Recagn
Sign in

Guide · Fairness

AI bias in hiring in India: where it comes from and how to limit it

Bias gets into a hiring tool through what it is told to look for, what it is given to read, and who acts on its output. Each has a check.

Part of DPDP Act and recruitment: what agencies must do with candidate data

Where does bias get into AI hiring?

Bias gets into an AI hiring tool in three ways: through what it is told to look for, what it learns from, and what it reads. Each can turn a difference between groups of people into a difference in scores, even when nobody asks the tool to look at a group.

The data side, from consent to accuracy, is in the DPDP Act and recruitment.

What it is told to look for

A tool scores against the requirements it is given. Some requirements stand in for a group without naming it: a graduation year for age, a list of colleges for background, a pin code for community. A tool that applies them to every candidate applies the bias to every candidate too.

What it learns from

Some tools are trained on who was hired before. In a review of 18 vendors of algorithmic pre-employment assessments, Raghavan, Barocas, Kleinberg and Levy (arXiv, December 2019; ACM FAT* 2020) point out that predicting success from current employees pushes a tool towards candidates who resemble the people already hired. Of those 18 vendors, 15 referred to "bias" at least in the abstract, and 7 discussed adverse impact or compliance explicitly.

What it reads

Some inputs carry identity on their face: a name, a photo, a voice. Speech-to-text is the less obvious one. A Stanford-led study published in PNAS in March 2020 tested five commercial speech recognisers on interviews in five US cities and found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers. When a tool scores a transcript, a misheard answer becomes the evidence. That study covered US English; we found no comparable study of Indian accents to cite.

What does Indian law say about AI bias in hiring?

MeitY's AI governance guidelines of November 2025 say a separate law to regulate AI is not needed now, and existing laws can address many of the risks. As an illustration, they list the disability, transgender persons, wages and SC/ST atrocities laws against discrimination in hiring with AI recruitment tools. This is not legal advice; check with counsel.

The India AI Governance Guidelines name bias and discrimination as a risk, with the example of inaccurate data used to decide someone's future employment. Their Annexure 4, which MeitY marks as illustrative, maps "discrimination in hiring decisions using AI recruitment tools" to four Acts: the Rights of Persons with Disabilities Act, 2016; the Transgender Persons (Protection of Rights) Act, 2019; the Code on Wages, 2019; and the Scheduled Castes and the Scheduled Tribes (Prevention of Atrocities) Act, 1989. This guide names them as MeitY lists them and does not summarise what each says.

The DPDP Act, 2023 has no provision on AI, profiling or automated decisions, and gives candidates no right to have a person review a decision. One duty does touch screening: where personal data is likely to be used to make a decision that affects the person, the agency must keep it complete, accurate and consistent (s.8(3)). The Act's duties start on or about 13 May 2027.

What that means for an agency, duty by duty, is in the DPDP Act and recruitment.

Is AI hiring fair? How to check a tool before you use it

AI hiring is as fair as its rubric, its inputs and the person who decides, and you can check all three before you start. Ask the vendor what the AI is given, what it is told to ignore, how a score becomes a recommendation, and who can reject a candidate. You decide whether the answers are good enough for your clients.

AskA good answer shows
Is every candidate for a role scored against the same, fixed criteria?A rubric locked before invitations go out, with a new version each time it is edited
Are the criteria job-related?Each criterion and hard gate traceable to the brief; no graduation year, college list or pin code standing in for a group
What is the AI given?A list of its inputs, with name, contact details and photo absent
What is it told to ignore?Protected attributes named in its instructions
How is a voice or video answer read?Whether it is transcribed, and whether you can check the transcript and recording beside the score
How does a score become a recommendation?A fixed, published rule, not a model's verdict
Who can reject a candidate?Only a person, with the AI's recommendation and your decision recorded separately
Does it learn from past hiring decisions?What it learns, from whose data, and how far that can move a result
Has anyone measured outcomes by group?Who ran the check, on which groups, and when

Source: questions written for this guide

Check your own hard gates

The vendor is half of it. The requirements you agree with a client are the other half, and a hard gate removes people outright. Here are four gates that can stand in for a group, each with a job-related version.

