An escalation route is a defined path for raising a serious concern about an AI system that does not run through the person who has a reason to suppress it, with real protection for whoever raises it. Written like that, it sounds like a formality every organisation obviously has. In practice, most of what we find called an “escalation route” is an informal habit: raise it with your manager, who raises it with theirs, through the same chain that has every incentive to let a launch date survive intact.
The word doing the real work in that definition is “protected.”
A route that technically exists but offers no protection from being quietly deprioritised, reassigned, or managed out afterward is not a route…
We have seen organisations point to a policy document as evidence of an escalation route that had never been used.
What makes a route real, in our experience, is three things holding at once:
- it bypasses the person with a stake in the outcome,
- it is documented well enough to be pointed to after the fact, and
- there is a visible instance, even one, of it having been used.
The third one is the hardest to build and the easiest to fake with a policy nobody has tested.
Without a working escalation route, an AI ethicist’s independence is just a claim.