A human in the loop has to be able to say no
Otherwise it isn’t oversight. It’s decoration.
“Human in the loop” has become background noise, so it’s worth saying plainly what it has to mean. Three things, all of them, or it doesn’t count.
The person can see what the system is about to do. They have enough context to judge whether it’s right. And they have the authority — and the time — to stop it. Drop any one of those and you don’t have oversight; you have a signature.
A reviewer who can’t understand the output is rubber-stamping. One who can’t override it is a spectator. One who’s measured on never slowing things down will wave everything through, and everyone will call it human oversight. The hardest of the three to protect isn’t the seeing or the understanding — it’s leaving the person genuinely free to say no.
It’s the same idea I keep coming back to in the longer pieces on governing AI that acts, and the thing Eloryn is built to make real rather than rhetorical.
Written by Davor Cukeric — an AI builder, systems integrator, and problem solver in Ottawa, Canada, working on AI that earns its trust. More about me.