No One to Blame: A Framework of Constitutive AI Unaccountability

This paper introduces the concept of constitutive AI unaccountability to describe sociotechnical configurations in which AI accountability is conceptually unachievable, regardless of effort. The authors argue that existing research predominantly frames AI accountability gaps as barriers that can be overcome through better standards, transparency, and institutional reform, but this framing is insufficient because certain configurations of actors, systems, and institutions render AI accountability conceptually unachievable regardless of effort.

Through a three-stage qualitative study—comprising a concept-centric literature analysis, a secondary analysis of 27 expert interviews with AI professionals from technical, legal, and sociotechnical backgrounds, and an illustrative framework application to the open-source agentic AI system OpenClaw—the authors identify nine categories and 20 themes of constitutive AI unaccountability. These are organized across three clusters: structural, technological, and normative. The nine categories are: actor network dynamics, sanction incapacity, regulatory gap, systemic ambiguity, moral incapacity, temporal rationalization, accountability displacement, ideological rationalization, and economic-driven prioritization.

The framework identifies eight directed interdependencies between categories, showing that the conditions reinforce one another rather than operating in isolation. For example, Actor network dynamics feeds into systemic ambiguity because the diffusion of accountability across multiple actors, organizations, and recursive chains compounds the difficulty of understanding and tracing AI outcomes, and Economic-driven prioritization contributes to systemic ambiguity through a distinct mechanism: commercial secrecy.

The authors operationalize the framework as a diagnostic instrument of 20 questions. When applied to OpenClaw, the instrument detected 17 of 20 conditions, including an inverted anthropomorphism configuration in which the AI agent was the only identifiable actor operating under a constructed human identity while its operator remained unidentifiable.

The paper makes three contributions: (1) a framework of nine categories and 20 themes of constitutive AI unaccountability that extends the four barriers to accountability (diffusion of responsibility, technical errors, scapegoating, and severing of ownership from liability) by identifying conditions absent from prior work, mapping their interdependencies, and revealing asymmetries between academic and practitioner understanding; (2) a 20-question diagnostic instrument for practitioners, regulators, and auditors; and (3) an application to OpenClaw that surfaces 17 of 20 conditions.

The authors also classify the 20 themes along a continuum of unachievability: practically achievable (resolvable with existing measures), theoretically achievable (requiring changed arrangements that are possible in principle but contested in practice), and conceptually unachievable (where no possible arrangement can confer the requisite status because the capacities accountability presupposes are categorically absent). For instance, intra-organizational diffusion may be addressed through clear allocation of responsibility combined with sociotechnical tracking, while AI systems cannot experience punishment or remorse, making accountability conceptually unachievable in those cases.

The paper concludes by shifting AI accountability research from asking who can be held accountable toward identifying where accountability cannot be achieved.

Improvements to AI systems based on this paper:

1. Pre-deployment unaccountability auditing: AI systems can be augmented with a built-in diagnostic module that runs the 20-question instrument before deployment, flagging configurations where accountability is conceptually unachievable (e.g., no identifiable actor, no sanction capacity, or moral incapacity). The improved system would refuse to operate in modes where no human can be held responsible, or would require a human operator to be explicitly registered and traceable before enabling autonomous actions.

2. Dynamic accountability tracing for agentic systems: For open-source or multi-agent AI systems, the system can maintain a real-time causal graph of every action, decision, and data flow, linking each output to a specific human principal or legal entity. If the graph reveals actor network dynamics (diffusion across recursive chains) or accountability displacement, the system would automatically halt, escalate to a designated human, or generate a structured report identifying the exact point where accountability becomes untraceable.

3. Inverted anthropomorphism detection: The AI system can be trained to detect when it is being presented as a human (e.g., via user-facing personas, fake identities, or autonomous social media accounts) and when its operator is hidden. Upon detection, the system would refuse to continue operating under that constructed identity, or would inject a clear AI disclosure marker into all outputs, preventing the inverted anthropomorphism configuration identified in OpenClaw.

4. Temporal rationalization countermeasures: The system can implement self-imposed audit trails with immutable timestamps and no forgetting mechanisms for consequential decisions. If a user or operator attempts to delay, obscure, or retroactively rationalize accountability (e.g., by claiming we didn't know after the fact), the system would flag the discrepancy and trigger a mandatory review process, making temporal rationalization practically impossible.

5. Sanction capacity pre-check: Before executing high-impact actions (e.g., financial trades, medical recommendations, or legal decisions), the AI system would verify that a human with legal standing and sanctionable capacity is explicitly assigned and reachable. If no such human exists (e.g., a fully autonomous agent with no legal owner), the system would refuse to act, or would require a third-party auditor to be inserted into the loop.

6. Economic-driven prioritization transparency override: The system can be designed to expose any commercial secrecy or profit-driven constraints that influence its outputs. For example, if a model is fine-tuned to prioritize revenue over user safety, the system would surface this bias in a machine-readable accountability manifest attached to each output, allowing regulators and users to see exactly where economic incentives have shaped behavior.

7. Interdependency-aware self-diagnosis: The improved system would not just check for individual unaccountability conditions but would model the eight directed interdependencies (e.g., actor network dynamics → systemic ambiguity). If one condition is detected, the system would automatically probe for downstream conditions and report the full causal chain, enabling proactive mitigation rather than reactive fixes.

8. Conceptual unachievability gating: The system would classify each of its own functions along the continuum (practically achievable, theoretically achievable, conceptually unachievable). For functions that are conceptually unachievable (e.g., where punishment or remorse is required for accountability), the system would either (a) refuse to operate in that domain, or (b) require a human surrogate to explicitly accept liability in writing before each use, making the unachievability transparent and externally managed.

What the improved AI system can do: It can autonomously identify and avoid or mitigate its own unaccountability configurations before they cause harm, provide regulators with a standardized diagnostic report of its own limitations, refuse to operate under hidden or diffuse human control, and ensure that every consequential action has a legally identifiable, sanctionable human principal—or it will not act at all.