A vendor demo for an AI-powered resume screener or interview-scoring tool almost always includes some version of the same line: our AI removes human bias from hiring. It’s a comforting pitch, and it’s rarely backed by anything you can independently verify. Fairness in an algorithm isn’t visible on a dashboard. It has to be tested, documented, and re-tested every time the underlying model changes.
Indian HR teams are buying these tools faster than they’re learning to interrogate them. NASSCOM’s State of Responsible AI in India 2025 survey of 574 senior executives found that unintended bias or discrimination was the fourth most commonly reported AI risk, cited by 29% of organisations already using AI in some part of their business. That’s not a hypothetical concern raised by sceptics. It’s something roughly three in ten companies say they’ve already run into.
Why “Unbiased” is a Claim, Not a Feature
An AI system doesn’t invent bias on its own. It inherits it: from the historical hiring data it was trained on, from the proxies it learns to associate with success, from the assumptions baked into how “good candidate” gets defined in the first place. Amazon discovered this the hard way in 2018, when it scrapped an internal recruiting tool after finding it had taught itself to penalise resumes containing the word “women’s” because a decade of prior tech hires skewed male.
That history matters for Indian buyers because most vendor claims of fairness rest on internal testing the vendor designed, ran, and controls the disclosure of. A tool can pass every test the vendor chose to run and still discriminate on a dimension nobody checked for. Caste-adjacent surname patterns, regional college tiers, employment gaps that correlate with maternity leave, English fluency scores that track socioeconomic background more than job capability. They don’t show up in a generic “gender and ethnicity” bias audit, which is often the only kind vendors run.
The honest position is that no AI hiring tool can claim to be unbiased in any absolute sense. It can only claim it was tested against specific criteria, using a specific method, at a specific point in time. Your job is to find out which criteria, which method, and how recently. The 29% figure cited earlier, from Nasscom’s State of Responsible AI in India 2025 survey of 574 executives, measures organisations reporting the risk. It says nothing about which of them have actually run an audit to check for it. The real gap is likely wider.
The Questions Worth Asking
A useful vendor conversation moves past “is it biased” and into specifics an HR leader can actually verify. The table below breaks down what to ask and why each question matters.
| Question | Why It Matters |
| Which protected attributes were tested, and which weren’t? | Vendors often test for gender and sometimes caste or religion, but rarely for proxies like PIN code, college tier, or career gaps. |
| Who ran the audit, an internal team or an independent third party? | Self-reported fairness testing carries an obvious conflict of interest. |
| What’s the adverse impact ratio by group, and against what benchmark? | A number without a comparison point (like the four-fifths rule used in US employment law) tells you little. |
| How often is the model re-audited after retraining? | A model tested as fair at launch can drift after retraining on new hiring outcomes. |
| Can you export a decision log for every candidate the tool screened out? | This is what makes the tool auditable long after the demo stage has passed. |
| Does the tool’s design account for India-specific proxies? | Bias research from the US and Europe rarely covers caste, regional dialect, or vernacular-medium education, all live issues in Indian hiring. |
| What happens when a candidate disputes an AI-driven rejection? | If there’s no defined escalation path, the fairness claim has no accountability mechanism behind it. |
Bring this list into the vendor conversation and watch how specifically they answer. A vendor with a genuine fairness process will have documentation ready. One that only has marketing language will start talking in generalities about “continuous monitoring” without naming a method, a frequency, or a person accountable for it.
Reading the Answers, Not Just Collecting Them
Getting answers to these questions is only half the work. The other half is knowing what a strong answer looks like versus a rehearsed one. A vendor who says their tool was audited “regularly” hasn’t told you anything. Regularly could mean quarterly, or it could mean once, two years ago. A vendor who can name the audit framework, the date of the last run, and the specific adverse impact ratios by group has given you something you can act on.
The same scrutiny applies to India-specific proxies. Global HR tech platforms built primarily for US and European markets often haven’t tested for the variables that matter most in Indian hiring: surname-based caste signals, tier-2 versus tier-1 college weighting, or resume gaps tied to the Maternity Benefit Act, 1961’s 26-week leave provision. Ask directly whether the vendor’s fairness testing included Indian labour market data, or whether it’s an import from a US-built model with a localisation layer on top.
The Regulatory Backdrop is Catching Up
For years, Indian HR teams evaluating AI vendors on fairness were doing so purely as good practice, with no legal framework compelling it. That’s shifting. The Ministry of Electronics and Information Technology released India’s AI Governance Guidelines on 5 November 2025, setting out a coordination-based framework for AI risk management across sectors, including employment-related use cases.