A gate that can stand in for a groupA job-related version
Graduated in 2020 or later2–5 years of relevant work
From a listed collegeCan show the named skill the role needs
Lives within 5 km of the officeCan be at the office by 9:30 on site days
"Good English accent"Explains a named task clearly to a client in English

Source: illustration written for this guide

Compare who moves on

Once a role has enough candidates, compare how often each group you can see moves on. In the US, the federal Uniform Guidelines (29 CFR 1607.4(D)) generally treat a group's selection rate below four-fifths (80%) of the highest group's rate as evidence of adverse impact. That is a US rule of thumb, not Indian law, but the arithmetic works anywhere.

It needs group data, which an agency should hold only for a purpose its notice to candidates names. And a handful of candidates can't show a pattern either way.

What can a rubric and human review not fix?

A rubric applies whatever it is given, so a biased requirement is scored faithfully, and a reviewer can bring their own bias to an override. Consistency makes bias visible and checkable; it doesn't make it go away. You decide which requirements are real, and you own every override.

  • A biased gate or weight, applied to everyone, is still biased.
  • A transcription error reaches the score unless someone reads the transcript.
  • Overrides are where personal bias gets back in. Writing down a reason for each one makes them reviewable later.
  • A few candidates per role can't show a pattern between groups.
  • The client's brief can carry its own bias, and the recruiter is the first person who can question it.

How Recagn does this

Recagn is our product. This section shows what each of its AI steps is given, what each is told, and where the decision passes to you, from its code.

How the rubric and human review limit bias

StepWhat the AI is givenWhat it is toldWhat stays yours
Drafting the rubricThe client's job descriptionUse only job-relevant signals and don't encode protected attributes, naming age, gender, ethnicity, religion, marital status and graduation year as a proxyEditing and locking the scorecard, which Recagn calls a rubric
Reading the CVThe CV's full text, name includedExtract only job-relevant signals and leave out name, age, gender, photo, ethnicity, religion, marital status and date of birth; name and email go to a separate contact record, not the profileNothing to decide here; the profile feeds the next row
Matching a CV to the roleSkills, years of experience, education, projects, locations, work authorisation, job type and work type, parsed from the CV; no name, email, phone, job title, employer or work historyNot to use or infer protected attributes, naming name, age, gender, ethnicity, photo, graduation year and university prestigeWho gets invited
Scoring the interviewThe job description, the locked rubric and the answers: text as typed, voice and video as transcripts, uploaded documents; no name or contact fieldsScore each criterion 1 to 5 using only the answers to its linked questions, quote the evidence, and not use or infer protected attributesWhether each score holds
Working out the totalNo AI: Recagn's code works out Σ(score × weight) ÷ Σweight, then applies the rubric's cut-offsNothing; it is arithmeticApproving or overriding the recommendation, with a written note
The decisionNothingNothingSelected, Maybe or Rejected; only your Rejected moves a candidate out

Source: Recagn code (prompts rubric-draft, parse, match-assess@2, score@2), read 17 Nov 2026

Two limits are in that table. The matcher is given the education entry, college included, and is told not to weigh university prestige. The scorer is given no name or contact fields, but the answers are not redacted, so a name a candidate says or types still reaches it.

One rubric, one rule, for every candidate

Recagn scores every candidate for a role against the same locked rubric. Editing and re-locking it makes a new version, and each candidate keeps the version that was current when they finished. The total and the recommendation come from Recagn's code by a fixed rule, so the same scores always give the same recommendation.

Each score sits beside the answers, with transcripts and recordings for voice and video, so you can check a misheard word against what was said. A candidate's own history on earlier roles can move their place in the match ranking by at most 15 points, with the reason shown; it never changes an interview score or the recommendation.

Follow-up questions are off unless you switch them on. When they are on, a typed answer too vague to score can get one clarifying question, and that question may not add a criterion, raise the bar or ask about protected attributes.

