The Second Pair Of Hands

The frontier training objective in artificial intelligence is single-form omnicompetence: one model, one pass, every problem, a complete and confident answer. This working paper argues the objective is impossible in principle — Ashby's law of requisite variety establishes that no single controller, human or machine, carries sufficient internal variety to govern an unbounded environment — and that humanity's working solution has never been a better individual. It has been the team. Read through that lens, the observable failure modes of deployed AI stop looking like separate problems. The paper partitions hallucination into specification, incentive and capability failures; reframes user dependency through the supported/substituted decision-making distinction codified in English capacity law; and identifies register and incentive filtering as the mechanism by which safety signal dies in human teams — a mechanism hybrid human-AI teams will inherit unless designed otherwise. It closes with a sketch of a team-insertion audit treating clarification, escalation, dissent, second signature, handover and supervision as first-class, measurable safety properties, and four falsifiable predictions. Two features of the record. Appendix A is a contributed note from the assisting AI system (Claude Fable 5, Anthropic), retained verbatim, with confidence calibration attached at the author's requirement. And the citation table was verified by an independent AI audit pass that itself produced two errors — caught by cross-check and recorded — a live instance of the paper's own argument about maker-checker loops. Part of the Heartbeat Framework series. Scope generalises to any industry that runs on team functions, wherever an AI system is inserted as a de facto team member rather than a tool.