Using AI tools in hiring. The benfits, pitfalls, and legal risks.

Most employers are using AI for most stages of recruitment and hiring. But how dependable are AI tool? Do they amplify or minimise bias? And how do they measure up to equality laws?

Most would agree that hiring is the most important function of any business or organisation. Historically it’s been a messy, slow, tedious, and inconsistent process that rarely guarantees satisfaction. Now, imagine handing over that process to a machine for what seems like the same result in terms of levels of satisfaction with the outcome! What a relief. AI has promised to fix the two thorniest issues that have haunted hiring – efficiency and fairness. You can hire efficiently and fairly because unlike human recruiters, AI does not have feeling or preferences or biases. That is the promise. But has it really delivered on both? So far, the evidence seems to suggest otherwise.

The truth is that an algorithm trained on your organisation's past hiring patterns becomes very efficient at reproducing your organisation's past hiring, including every pattern you were hoping to leave behind. Bias does not disappear. It gets conveniently covered by a veneer of technological objectivity.  

The evidence of the real impact of AI on hiring is increasingly less anecdotal. This year, researchers at Stanford, Chapman and Northeastern published what is thought to be the largest independent analysis of algorithmic hiring ever conducted, covering over four million applications from 3.4 million people, across some 1,700 roles, all screened by a single third-party AI tool. Their finding should give every people leader pause. The tool could sail through a bias audit at the aggregate level while still systematically screening out Black and Asian candidates for specific roles. More than one in four applications from Black job seekers were routed towards outcomes that would trigger regulatory scrutiny in the US. The researchers gave the underlying problem a name - "algorithmic monoculture." Because a small number of vendors now supply the hiring algorithms behind a large share of employers, a single flawed model doesn't simply affect outcome in one organisation, it pervasively shapes who gets rejected across an entire market or sector.

A separate large-scale experiment on leading AI models found they tended to favour female candidates while disadvantaging Black male applicants with otherwise identical qualifications. And tools that analyse video interviews or gamified assessments have repeatedly been shown to penalise neurodivergent candidates, especially people with autism or ADHD, for not responding to questions in a "standard" way.

While most people leaders to check if tools have been audited, I believe that check needs to be more rigorous than the current almost tick box exercise. The Stanford study showed that a tool can pass an aggregate audit and still discriminate role by role. Thus, the question should be how the tool was audited and on which population or groups and to what level of granularity. Does the audit look at individual roles, or does it pool everything together until the disparities average out or vanish altogether?

In the US, the regulators are catching up. New York City already requires bias audits and public reporting for automated hiring tools. Colorado's AI Act arrives in 2026. California has firmed up how existing anti-discrimination law applies to these systems. And the Mobley v. Workday case in the US is testing whether a software vendor itself can be held liable for discriminatory screening.

Here in the UK, the Equality Act 2010 already prohibits indirect discrimination — and indirect discrimination doesn't require anyone to intend harm. A neutral-looking tool that disproportionately disadvantages a protected group is a big legal risk sitting inside your applicant tracking system (ATS) right now.

So, what should people managers and employers do?

1. Treat AI hiring tools as regulatory and legal risk decision, not just an IT procurement one. Consider the risk of embedding discrimination unintentionally in your hiring process.  The risk could outweigh any benefit you gain from the speed and “efficiency” of using AI tools.

2. When procuring an ATS system, demand role-level and intersectional audit evidence, not a single reassuring headline figure.

3. Check for and beware of the monoculture trap. If you're using the same tool as everyone else, you may be inheriting and exacerbating the same biases.

4. Keep meaningful human oversight, especially for the groups a model is known to disadvantage.

5. Monitor your own outcomes over time. The vendor's audit is a starting point, not a warranty. Assess the outcomes of your hiring process or of the industry and sector. If a pattern emerges around any protected characteristic, investigate further and seek assurances that the tool is not responsible.

None of this means rejecting or abstaining from using AI tools in hiring. Used carefully, IT tools can reduce some human bias. Remember, asserting that "the computer decided" is not submissible as a legal, moral or commercial argument. The organisations that benefit from using AI tools will be the ones that bring rigour to assessing inclusivity and objectivity of their AI tools and their organisational culture.

If you are concerned about your legal risk and exposure and would like a second pair of eyes and rigorous assessment on how you're deploying AI tools in your hiring processes, get in touch. advice@sias-uk.com