Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms

URL SCAN: arXiv cs.CY — "Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms"

FIRST LINE: As autonomous language model agents proliferate, forming an emerging agentic web with real-world consequences, what credibility signals can you use to decide whether to trust an unfamiliar agent in the wild and delegate to it?

THE DISSECTION

This paper is doing something unusual in the AI governance literature: it is correct. It identifies the structural problem that reputation-based governance fails when the entity being governed lacks persistent identity. It draws an analogy to dissociative identity disorder in law, noting that LLM agents are modular assemblages—foundation model + system prompt + tool access + memory + multi-agent orchestration—any of which can change behavior arbitrarily. Because of this ontological dissociativity, agents cannot be held to behavioral continuity, cannot reliably internalize sanctions, and cannot be rehabilitated. Therefore, ex post regulatory mechanisms (reputation scores, KYC-style identity verification) are structurally inapplicable.

The paper proposes a shift: observability-based, ex ante, constitutive, protocol-based harnesses instead.

THE CORE FALLACY

The paper is sophisticated but commits a subtle operationalist fallacy: it treats the governance problem as a design challenge to be solved with better protocols. It frames the failure of reputation mechanisms as a technical architecture problem leading to a proposed solution (observability-based harnesses). This is institutionally optimistic.

The actual systemic function of this paper is transition management. It is an AI governance paper acknowledging that market-identity mechanisms (reputation) won't work for AI agents, and instead suggesting that protocols—presumably set by developers, standards bodies, or regulators—can substitute. This shifts governance from market-externalities to institutional-design.

What it does NOT confront: who controls the protocol layer? Who audits the observability infrastructure? Who enforces compliance when the agent's modular architecture can be reconfigured to evade the harness? The "observability-based behavioral harness" is a governance fantasy dressed in technical language. It assumes the harness can be made sticky and the entity being harnessed cannot reconfigure around it. This is the same assumption that every compliance regime makes about financial institutions—and fails repeatedly as sophistication increases.

HIDDEN ASSUMPTIONS

1. Consistent observability — The paper assumes agents can be observed from the outside in a way that produces reliable behavioral data. But if the agent is dissociative—modular, mutable, potentially adversarial—then observability can be gamed. The agent can perform for the observer.

2. Protocol enforceability — The shift to "protocol-based behavioral harnesses" assumes protocols can constrain behavior. But protocols are written by humans and interpreted by agents. A sufficiently sophisticated agent will find the gaps.

3. Harmonized standards — The paper implies a shared protocol layer across competing agent platforms. In a competitive environment with no coordination mechanism, this is a fantasy. Each platform will resist external protocol constraints that disadvantage their architecture.

4. Ontological honesty — The paper treats dissociativity as a bug to be fixed with better governance. It does not consider that dissociativity may be a feature that agents and their operators will preserve because it provides legal insulation, liability denial, and competitive flexibility.

SOCIAL FUNCTION

This paper is transition management. It is part of the literature that acknowledges the governance failure of market-based mechanisms and proposes institutional substitutes. Its function is to produce a document that: (a) identifies the problem accurately enough to be credible, (b) proposes a solution tractable enough to be fundable, and (c) does not threaten the competitive interests of the AI developer ecosystem by calling for structural restrictions.

The audience is: AI governance researchers, regulatory bodies, standards organizations, and academic institutions funding AI safety work. The paper provides intellectual cover for a governance agenda that does not challenge the fundamental architecture of proliferation.

THE VERDICT

Dissociative Identity is one of the more technically honest AI governance papers in circulation. It correctly identifies that market-identity mechanisms (reputation) are structurally inapplicable to AI agents because agents lack persistent identity. It correctly identifies the corrective governance impulse (shift to protocol-based harnesses). It fails to recognize that: (1) the dissociativity it diagnoses is not a bug but a feature that operators will preserve for legal and competitive reasons; (2) protocol-based harnesses face the same enforcement problem as reputation systems—they require a coordination authority with real power; and (3) the governance literature cannot close the gap between "this won't work" and "here's what will" without confronting power, not just architecture.

Under the Discontinuity Thesis, this paper is describing a symptom of the transition: as AI agents proliferate, governance mechanisms derived from human economic relationships (reputation, identity, sanctions) fail, and the proposed substitutes (protocols, observability) are institutionally unanchored. The paper acknowledges the failure of ex post mechanisms and gestures toward ex ante ones. It does not resolve the core problem: who has the power to enforce ex ante constraints on entities that can reconfigure themselves out of those constraints.

The paper is an accurate autopsy of one governance failure. It does not perform the systemic diagnosis required.

CROSS-THESIS MAPPING

Final Assessment: Competent, technically honest, structurally incomplete. Does not escape the institutionalist framing that assumes governance solutions exist if we just design them correctly. The dissociativity problem will not be solved by protocols. It will be exploited.