AI does not decide a candidate's fate

In the recruiting app we run, a model can summarise CVs, match them against the job criteria and draft a first assessment. Technically it would be very easy to wire that assessment to an automatic reject: below a certain score, the candidate gets a rejection message straight away. We deliberately do not. AI never rejects a candidate on its own. A recruiter opens the candidate's file, reads, weighs, then decides, and that decision is recorded in their name.

There are three reasons. First, being turned down for a job is a decision with a significant effect on someone's life. Article 10(1) of the PDP Law gives Personal Data Subjects the right to object to decisions based solely on automated processing, including profiling, that have legal effects or a significant impact. The full explanation of this right is in the lesson on automated decisions and people's rights. Second, a model can misread a CV: two-column layouts, work histories with gaps, or job titles from another industry are often misunderstood. Third, a CV can carry text deliberately written to sway the model, and only a person reading it can catch that.

This does not mean AI is banned from screening. AI is useful for summarising, flagging evidence that matches the criteria and drafting interview questions. What we guard is the line between a suggestion and a decision. A suggestion may come from a model. A decision comes from a person who is authorised and can be asked why. This principle is called human in the loop.

Human review only means something if it is real. A recruiter who receives 300 suggested rejections and approves them all in ten minutes is not deciding; they are just a button. In practice such a decision is still automated. Real review has four features: the recruiter opens the original file, not only the summary; they have a reasonable amount of time per candidate; they are authorised to decide differently from the model's suggestion; and they write the reason for their decision in their own words.

An easy way to test it: look at how often recruiters disagree with the model. If they never do, there are two possibilities. Either the model is perfect, which is rare, or the recruiters are not really reading. The second is far more common.

The same applies to psychometric scores. A personality or reasoning test gives a picture of one side of a candidate on one day, with limited precision. A psychometric score is one input among several, not the decision. It is read together with evidence from the CV, structured interview results, work samples and references. A score threshold that knocks candidates out directly makes the test the sole judge, and that is the same as an automated decision.

A score can also be wrong because of how it is computed, not only because of the candidate. Some test sections, such as a forced-choice block, have a mathematical property that makes their scores not comparable across people like ordinary scores. If the composite formula is wrong, every candidate can be pulled to the middle. That is another reason not to hand the decision to one number.

The decision record also protects candidates and recruiters. If a candidate objects or asks why they did not move forward, the organisation can show who decided, what material was read and what reason was written. A record with only the model's score cannot answer that question. Keep this record according to the hiring round's retention note, no longer than needed.