Is an AI screening interview fair to the candidate?

The question is what the AI is allowed to decide

“Is AI screening fair” is not answerable as asked, because it covers two products that have almost nothing in common. One asks a set of questions, records what was said, and hands it to a person. The other reads a face on video, scores confidence or enthusiasm from tone, and returns a ranking nobody can explain. Both get called AI screening. Only one of them is a question about fairness in any interesting sense.

The useful version of the question is narrower. What is the system allowed to decide on its own, what evidence is that decision made from, and can a person see the reasoning afterwards. Nearly everything that matters follows from those three.

Where it is genuinely fairer than the alternative

The comparison that matters is not against an ideal interview. It is against what actually happens to applicant number one hundred and forty on a Friday afternoon.

Everyone gets asked. When two hundred people apply, most are rejected on a CV without a conversation. A first round that runs at volume means the person with the badly formatted CV and the right experience gets the same four questions as everyone else. That is a larger fairness gain than anything downstream of it, and it is the one most often left out of the argument.

The questions do not drift. A human first round at hour six is not the same interview as at hour one. Consistency is the thing automation is genuinely good at.

There is a transcript. A phone screen leaves a few lines of notes written from memory by someone who has since done nine more. A written first round leaves exactly what was said, which is reviewable, disputable and auditable in a way notes never were.

No calendar. Scheduling filters hard on people with inflexible jobs, caring responsibilities or a different time zone. Removing the slot is quietly one of the more equalising things you can do to a first round.

Where it is worse, and this part is real

Language. Any text-based screen rewards written fluency. For a role where writing is not the job, that is a bias the format introduces on its own, and it is not fixed by anything in the model. It is a reason to keep questions concrete and to read the transcript rather than trusting a score.

Scoring what was not asked. A system that infers personality, confidence or culture fit from phrasing is making a judgement it has no evidence for. Video-based inference of emotion or enthusiasm is the clearest example, and it is the practice most likely to be indefensible.

Silent rejection at scale. A human doing this badly rejects fifty people. A system doing this badly rejects fifty thousand, identically, and nobody notices because there is no moment where a person felt uneasy about it.

Candidates who will not use it. Some people will not do an interview with software, and they are entitled not to. If there is no route to a human, the screen is a filter on that preference too.

What to check before you put one in front of candidates

Are candidates told, before the first question? Disclosure after the fact is not disclosure. In several jurisdictions this is becoming a legal requirement rather than a courtesy, and it is worth checking what applies where you hire.

Does it reject anybody by itself? The safest design is one where the software never issues a rejection. It produces a profile, a transcript and a ranking, and a person decides. Ask the vendor directly, because “automatically shortlists” and “automatically rejects” are often the same feature described from different ends.

Does every score come with its evidence? A number with no quotation under it cannot be checked by you and cannot be explained to a candidate who asks.

Does it know the criteria while it is interviewing? This one sounds technical and is not. A system shown the must-haves during the conversation will steer toward them, and a candidate who does not match starts being asked the same thing in different words until they give up. Keeping the criteria out of the interview and using them only to rank afterwards produces a fairer conversation and a more honest transcript.

What does it need from the candidate? A camera, a quiet room and good upload speed are requirements dressed as defaults. Text on a mid-range phone is not.

Where we stand

It would be dishonest to write this without saying that we sell one of these. Blue runs first-round screening interviews, so treat the list above as the standard we are asking to be held to rather than a neutral survey.

For what it is worth: candidates are told they are talking to an AI before the first question and have to agree. The interview is typed by default, with no camera unless the employer offers a video option and the candidate chooses it, and no scheduled slot, and it can be closed and resumed. Blue is deliberately never shown the role’s must-haves while it is interviewing, for the reason in the checklist above. Every score carries its reasoning and the full transcript underneath it. Blue does not reject anybody. The recruiter decides who moves forward.

If you are weighing this up because of volume rather than principle, the arithmetic behind that pressure is in how to screen two hundred applicants for one role.