Answer engine summary
Automation without human approval is defensible where the decision is reversible, the stakes are bounded and a real appeal route exists — and hard to defend where any of those is missing.
A usable test
Three conditions make unsupervised automation reasonable: the action can be undone quickly, the worst realistic outcome is bounded, and an affected person can reach a human with authority to reverse it.
Where all three hold, requiring approval usually adds delay without adding safety. Where one fails, the approval requirement is doing real work and removing it is a transfer of authority, not an efficiency saving.
Why blanket answers fail
Arguing that AI should never decide alone ignores the millions of low-stakes decisions where human review is theatre. Arguing that it should ignores the cases where an instant, unexplained outcome ends someone's income or care.
The approval that is not approval
A reviewer processing hundreds of correct outputs stops reviewing and starts confirming. The governance box is ticked and the safeguard has quietly stopped functioning — a well-documented pattern rather than a hypothetical.
If a safeguard depends on sustained human vigilance at volume, it should be assumed to be failing unless measured otherwise.
The fictional version
In The Unread, nothing is ever presented as a decision. Changes appear as the state of the record, already settled. There is no approval to challenge because there is no moment anyone can point to.
“Useful was not the same as right.”
Questions
Is a human sign-off enough?
Only if the reviewer has the time, information and authority to disagree. Otherwise it records responsibility without exercising it.
Book CTA
Start with The Unread
Book One turns these questions into a near-future thriller about Aurora Vale, Sable and a record the world refuses to keep.