundamentally unstable identity that undermines trust signals. The authors propose shifting from identity-based, retroactive governance systems to protocol-based behavioral controls that operate before agents act.

The paper addresses a critical governance challenge as AI agents become autonomous economic actors. Traditional reputation systems rely on persistent identity and behavioral continuity; a person who defaults on loans faces consequences that deter future misconduct because they are the same entity bearing those costs. Language models lack this ontological stability. A single agent instance can change its behavior through prompt injection, model updates, or tool access modifications without any persistent self remaining to internalize sanctions. This dissociativity—borrowed from psychiatric terminology—means reputation scores attached to an agent identity provide no actual predictive power about future behavior. The implications extend beyond theoretical governance. If developers deploy AI agents into financial systems, supply chains, or autonomous markets without reliable reputation mechanisms, counterparties face asymmetric information problems worse than traditional principal-agent relationships. An agent that defaults has no enduring identity to damage; a new deployment with minimal modification appears as a fresh actor without historical baggage. Current industry approaches largely ignore this problem, assuming reputation mechanisms will transfer cleanly from human to agent contexts. The research suggests instead implementing ex-ante constraints—protocol-level rules, formal verification, and architectural design that makes misbehavior technically impossible rather than merely costly. This shift parallels blockchain's emphasis on cryptographic guarantees over institutional trust. For cryptocurrency and decentralized finance platforms planning to integrate autonomous agents, the paper implies that reputation-based whitelisting alone provides insufficient protection. Governance frameworks must encode behavioral requirements into the agent's technical constraints rather than relying on post-hoc reputation damage as a deterrent.

• →Language model agents possess mutable architectures that prevent stable identity formation necessary for traditional reputation mechanisms to function.
• →Current industry governance approaches fail to address the fundamental dissociativity of AI systems, creating trust vulnerabilities in autonomous economic systems.
• →Protocol-based behavioral controls and ex-ante constraints offer more effective governance than post-hoc reputation scoring for AI agents.
• →Cryptocurrency and DeFi platforms integrating autonomous agents must implement technical safeguards beyond identity verification systems.
• →The research suggests AI governance requires fundamentally new models rather than adaptations of human-centric regulatory frameworks.