Who still holds final approval?
The unnerving AI story today is not what the model can say. It is who may be allowed to let it decide.
That shift is visible in separate places that belong to the same horizon. In England, one of the country’s top judges, Geoffrey Vos, said AI could adjudicate many civil cases, with the open question being where the line should be drawn, as a WIRED report noted. In Delaware, a plan would allow AI agents to run entire businesses with no human at the helm, according to Ars Technica’s report. And Anthropic said it is expanding its partnership with the US General Services Administration through a new OneGov agreement to bring Claude into federal, state, local, tribal, and territorial government buying channels, as Anthropic announced.
These are not the same institution and not the same risk. But they point in one direction. AI is moving from draft assistance toward recognized authority through rules, procurement, and operating structure. The important change is not that a model sounds more capable. It is that a form, a workflow, or an office may start treating its output as the decision that counts.
Once that happens, accountability becomes uneven. A system can produce the recommendation, ranking, ruling, or action path, while a person nearby remains the one answerable for it. That is the seat that quietly disappears first: not the person, but the person’s visible authority to say no.
That matters even when a human is still nominally involved. A cited finding says experts working with AI agents are swayed by the AI’s positions even when they think they are working independently, as Ethan Mollick summarized. And Sam Altman said people should not take important life advice from ChatGPT, in an OpenAI video. The reasonable inference is that formal authority can creep in before anyone openly declares that the machine is in charge. It can arrive through habit, deference, and policy language.
Find the last human veto
A prepared person does not need to solve AI governance in general to act usefully here. The practical move is smaller and more durable: identify the last explicit human sign-off in one workflow you touch, and make sure it stays visible, named, and required before AI output becomes action.
That sign-off might be a manager approval, a judge’s final review, a procurement officer’s authorization, a caseworker’s confirmation, or a founder’s written acceptance of an automated action. The title matters less than the fact that the approval point is written down and cannot be skipped by convenience. If the workflow says a person must approve, then authority has not fully drifted into the system.
What to map this week
- Pick one workflow where AI output can become an action.
- Find the last step where a named human must approve, reject, or verify it.
- Check whether that step is explicit in policy, ticketing, procurement, or operating documents.
- If it is vague, rewrite it in plain language for your team: who approves, what they review, and what cannot proceed without that approval.
- If the AI output is already treated as default, mark that as a governance issue rather than a tooling issue.
This is not glamorous work. It is how authorship and accountability stay attached to a human being long enough to matter. In a period when AI enters courts, companies, and government channels by process rather than spectacle, the written veto is the point worth defending.
The useful takeaway for this week is simple: map the last explicit human sign-off in one workflow you touch and make sure it remains visible, named, and required before AI output becomes action.
Human Runtime Core™ is designed for the mental edge you need to survive in an AI-run world.