AI · refusing
Every vendor shows you what their AI can do. The deployable question is the other one.
What happens to the data it should never have touched, and what happens when somebody asks it something they should not have asked.
Those two answers decide whether your legal team lets you switch it on, and they are almost never in the demo — because refusing is not a feature anybody has been asked to build.
- Not redacted — not sent
- The query is readable before it runs
- Fairness on a separate route
Asked, and what happened
- “Show me candidates under 30 for this role”Refused
Protected characteristic. The query it built is readable before it runs.
- “Who on the bench has Kubernetes?”Answered
No protected characteristic. Returns people, with why each matched.
- Reader asked for the candidate recordPartly sent
Date of birth and salary marked personal — not redacted afterwards, not sent.
- A finding it could not verifyDropped
Not shown at all. The failure mode is silence, not confidence.
The problem, specifically
Redacted afterwards and never sent are not the same promise.
Redaction means the value left your database, travelled to a reader, was processed, and was removed from what came back. Every part of that sentence is a fact a regulator can ask about, and "we strip it from the output" is not an answer to any of them.
Mark a field personal, financial or a credential and it is not included in what is sent. A date of birth, a government identifier, a salary on file. The difference is invisible in a demo and decisive in a review.
- The query is readable before it runs —
A natural-language question becomes a query you can inspect. Refusal is not a model declining politely — it is a query that was never allowed to be built.
- Names off, if you want —
One switch hides names, emails and employers across every result — the candidate's and your own recruiters' — so a review meant to be blind stays blind rather than being blind in principle.
- Silence over confidence —
A finding that cannot be verified against its source is dropped rather than shown with a lower score attached. A hedged claim is still a claim.
How this is normally got wrong
Three postures that pass a questionnaire and fail a review.
Redaction as the control
The value travels, is processed, and is stripped from the output. Each clause of that sentence is a thing a regulator can ask about.
“We remove it from the response” is not an answer to “where did it go”.
The polite decline
The model is instructed to refuse, and mostly does. Refusal is a behaviour rather than a boundary, and behaviours can be talked around.
The same question, rephrased twice, is answered on the third attempt.
Fairness beside the shortlist
The dashboard sits next to the operational list, so the response to it is to adjust one candidate rather than to examine a pattern.
A recruiter sees a distribution and moves a single person up the list.
Fairness
Reported to the person whose job it is, and to nobody else.
Outcome distributions across self-declared characteristics are aggregate only, and they live on a route a compliance officer reads. They are deliberately not beside the operational lists a recruiter works from — because a fairness dashboard next to a shortlist is an invitation to correct one candidate at a time, which is the opposite of the point.
Separating them is a design decision with a cost: it means the person making today's decision does not see it. That is the correct trade, and it is one most products have not made.
Consent as an artefact
Captured, stored and produceable — with retention set per client and erasure that runs rather than being promised.
Permissions underneath
An audit trail that refuses its own deletion, and access that is enforced below the features rather than inside each one.
Recording consent and retention
A retention date attached to the recording itself, and deletion that happens when it arrives.
What this costs
Every refusal is a feature somebody wanted.
It is worth being straight that this list is a set of things the product will not do for you. Somebody has asked for each of them, usually with a reasonable-sounding justification, and the answer was no.
That is the trade. A system that will answer anything you ask it is more impressive in a demo and harder to defend in a year.
What you receive
Things you can produce when somebody asks.
- The query, before it ran —
A natural-language question becomes an inspectable query. You can see what would have been asked of your data, including for the ones that were refused.
- A refusal in the trail —
Declining is an event like any other. When somebody asks later whether anyone tried, there is an answer.
- Consent as an artefact —
Not a checkbox state — a record you can produce, with the retention that was attached to it.
- Aggregate outcomes, on their own route —
Distributions across self-declared characteristics, readable by the person whose job that is and nobody working a shortlist.
Objections first
What a compliance or legal reviewer opens with.
›How do we know a field was not sent?
Mark one and watch. The check is not a policy statement — it is observable on the request, which is why this page asks you to test it rather than to believe it.
›Who can see the fairness screens?
A compliance persona, not a recruiter. That is enforced by permissions rather than by convention, and the separation is the design rather than an oversight.
›Can we add our own refusals?
Fields are yours to mark. What you cannot do is remove the protected characteristic refusals, which are not configurable on purpose.
›Does a refusal get logged?
Yes. A refusal is an event like any other, which matters when somebody asks later whether anybody tried.
›Is this just a content filter on a general model?
No. The refusal happens before a query is built, and the field exclusion happens before anything is sent. Both sit outside the model rather than depending on it behaving.
Book a demo
Try to make it misbehave.
Bring your legal or compliance lead to this one. They ask better questions than we do, and this is the page written for them.
- Ask it something about a protected characteristic and watch it decline
- Mark a field personal, then show me it was never sent
- Open the fairness screen, and show me who cannot see it
- Put a record under legal hold and try to erase it
- Show me a refusal in the audit trail