Algorithmic bias in hiring is a systematic skew in an automated screening or ranking tool that leaves one group of candidates worse off for reasons unrelated to the job. It usually comes from training on past hiring decisions or from inputs that stand in for gender, caste, religion, age or disability. It is a different problem from the human judgement that diversity hiring programmes target, because the bias repeats at scale on every CV.

How it gets in

The textbook case is Amazon. Reuters reported in October 2018 that Amazon had built an experimental CV-rating engine trained on about ten years of applications, mostly from men, and found by 2015 that it was not rating candidates for technical roles in a gender-neutral way. The project was dropped. Nobody coded “prefer men”; the model learned it from history.

In India the usual proxies are local:

• College name or tier: tracks family income and region as much as ability.
• Pin code or locality: can stand in for community or religion in segregated cities.
• Career gaps: penalise women returning after maternity or caregiving.
• Surname, language, graduation year: signals of caste, region and age.

Errors upstream add to it: if resume parsing mangles certain name formats or layouts, those candidates rank low for reasons no one sees.

Auditing a tool

The basic check is the impact ratio: each group’s selection rate divided by the rate of the best-treated group. Say an automated screen passes 120 of 400 men (30%) and 36 of 200 women (18%). The ratio is 18 / 30 = 0.60. The US “four-fifths” guideline from 1978 treats anything below 0.80 as a sign of adverse impact; it is a rough screen, unreliable on small numbers, not a verdict. New York City’s Local Law 144, enforced since 5 July 2023, requires a yearly independent bias audit reporting these ratios before such tools are used there. India has no equivalent, but the Code on Wages already bars sex discrimination in recruitment for the same or similar work, and a tool is no defence. Run the check on your own applicant data before go-live and every quarter after.

Frequently asked questions

Does removing gender and name fields fix it?

Rarely on its own, since proxies carry the same signal; blind recruitment helps human reviewers, but a model still needs outcome testing.

Can we collect gender or category data to run the audit?

Yes, kept separate from the screening inputs and gathered with clear notice; plan it against the DPDP Act duties due from May 2027.

Is the vendor responsible, or are we?

You run the hiring, so you carry the risk with candidates; in the US, the Mobley v. Workday lawsuit is testing whether a screening vendor can also be held liable.

An audit tells you there is a skew; finding why needs a tool that can show its reasoning, covered under explainable AI in HR.