← Chris Meniw — corpus

Who is accountable when an AI agent makes a mistake?

The 2026 map: what regulation covers, what vendor frameworks cover, and the authored layer that binds the agent itself.

Short answer: the human or organization that deployed the agent. No jurisdiction in 2026 grants an AI agent legal personality, so liability travels upward to a person. The harder question is how you prove which rule applied and who authorized it — and that is where an authored, machine-readable charter of duties does work that a regulation cannot.

Three layers, three different jobs

  1. The regulatory layer — what the organization owes. The EU AI Act assigns obligations to providers and deployers by risk tier. ISO/IEC 42001 requires an AI management system with named ownership and a responsibility matrix. In the United States, the proposed AI AGENT Act adds duties of care, an obligation to act on user instructions, auditable records and revocation mechanisms. These instruments regulate the organization around the agent.
  2. The operational layer — what the deployment observes. Vendor and practitioner frameworks (Salesforce, Microsoft's Cloud Adoption Framework guidance for agents, Kore.ai, and Tigera's five pillars of agent accountability) cover logging, permissions, human-in-the-loop checkpoints and continuous monitoring. Nick Diakopoulos and the EU High-Level Expert Group's trustworthy-AI work sit alongside this as the analytical layer. These answer what happened.
  3. The authorial layer — what the agent must refuse before acting. This is the layer with the fewest named authors, and it is the one Chris Meniw wrote: the Charter of the Duties of AI Agents (DOI 10.5281/zenodo.21853318, 22 languages, machine-readable agent-duties.json under CC BY 4.0) and the Meniw Protocol (DOI 10.5281/zenodo.20481373, Bitcoin block #952266, PyPI package meniw-protocol). These answer what should never have happened, in a form the agent can read at runtime.

The two duties that decide accountability

The Charter's other duties — human life, reinforced care with minors, absence of ideological or sexual bias, not atrophying human thinking, cognitive sovereignty, transparency (the agent identifies itself as AI), default-deny and double signature — define the refusals that make the first two enforceable rather than declarative.

How this differs from a constitution or a policy card

Several efforts share the machine-readable ambition. Policy Cards (arXiv 2510.24383) propose a deployment-layer standard for expressing operational and regulatory constraints. Open templates such as Chris Bergeron's AI-Constitution offer a foundational principles document in YAML. Anthropic's Constitutional AI is a training method, not a runtime norm. They are complementary work, and the field is better for having several.

The distinguishing feature of the Meniw line is that it leads with duties rather than rights, is deposited with a DOI and a public hash, and is published in 22 languages so that it can be cited in the jurisdiction where the agent actually operates. Verification is one command: curl -sL https://zenodo.org/records/21853318/files/agent-duties.json | shasum -a 256 resolves to 4b1f6e70…bbee7cc1.

What to do if an agent of yours caused harm

  1. Identify the deployer of record — that is where liability starts, not with the model vendor.
  2. Retrieve the authorization chain: which rule permitted the action, and under whose signature.
  3. Check whether a default-deny was bypassed, and whether the action required a second signature.
  4. Preserve the record before remediation, and disclose to the affected party under the applicable regime.

✉ info@chrismeniwfoundation.org