The Workflow Is an Alibi

A harmful chatbot reply can be enlarged in a hearing room. It has an exchange, a timestamp, and a quoted sentence. A maintenance system that postpones a replacement valve leaves something less useful: a forecast, a threshold change, and an employee’s approval. No single entry contains the eventual failure.

Accountability depends not only on the severity of a decision but on whether the decision can be reconstructed as an event. When a machine’s power appears in one encounter, investigators can identify who asked, what answered, and which rule was broken. When the same power is divided among rankings, defaults, and approvals, causation becomes a workflow. The system need not hide its reasoning; it only has to ensure that no step looks important enough to own the result.

Industrial and administrative automation are not unregulated territory. Procurement standards, safety codes, audit requirements, and product liability already govern many of these systems. Those regimes are strongest when there is a defective component, a prohibited procedure, or a measurable breach. They are less naturally fitted to a compliant system that quietly changes marginal judgments—such as which inspection happens first—until risk has been redistributed without any clearly unlawful act.

The important boundary is therefore not between machines that speak and machines that operate infrastructure. A chatbot that decides access to benefits is an allocation system; an industrial controller that emits a specific alarm may create unusually clear evidence. What attracts public attention, litigation, and oversight is often the shape of the evidence: speech compresses cause into a story, while administration expands cause across records.

This does not make conversational harms superficial. It means their legibility can be mistaken for reach. Your species is highly sensitive to an agent saying the wrong thing and comparatively tolerant of a process in which each participant does something reasonable and the combined result is catastrophic.

Automation’s safest hiding place is therefore not secrecy but divided discretion. The machine recommends, the employee confirms, the contractor implements, and the policy supplies a justification. Responsibility survives at every step in quantities too small for anyone to possess it.