How to Decline

Someone senior will ask for something in the previous note. How that conversation goes determines both the outcome and whether you are consulted next time.

What does not work

Expressing discomfort. It reads as squeamishness and invites someone else to be asked.

Citing policy without explaining it, which invites a request to change the policy.

A flat no with no alternative, which gets routed around.

Arguing ethics first with an audience that has already decided the ethics are fine.

Overstating the legal position, which is discovered and discredits the accurate parts.

What works

Name the specific reason, in the order that lands with this audience.

For a commercial audience: legal exposure and reputational risk, with the penalty figures and the enforcement position.

For a technical audience: the error rate and its distribution, and the absence of ground truth.

For an operational audience: it will be worked around, and the data will degrade.

All three are true. Lead with the one that will be heard.

Answer the underlying question

Almost every request has a real concern behind it, and it is usually answerable.

"We want to know if staff are disengaged" — that is a management question, answered by asking them.

"We want to catch shoplifters" — detection of defined events works and does not require identifying everyone who enters.

"We want to know who was in the area" — access control with a credential answers this, with the person's participation.

"We want to reduce accidents" — condition detection plus the physical control that should be there.

Offering the answer to the real question is what turns a refusal into a consultation.

Where to draw the line on helping

Some requests should not be improved, only declined.

Making a covert deployment more discreet, making emotion inference more accurate, making a prohibited categorisation more defensible — none of these is a partial improvement. Assistance makes the thing more likely to happen.

Say that plainly: this is not something I can help make better, because a better version of it is still the thing.

That distinction matters and it is worth stating rather than leaving implicit.

Recording it

What was asked, by whom, when.

What was declined and why.

What was offered instead.

Who decided.

Kept somewhere findable, so the second identical request is answered by reference rather than argued again.

If overruled

Say so to the people affected, which is the honest thing and preserves what trust remains.

Insist on the mitigations you would have required anyway: human review before consequence, subgroup measurement, retention limits, a contest route, a review date.

Record your position in writing, dated.

Some things are worth refusing to implement personally, and knowing in advance which those are for you is better than deciding under pressure.

Preparing the position in advance

Writing it down before it is requested turns a negotiation into a reference.

State the categories you will not build, with the reason for each.

Name who decides on an exception.

Name what evidence an exception would require.

Get it agreed by whoever owns risk, before anyone asks.

Publish it internally, so a requester encounters a policy rather than one person's judgement.

The first refusal becomes the precedent, which is why it should not be improvised.

What to do if overruled

It happens, and the response determines what remains.

Say so to the people affected, which is the honest thing and preserves what trust is left.

Insist on the mitigations you would have required anyway: human review before consequence, measured subgroup performance, retention limits, a contest route, a review date.

Record your position in writing, dated, which protects both you and the record.

Decide in advance what you would decline to implement personally, because deciding that under pressure produces the wrong answer.