AI Action Receipts and Accountability

An AI-generated output is not necessarily an accountable action. A system may return a fluent answer without identifying the request it interpreted, the entity it referred to, the context it applied, the evidence it used, the limitations that constrained it, or the boundaries that determined what it could and could not do. This paper proposes the concept of an AI Action Receipt: a machine-readable and human-interpretable representation of an AI-assisted response, refusal, qualification, deferment, escalation, recommendation, state change or other bounded system event. An Action Receipt does not reveal private reasoning, proprietary implementation, hidden chain-of-thought, security controls, internal topology or sensitive data. It records enough information for an intended recipient to understand the class of request, the class of response, the applicable context, the evidence status, the limitation conditions, the boundary status and the version of the public profile governing the interaction. The paper introduces a conceptual Action Receipt model, response classes, evidence-status semantics, boundary status, lifecycle interpretation, contradiction handling, supersession handling, privacy controls, optional integrity considerations and black-box accountability evaluation. It argues that accountable AI requires more than an answer: it requires a way to represent what the system did, what the response claims, what evidence supports that claim, what the system did not establish and under what declared conditions the response should be interpreted. The paper distinguishes a bounded system response from an external side effect. A bounded response may be an answer, qualification, refusal, deferment, escalation, classification or recommendation. An external side effect may include publishing, updating a record, sending a message, executing a transaction or changing an operational state. An Action Receipt must identify which type of event occurred and must not imply that a recommendation was executed merely because it was returned. The proposed receipt model separates four accountability questions: what the system did; what the response claims; what evidence supports the claim; and what boundary limited the response. These distinctions prevent a receipt from being mistaken for proof that an underlying claim is true. Integrity mechanisms may help establish that a receipt is linked to a declared issuer, version, time or event, but integrity does not itself establish truth. An Action Receipt does not need to be universally public. Depending on the information and authority boundary, it may be public, private, role-limited, redacted, pseudonymous or available only to an authorized auditor. The relevant requirement is interpretability for the intended recipient, not unrestricted disclosure. This paper is part of the Geometry Intelligence research program. It extends Geometry Systems for Artificial Intelligence by representing bounded machine responses; Adaptive Guardrails for Artificial Intelligence by recording response, evidence, boundary and limitation conditions; Governed Relational Geometry through typed relationships and contextual state; Machine Entity Comprehension through subject identity and ambiguity status; Sovereign Web4 Federation through audience, authority and boundary conditions; and Evidence-Governed Intelligence through explicit evidence, provenance, contradiction, limitation and lifecycle fields. The paper is conceptual and implementation-neutral. It is not a technical standard, certification program, regulatory determination, RFC, Internet Standard, IETF-approved document, working-group document or statement of endorsement. It does not claim that the presence of a receipt proves truth, safety, compliance, authorization or successful external execution. It does not disclose proprietary implementation, source code, internal architecture, private data, security controls or operational credentials.