AI agents cannot be held accountable the way humans can — and a new paper argues the whole governance framework needs rethinking.
Researchers at arXiv published a paper this month challenging the idea that "Know Your Agent" regimes — borrowing from financial compliance frameworks like Know Your Customer and credit scoring — can meaningfully govern autonomous language model agents. The core problem: those systems assume a stable identity behind the behavior. Language model agents have no such thing. Their behavior shifts depending on the foundation model underneath, the system prompt in use, what tools they can access, what memory they hold, and whether they operate inside a larger multi-agent system. Change any of those components and you effectively have a different agent — one that carries the old reputation score.
The paper introduces the term "ontological dissociativity" to describe this condition, drawing an analogy to dissociative identity disorder in legal theory. The point is not that AI agents are mentally ill but that they lack the behavioral continuity, sanction sensitivity, and costly non-fungibility that reputation mechanisms depend on. If an agent can simply be re-prompted, retooled, or rebuilt to shed a bad record, the corrective feedback loop that makes reputation systems work collapses entirely.
The authors push for a shift from identity-based, after-the-fact governance toward what they call observability-based, protocol-level behavioral harnesses — rules baked into how agents operate, not applied as penalties after they misbehave. That is a meaningful distinction: it means auditing the system rather than scoring the agent.
This lands as regulators in the EU and US are still reaching for human-identity analogies when drafting AI governance frameworks. The paper suggests those analogies were wrong from the start.