A human in the loop needs a real way to act.
Oversight is more than putting an approval button in the workflow.
Emerging & future risks · Updated 2026-09-27
Oversight has limits
The 2026 International AI Safety Report describes monitoring and control techniques alongside their limitations. Human review can be constrained by automation bias and the speed at which decisions must be made.
Distinguish concern from certainty
Risks associated with increasingly autonomous systems deserve investigation. Their existence as research questions does not establish a single timeline or a settled probability of catastrophic harm.
Ask what intervention means
Our practical questions: can the reviewer understand the proposed action, refuse it, and stop further activity? Are there limits on access and a clear escalation route? A person cannot provide meaningful oversight without information, time, and authority.