AI resume screening can reduce bias when it ranks on skills and experience and keeps name, photo, and contact masked until unlock. It can worsen bias when it is trained on biased past hires or uses names in the score. For Indian staffing desks, the honest stack is masking plus must-have gates — not a claim that the algorithm is “fair by default.”
A model that never sees a name cannot prefer a name. A model trained on last year’s shortlists can prefer last year’s habits.
Two ways screening AI can shrink bias on a desk
Unconscious bias in Indian recruitment rarely looks like a policy. It looks like a skip: a name that “doesn’t feel client-facing,” a photo that triggers a guess, a college that is not on the hiring manager’s mental list. Ranking before identity is visible changes the order of that decision. How masking works in practice is covered in how masked profiles fix unconscious bias in recruitment.
If the match percentage is built from resume text and the JD — skills, experience, location, CTC, notice when present — the first sort is about fit. Names do not belong in that calculation.
Profiles stay masked until unlock. The recruiter sees a code, role, company, location, and score. Unlock is a deliberate step, not the default view of the pile.
That combination does not make humans unbiased. It makes the first pass harder to contaminate. In a market where recruiters still spend 4–6 hours a day triaging, fatigue is itself a bias amplifier: late-pile resumes get a worse read. Ranked, masked review at least puts fit at the top before tired eyes take over.
Two ways the same AI can make bias worse
Vendors who say “AI is objective” are selling a slogan. Objectivity is a design choice.
| Design choice | What happens | Risk on an Indian desk |
|---|---|---|
| Train on past placements | The model copies who you used to send | If last year’s clients favoured one campus or community, so will the rank |
| Use name, photo, gender cues | Identity becomes a feature | The score inherits the skip you were trying to remove |
| Equal-weight every skill | Pedigree keywords can outscore a missing core skill | A brand-name intern with AWS listed outranks a Java developer missing Java on paper — until a 55% cap exists |
| Mask + must-have gates | Identity delayed; deal-breakers enforced | Bias is reduced at review time; false “strong” scores are capped |
College, company brand, and city can still carry assumptions after a name is hidden. Masking is necessary. It is not sufficient. Legal questions around masking and screening sit in is candidate masking legal, and does it reduce bias and is AI resume screening legal in India — neither of those pages is legal advice.
Masking plus must-have gates is the practical stack
Fairness and quality are not two projects. A must-have gate stops a candidate who cannot do the job from looking “diverse and strong” on a padded skill list. Masking stops a candidate who can do the job from being skipped because of a name. You want both. One without the other is either unfair ranking or polite ranking of the wrong people.
Ask three questions: Does the score use name or photo? Are profiles masked until unlock? If a must-have is missing, is the score capped (55% in a gate-based model) so good-to-haves cannot compensate? If the answers are fuzzy, the bias story will be fuzzy too.
What agencies should audit before they trust a percentage
Run a week of dual review: recruiters shortlist from masked, ranked lists, then compare who they would have called from the raw Naukri pile. You are not looking for a published study. You are looking for your own pattern — who rose, who dropped, and whether the drops were skills or identity. Log unlocks. If unlocks cluster on a name pattern after the score was already high, the bias moved from screening to the unlock click. That is still a process you can see, which is more than an unranked inbox gives you.
AI screening reduces bias only when identity is kept off the score and off the first screen, and when the job’s real gates still bite. Anything else is automation of the same desk, faster.
Rank on skills. Unlock on purpose.
TeamOB - SafalHires masks profiles until unlock and scores against the JD — so the first pass is fit, not a name.
Free trial