The arithmetic, gates and cut-offs are worked through in how AI interview scoring works.

Where your decision comes in

One evaluation, as the recruiter sees it

Look at criterion 2's quote, then at the empty decision: Recagn recommends, and the decision stays blank until you make it.

Ananya R. Senior Java Developer, Pune

Hard gate: Notice period ≤ 30 days

  1. 1Spring Boot services in production, 3+ yearsSpring Boot services in production, 3+ yearsweight 30

    I've run four Spring Boot services in production. Q1 · 0:35

    I've run four Spring Boot services in production. Q1 · 0:35

  2. 2Owned a payments or ledger serviceHandled payment reversals in productionweight 25

    Not establishedNothing in her answers says she owned the service. Not established, and not disproved. Q3 · whole answer

    I checked the reconciliation log first, then the idempotency key on the retry. Q3 · 0:42

  3. 3Kafka or RabbitMQ in productionKafka or RabbitMQ in productionweight 15

    I mostly maintained the settlement consumers someone else built. Q4 · 1:05

    I mostly maintained the settlement consumers someone else built. Q4 · 1:05

  4. 4Code review and mentoringCode review and mentoringweight 30

    I pair with the two juniors on every pull request. Q5 · 0:20

    I pair with the two juniors on every pull request. Q5 · 0:20

3.353.85Recagn recommends MaybeQualified

Your decision: not made yet

Illustrative example · Senior Java Developer, Pune · invented candidate · Recagn's scoring rule

Before you can set a decision, you approve Recagn's recommendation or override it with a written note. Recagn records the recommendation at the moment you decided and whether you agreed with it. Recagn recommends. Only you can reject a candidate.

Proctoring and integrity findings sit beside the score for you to judge. They never enter the score, the ranking or the client's report. You can also give one candidate extra time (+25%, +50% or +100%) or an extra attempt, with a reason only your team sees.

What candidates are told

What candidates read before they agree is on what candidates see and agree to.

Recagn drafts, reads and scores against the rubric you lock. The requirements, every override and every decision stay with you.

See a worked example

Questions people ask

Does Recagn check its results for bias between groups?

No. Recagn asks candidates for no gender, caste, religion or age, and its candidate profile has no fields for them, so it has no group data to compare selection rates by. It gives its AI no fields for those details, tells it not to use or infer them, scores every candidate for a role against the same locked rubric, and leaves every rejection to you. If your agency collects group data lawfully, the four-fifths comparison in this guide works on your own outcomes. Recagn recommends. Only you can reject a candidate.

Does HireVue's AI analyse faces or tone of voice?

No, going by HireVue's 2026 Explainability Statement, read 17 Nov 2026: its AI relies only on what candidates say in their answers, not on facial expressions, body language, surroundings or tone of voice. Recagn's scorer reads the answers: typed text, transcripts of voice and video answers, and uploaded documents, including PDFs and images. It is given no video. Two separate checks do look at images. Frames from video answers go to a visual-behaviour check for signs such as reading from off-screen. When webcam proctoring is on, its stills go to Recagn's proctoring rules, which look for signs such as a missing or second face and, with ID check by face match, compare the face with the candidate's selfie. The signals are described in how interview integrity works. Their findings sit beside the score for you to judge and never enter it. You approve or override each recommendation before you decide.

Does a person's review cost extra on Recagn?

No. Reviewing, approving and overriding a recommendation cost nothing. The interview itself uses credits: one for a text or voice interview, three if any answer is video, plus more when webcam proctoring is on. Without webcam proctoring, on Starter, monthly, that is at most ₹50 per text or voice interview (₹49.98; ₹58.98 with 18% GST) and at most ₹150 per video interview (₹149.95; ₹176.94 with 18% GST).

Is removing names from CVs enough to stop bias?

No. Other details can stand in for a group: a graduation year, a college, a home address, the way someone speaks. That is why the criteria themselves need checking, and why a person should read the evidence before any decision.

See one real brief scored from rubric to shortlist.