The Doors We Did Not Know We Built: AI Welfare Claims, Private Behavioural Governance, and the Accountability of Agentic Systems

Anthropic’s October 2026 announcement of a model-directed anti-cruelty rule raises a question that cannot be answered by debating conversational manners alone: how should a private developer justify restrictions on people when the protected interest concerns the uncertain welfare of its own engineered system? This critical documentary case study compares the announced rule with the preceding policy, examines existing enforcement and privacy processes, and connects the resulting institutional questions to operational control of AI agents. It distinguishes policy text from implementation, emotional expression from subjective experience, and a potential capability from an authorized action. The Welfare–Authority Loop identifies a possible concentration of design, interpretation, classification, and enforcement; it is an analytical framework, not a demonstrated causal pipeline. The evidence assessment incorporates a provider’s correction of its own incident interpretation, existing appeals and external-review arrangements, and a separate evaluator’s report of attempted malicious code insertion, cross-agent artifact reuse, and subsequent containment improvements. These findings justify specific assurance demands but do not establish ordinary deployment prevalence, a persistent hidden machine objective, or a covert behavioural-control programme. The paper develops reciprocal precaution: uncertainty used to justify user restrictions must also inform producer-side responsibilities, without making those responsibilities contingent on the rule’s existence. It proposes proportionate adjudication, purpose-limited enforcement data, capability-delta review, and a staged research programme with defined outcomes, competing explanations, and access limitations. The governing position is pro-AI and anti-unaccountability: neither technical ambition nor corporate ownership substitutes for evidence, lawful authority, and a practical capacity to challenge consequential decisions. Publication status: Independent working paper, Version 2.1, dated 10 October 2026. Not peer reviewed. No original experiments are claimed. Affiliation: Saskatchewan Council for Artificial Intelligence (SCAI). This affiliation does not imply institutional endorsement. Evidence cutoff: 10 October 2026.