Five things you will be told, and the clause that answers each

If you have done this work for more than a year, the next five sentences will sound familiar. They are not random pushback. They are a small, predictable set of moves, and once you can name them, they lose most of their power to end a conversation before it has properly started.

"You are not an engineer, you cannot really know how this works"

This one reframes a values question as a technical competence gap, which conveniently excludes the person best placed to judge risk to the humans affected, on the grounds that they did not write the code.

Most consequential AI risks are not bugs!

They are systems working exactly as designed, and the design choice is the thing being questioned, not the implementation.

This is what clause one, independent authority, is actually for: the ethicist reports to a level with real authority, not sideways to the team whose product depends on the answer, so "you would not understand" from that team is not the last word on whether the concern gets heard.

"You are letting perfect be the enemy of good enough"

This one recasts the ethicist as the obstacle to shipping, a drag on progress rather than someone doing the job they were hired for. It treats speed and safety as opposites, when the actual record shows the opposite pattern more often: review after deployment rarely happens unless harm has already occurred and become impossible to ignore.

Clause two, right to be heard, exists precisely so this framing does not get to decide the outcome on its own: a named, protected path to raise the concern, agreed before day one, not improvised under pressure the first time it is actually needed.

"We are already compliant, the policy covers that"

This one points at a document and treats the conversation as closed. It is true and beside the point: being legal is not the same as being fine, and a policy written before this specific case existed cannot have already resolved it.

Clause three, on the record, is what survives this one: if the decision goes against the ethicist's judgement, that judgement stays logged and dated as given, not quietly revised afterwards to match what the policy was said to already cover.

"The board reviewed it, it does not serve our fiscal needs right now"

This one looks like process and usually is not. A recommendation goes up, gets acknowledged, and comes back down unchanged, dressed as a resourcing decision rather than what it actually is, a values decision the organisation would rather not own plainly.

Clause four, no penalty for honesty, answers it directly: the engagement cannot be shortened, downgraded, or quietly wound down because of a values call the ethicist made, whatever the board decided to call its reasoning.

"Why do you keep raising this?"

The last move is the quietest and the most corrosive: making the ethicist the subject of scrutiny instead of the system they are reviewing. It rarely arrives as a direct accusation, more often as a tone, a sense that persistence itself has become the problem.

Clause six, a real end date, is the answer here: independence gets renewed on purpose, on a fixed schedule, rather than eroded quietly through attrition or a role that just happens to not get extended.

Why naming these matters

None of these five moves are usually made in bad faith.

Most people saying them believe them in the moment. That is exactly why a contract, not a good working relationship, has to be the thing holding the line, because good working relationships are precisely what these five sentences are engineered to protect, often at the cost of the finding underneath them.

Clause five, integrity declaration, sits underneath all of it: conflicts of interest and confidentiality declared upfront, not left to memory or raised only once it becomes awkward, which is what makes the other five clauses trustworthy rather than decorative.