Amazon’s engineers spent years building a tool that would score CVs automatically. When they tested it against real outcomes, they found it had been downgrading applications that mentioned women’s groups or all-women colleges. The model had been trained on a decade of hiring decisions from a male-dominated technical function. It had learned that male-coded signals correlated with success and encoded that correlation into its scoring. Amazon scrapped the tool before it was used at scale.
That was 2017. The legal landscape around AI-assisted hiring has moved considerably since then.
What is an AI hiring discrimination claim?
A discrimination claim against an AI hiring tool is a legal challenge that an automated system produced outcomes that disadvantaged candidates from a protected group: age, race, disability, or sex. Under the Equality Act 2010, the mechanism of selection doesn’t change employer liability. If a hiring process has a discriminatory effect, whether produced by a manager or a piece of software, the employer remains responsible.
The Mobley v Workday case, proceeding as a class action in the US, illustrates how the liability question is being framed. An African-American man over 40 with disabilities alleges that Workday’s AI-driven screening tools rejected him across more than 100 roles at multiple employers. In 2024, a court held that Workday could be treated as an “agent” of those employers for discrimination purposes. The employers had outsourced the screening; they hadn’t outsourced the liability.
EEOC v iTutorGroup was less ambiguous. Online tutoring firms had configured their recruitment AI to automatically reject female applicants aged 55 or older and male applicants aged 60 or older. The EEOC settled the case in August 2023 for $365,000 and secured an injunction to stop the age-based screening.
The pattern across these cases is consistent: a tool that looks neutral applies a rule or learns a bias that produces lower selection rates for a protected group. The claim follows the pattern, not the intention behind the tool.
Why does this matter for an owner-managed business?
Owner-managed businesses rarely build their own AI hiring tools, but that doesn’t reduce the risk. Modern HR and recruitment platforms bundle AI scoring and shortlisting features that switch on by default. Founders using platforms like Workday, LinkedIn Recruiter, or video-interview services may have no clear picture of how their shortlists are being produced. Under the Equality Act 2010, that gap doesn’t reduce your legal exposure.
Surveys suggest almost 70% of companies now use some form of AI in hiring, from CV screening to chatbots and automated interviews. Many of these features come packaged into off-the-shelf software bought without scrutinising what the AI components are actually doing in the background.
The ICO is explicit that organisations using AI in hiring must meet data protection obligations under UK GDPR, including fairness, transparency, and honouring candidates’ rights under Article 22. Where AI significantly influences a hiring decision, candidates have the right to request human review. Many businesses deploying these tools are unaware of that obligation, which means they’re not meeting it.
The Equality and Human Rights Commission has also warned that automated decision-making in employment can lead to discrimination if not carefully designed and monitored, and that employers should consider equality impact assessments when adopting new technology. That expectation doesn’t scale away for smaller businesses.
Where will you actually encounter it?
The three main points of liability in an AI-driven hiring pipeline are targeted job advertising, CV screening, and video-interview scoring. Advertising algorithms can direct roles to narrower demographics than intended. Screening models trained on historical data replicate whoever was hired before. Video-interview systems that score speech and expression can disadvantage candidates with disabilities or those outside the model’s training distribution.
The Amazon case is the most-cited example of biased CV screening: the model learned to treat male-coded signals as positive because its training data was predominantly male. In D.K. v HireVue, filed in 2024, an Indigenous, deaf employee alleges that HireVue’s video analysis system penalised her speech pattern and signing, raising both disability and race discrimination claims. HireVue’s tools have been marketed to assess facial expressions, tone of voice, and word choice as predictors of job fit.
NIST’s research on face-recognition systems has documented significantly higher error rates for certain demographic groups, particularly women, Black and East Asian people. That evidence is now being cited in video-interview cases where similar systems are used to score candidates on expression or appearance.
On job advertising, enforcement actions have shown that optimisation algorithms can serve high-paying role ads predominantly to younger or male users, not because anyone programmed that preference, but because the system learned from historical engagement patterns and replicated them.
When does the risk actually apply to you?
The size of your business doesn’t determine your legal exposure: the Equality Act 2010 applies regardless of headcount. What shapes your risk is how much of your hiring process is being filtered or scored by an AI system you haven’t audited. A fifteen-person firm using an AI shortlisting tool carries real liability. A larger firm reviewing every application manually carries far less.
Legal commentary on AI-bias litigation identifies two consistent factors where liability is established: the employer didn’t ask the vendor how bias was addressed, and the employer never checked whether the tool produced substantially different selection rates across protected groups. Both are gaps an owner-managed business can close with modest effort.
The risk is also lower if you’re using AI only as a minor input, such as grammar suggestions for job descriptions, rather than a system that ranks or eliminates candidates. The cases that have attracted liability involve tools materially influencing who is seen and who is not.
New York City’s Local Law 144, effective from 2023, now requires annual independent bias audits of automated hiring tools and public disclosure of audit summaries. No equivalent UK law is in force yet, but the ICO’s published work on AI assurance signals that regulators are moving in that direction. The practical question for a founder isn’t whether to wait for regulation, but what a diligent employer would reasonably do given the current state of guidance.
What regulatory frameworks does this connect to?
AI hiring discrimination sits at the intersection of three overlapping frameworks. The Equality Act 2010 governs the discrimination claim itself. UK GDPR Article 22 gives candidates rights over automated decisions that significantly affect them. The EU AI Act classifies recruitment AI as high-risk and requires bias audits and human oversight, with reach into UK businesses recruiting in the EU. These frameworks can all apply simultaneously.
Each framework creates a different documentation obligation. The Equality Act calls for evidence that you considered discriminatory impact when adopting a process. UK GDPR requires you to explain how automated decisions are made and to offer human review when requested. The EU AI Act demands technical documentation, records of bias mitigation, and evidence of meaningful human oversight, not rubber-stamping of AI outputs.
For an owner-managed business, this doesn’t mean becoming an AI governance specialist. It means asking vendors three questions before deploying a tool: how they have tested for bias across protected groups, whether they have independent audit reports available, and what human review their workflow requires. Vendors who cannot answer those questions clearly are worth treating with caution.
If something goes wrong and a claim is made, documentation becomes central. What you asked, what you were told, and what checks you ran is often the difference between a manageable situation and a protracted one. That paper trail is cheap to build at the point of purchase and expensive to fabricate after the fact.
