Coordinate and act
The wrong order, followed
People are expected to flag an instruction or AI output they believe is wrong instead of carrying it out simply because it came from an authority.
Borrowed from former US Navy submarine captain David Marquet, in his book Turn the Ship Around! Read the original source.
One session · Minutes to start · Anyone on the team can start it
Try this first
Set one expectation out loud. If you are asked to do something you believe is wrong, by a person or by the AI, you are expected to say so.
Use it when
People execute known-wrong instructions or outputs.
Skip it when
This is not a license to reopen every instruction someone dislikes. Keep the bar at wrong or unsafe, or it becomes cover for ignoring direction.
How to introduce it
Make it explicit that nobody executes an instruction they know or suspect is wrong, whether it came from a senior person or from a model. Raising it is part of the job. Say the reason plainly: trained compliance can walk a capable person straight through a bad order.
How to show up
State the expectation, then back the people who use it. The person who questions an instruction and turns out right needs visible support. Punish that once and people will stop trusting the norm.
How long it takes
A few minutes to set the expectation. It lives as a standing norm.
What makes it hard
Hierarchy and compliance run deep, so people need repeated proof that flagging a bad order is safe. The same deference attaches to confident AI output, so extend the norm out loud to "the model told me to."
What it looks like when it's working
People stop and question instructions and AI outputs they believe are wrong, and get thanked even when they turn out mistaken. Wrong-but-authoritative instructions still getting executed means compliance still rules.
How long until it sticks
It becomes real over repeated instances of flagging being backed rather than punished.
How you know it stuck
Questioning a wrong instruction, from a person or a model, reads as doing the job well.
The idea behind it
Authority does not make a wrong instruction safer. People should be expected to flag an instruction they believe is wrong or unsafe.
Where it comes from
David Marquet's work on submarine command. The anti-compliance norm. People are expected to flag an order (or an AI output) they know is wrong, not execute it. Marquet's turning point was learning a crew member had passed along an impossible order because he was trained to comply.
Why it matters for senior teams
Seniority can make this problem worse because people are less likely to challenge a high-status colleague
Also fits: The people who know something don't say it
Home problem: Everything escalates to the top