The Digital Personal Data Protection Rules, 2025, notified by MeitY on 13 November 2025, add a more direct obligation. Under the Rules, organisations classified as Significant Data Fiduciaries must conduct due diligence to verify that their algorithmic software doesn’t pose a risk to the rights of data principals, a category that extends to job candidates whose data an AI hiring tool processes. The substantive provisions phase in over 18 months, with full enforcement landing 13 May 2027, but the direction is clear: “the vendor said it’s fine” won’t be a defensible compliance position much longer.
This is where the earlier questions turn from best practice into risk management. An HR leader who can produce a documented vendor fairness assessment, dated and specific, is in a materially different position than one relying on a sales deck slide. The DPDP Act doesn’t yet create an Indian equivalent of the EU’s right to challenge purely automated decisions, but Significant Data Fiduciaries carry algorithmic due diligence obligations regardless, and enforcement expectations tend to tighten well before enforcement dates arrive.
What This Looks Like in Practice
A mid-sized IT services firm running an AI-based resume screener at scale, processing thousands of applications a month for entry-level engineering roles, shows how this plays out. If a bias audit only checked pass rates by gender, it could look clean while systematically down-ranking candidates from Tier 2 and Tier 3 engineering colleges, a proxy that correlates strongly with caste and regional background in India’s education system. The tool would appear unbiased on paper and still produce a workforce that skews heavily toward a narrow set of institutions and backgrounds.
This is where the algorithmic bias in hiring conversation gets specific rather than theoretical. A generic fairness claim doesn’t tell you whether your tool would catch this pattern. Only a testing methodology that explicitly checks for education-tier proxies would.
HR teams that want a structured way to run this evaluation themselves, rather than relying entirely on vendor self-reporting, can build it around an internal AI bias audit process: pulling outcome data by demographic group, checking it against a defined adverse impact threshold, and repeating the check after every model update rather than once at procurement.
Explainability Changes What You Can Even Check
None of this works if the tool can’t explain its own decisions. A model that outputs a candidate score with no visibility into which inputs drove that score can’t be meaningfully audited for bias, because there’s no way to trace a discriminatory outcome back to its cause. This is the explainable AI problem, and it belongs early in the vendor conversation. Ask whether the tool can produce a reason code for every rejection. Without one, every other fairness claim rests on trust rather than verification.
In the End…
The real question isn’t whether an AI tool is unbiased. It’s whether your organisation can prove that it is being used fairly.
That means looking beyond the vendor’s assurances and asking for evidence: what was tested, which Indian hiring proxies were considered, who conducted the audit, when it was last done, and what happens when the model changes. If the answers aren’t specific, the fairness claim isn’t either.
AI hiring tools may make decisions at scale, but accountability still sits with the people who choose, deploy, and oversee them. For HR, that makes bias testing less of a one-time procurement exercise and more of an ongoing part of responsible technology governance.
Don’t buy the promise of unbiased AI. Buy the ability to verify it.
FAQs
Can AI hiring tools really be unbiased?
No AI hiring tool can reasonably be treated as universally unbiased. Fairness depends on factors such as the data used, the variables considered, the testing methodology and how the model is updated. HR teams should ask vendors for specific evidence of how bias was tested rather than relying on a general claim of being “unbiased.”
What should HR ask an AI hiring vendor about bias?
HR teams should ask which protected attributes and potential proxies were tested, who conducted the audit, when it was last completed, what adverse impact was found, how often the model is re-audited and whether decision logs can be accessed. For Indian hiring, teams should also ask whether testing considered factors such as college tier, regional background and career gaps.
How often should an AI hiring tool be audited for bias?
A bias audit should not be treated as a one-time exercise. HR teams should understand how frequently the vendor tests the model and whether a new audit is conducted after significant retraining or model updates. A tool that passed a fairness assessment at launch may produce different outcomes after its underlying model changes.
Why do AI hiring tools develop bias?
AI systems can inherit patterns from historical hiring data and learn associations between candidate characteristics and past hiring outcomes. Bias can also enter through the variables selected by the model or through proxies that correlate with protected or sensitive characteristics.
What is an adverse impact ratio in AI hiring?
An adverse impact ratio compares the selection rate of one group with that of a comparison group. It can help HR teams identify whether a hiring system produces materially different selection outcomes across groups. The figure needs to be interpreted against an appropriate benchmark and testing methodology.
Why is explainability important when evaluating AI hiring tools?
Explainability helps HR teams understand why a tool produced a particular candidate score or rejection. Without visibility into the factors influencing a decision, it becomes much harder to investigate potentially discriminatory outcomes or determine whether a vendor’s fairness claims can be independently verified.
Should HR rely on a vendor’s own AI bias audit?
Vendor testing can provide useful information, but HR teams should understand who conducted the audit, what methodology was used, which groups and proxies were tested and when the assessment was performed. Independent testing or internal validation can provide an additional layer of scrutiny where appropriate.
