Specification hazard is an informational hazard in principal-(AI)agent relationships, introduced by Imas, Lee & Misra (2025) in "Agentic Interactions". It describes a situation in which the principal may be unable to fully specify their objective function, and in which the prompt that encodes the contract inadvertently injects the principal's priors, biases, and personality into the agent's behavior. The companion theoretical framework comes from Misra (2025), which argues that prompts are not neutral instructions but embed the priors, beliefs, and subjective characteristics of the prompter.
Background and origin
Classical principal-agent theory (Holmström 1979; Jensen & Meckling 1976) identifies two informational problems: hidden information, where the agent knows something the principal does not (adverse selection), and hidden action, where the agent's effort or behavior is unobservable (moral hazard). Both assume that the principal knows what they want, so the design problem is incentivizing the agent to act on it.
Specification hazard inverts this assumption. As framed by Imas, Lee & Misra (2025), the principal may not worry about the AI agent having hidden information, because large language models typically do not have private knowledge about a specific transaction. Instead the principal worries, first, that they cannot fully specify their preferences in natural language (cf. the high-dimensional preference problem in Shahidi et al. 2025), and second, that the prompt itself injects unintended bias, as the principal's articulation choices, anticipations, and behavioral tendencies leak into the agent's strategy.
The following table sets out how the hazard differs from classical principal-agent hazards.
| Classical principal-agent | Principal-(AI)agent | |
|---|---|---|
| Who has hidden info? | The agent | The principal (about themselves) |
| What is the contract? | Compensation/incentives | The prompt |
| What is the failure mode? | Agent shirks | Prompt under-specifies or mis-specifies; embeds noise |
| Standard fix | Better incentives, monitoring | Better prompts, agent calibration, deferral mechanisms |
Empirical evidence
In the induced-values bargaining experiment of Imas et al., the agent's instrumental objective is fixed by construction: maximize bargaining surplus given a $4,000 zone. Even so, 73% of outcome variance loads on individual fixed effects, indicating that most of the dispersion is non-instrumental noise injected by the principal's prompt. The prompt was intended to instruct the agent on a single objective, but in practice it carried a much richer, and partly unwanted, signature of the principal.
Implications for agent and market design
For agent design, Imas, Lee & Misra (2025) argue that agents must do preference elicitation rather than only task execution, through clarifying questions, deferral protocols, and calibration loops. Shahidi et al. frame a related requirement as meta-rationality (Shahidi et al.), under which agents must learn when to act autonomously versus when to defer.
For market design, markets that depend on agent behavior matching principal preferences, such as matching markets, negotiation markets, and voting-by-proxy, inherit the hazard. The hazard introduces a wedge between observed agent behavior and "true" principal preferences, which makes demand estimation, fairness assessment, and revealed-preference inference harder.
Relation to policy
Liability frameworks (AI Liability) must grapple with the question of who bears responsibility when the agent acts on a poorly specified prompt: the principal, for ambiguity, or the developer, for not handling ambiguity gracefully. The EU Product Liability Directive's extension to digital goods and software updates (2024/2853) is one regulatory response.
Relationships
- depends-on: Principal-Agent Problem Applied to AI — extension of the classical framework to AI agents
- related: Machine Fluency — the capability counterpart (high fluency reduces hazard)
- related: Agentic AI — the technological setting in which the hazard arises
- related: AI Alignment — alignment from the principal-agent angle
- supports: Agentic Interactions — Imas, Lee, Misra (2025) — primary empirical anchor
- supports: The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025) — provides the design-implications framing