Is AI screening biased?
Any system that scores people can encode bias, and a vendor claiming otherwise is the one to distrust. Haira's controls are procedural rather than statistical: you set the ranking priorities, so the criteria are visible and yours; every candidate is interviewed in the same format and scored on the same four dimensions with the same rubric, on questions personalised to their resume; and every interview score ships with the transcript so a decision can be reviewed and overruled. By default Haira recommends and a person decides; Auto-Process, if you switch it on for a job, rejects and advances candidates on the score.
What is actually in place
The criteria are yours. Ranking priorities are set per role by the hiring team and can be read back afterwards, so a disputed shortlist can be examined against the criteria that produced it rather than against a model's private judgement.
The interview has the same structure for everyone in a role: the same four scored areas and the same scoring rubric. The questions themselves are personalised — prepared from the same job description and each candidate's own resume — so two candidates are not asked identical questions.
Scores rest on evidence in the resume. A skill is not assumed from a similar tool or a senior-sounding job title, because guessing from surface similarity is one of the ways proxies for background creep into a score.
Every recommendation is reviewable. The transcript is kept with the score and its reasons, so a human can see why and disagree; on the Growth and Enterprise plans the interview can also be recorded.
What is not in place, stated plainly
There is no bias audit. No third-party fairness assessment has been run against this system, and no adverse-impact analysis exists in the codebase.
Haira asks candidates for no demographic data, but a date of birth that appears on a resume is stored with the parsed profile, and nothing measures outcomes across groups. That means Haira also cannot show you that it is unbiased — the honest position is that the controls above are reasonable design, not evidence of a fair outcome.
If you need documented fairness testing for a regulator, an internal audit or a jurisdiction that requires it, Haira does not currently provide it. Ask us rather than assume, and if the answer matters to your hiring, weigh it before you buy.
What to do about it on your side
Look at your priorities before you trust the order. Priorities that put education and company tier high will rank for background rather than capability, and that is a choice you made rather than something the model did to you.
Read the transcripts of the candidates the system ranked low. That is the cheapest bias check available to you and it needs nothing from us, because the transcript ships with every score.
Keep the decision with a person. Auto-Process is off unless you switch it on for a job; once on, it rejects and advances candidates on the score without anyone reviewing each one. Left off, every score is a reading order and the recommendation is there to be overruled.
Where this stops
This page describes controls, not proof. Without demographic data and outcome measurement, neither Haira nor any similar tool can demonstrate an absence of bias, and nothing here should be read as a compliance statement.
How do you stop the AI from being biased?
Three things. You set the ranking priorities for each role, so the criteria are yours and visible rather than a black box. Every candidate for a role is interviewed in the same format and scored on the same four dimensions. And every score comes with the full transcript (and the recording, on plans that include recording), so any decision can be reviewed and challenged after the fact. By default Haira scores and recommends, and your team decides. Auto-Process, if you switch it on for a job, rejects and advances candidates on the score.
Last checked against Haira's own implementation on 2026-09-30.