Photo by Vitaly Gariev on Unsplash
UK employers now hand a growing share of recruitment to software. Applicant tracking systems screen CVs before a human sees them. Video interview platforms score tone, word choice and facial expression. Chatbots run first-round screening for high-volume roles. The pitch to employers is that automation strips bias out of hiring decisions, cuts admin time, and gives every candidate the same consistent assessment. The evidence published this year points to a different conclusion.
A 2025 study testing GPT-4 and Microsoft Copilot found both tools favoured male candidates for senior engineering roles. When asked to generate an image of a "typical" candidate, the same tools produced people who were younger, slimmer and lighter-skinned than the applicant pools they were meant to represent. Separately, Workday is facing US allegations that its AI-powered applicant screening tools discriminated against candidates by age, disability and race, in a case now being watched closely by employment lawyers on both sides of the Atlantic.
Writing in The Conversation, researchers Muneera Bano and Didar Zowghi describe AI bias in hiring as a chain of human decisions, made well before the software screens a single CV. Someone decides which qualities count as senior experience. Someone decides which CVs form the training data. Someone decides what a good culture fit looks like. The model learns and repeats whatever pattern it was handed, at far greater speed and scale than any single recruiter working alone.
Photo by Christina @ wocintechchat.com on Unsplash
Every one of those upstream decisions tends to encode the same assumptions that already exclude neurodivergent and disabled candidates from traditional hiring. Keyword-matching software penalises applicants who describe their experience in precise, literal language, a common pattern among autistic candidates, over the buzzwords a parsing tool is trained to reward. Video interview platforms that score eye contact, vocal tone and facial expression disadvantage autistic and ADHD candidates, people with speech differences, and non-native English speakers, none of whom are being assessed on their ability to do the job.
"Culture fit" criteria create the same problem with or without AI involved. Automating a hiring process built on narrow assumptions scales those assumptions to every application that follows. It also removes the safeguard a human interviewer could still provide: the judgement to notice a strong candidate who does not fit the pattern. Once that judgement is designed out of the process, it rarely gets designed back in.
Photo by Danial Igdery on Unsplash
Employers evaluating or already using AI recruitment software should start with the criteria, not the code. Four questions are worth asking now, ideally before a contract is signed rather than after a complaint arrives.
What data trained this tool, and whose past hiring decisions does that data reflect? A vendor should be able to answer this in specific terms, not marketing language.
Where does a human review every automated rejection? A tool that screens out candidates with no documented human check removes accountability along with the workload.
Has the tool been tested for disparate outcomes by protected characteristic, including neurodivergence and disability where that data exists? Accuracy against a benchmark is not the same test, and a supplier's confidence in its own product is not evidence either.
Would the criteria this tool applies survive being read aloud to a tribunal? UK employers already have Equality Act 2010 duties on indirect discrimination, and the reasonable-steps duty on preventing harassment and third-party discrimination that lands in October raises the bar on evidencing that hiring processes are fair as well as efficient.
An AI hiring tool moves these judgements into new places. It does not remove the need for someone to own each one and explain the reasoning behind it. Human by Practice works with employers to audit recruitment processes and screening criteria for exactly this kind of hidden exclusion, before it becomes a discrimination claim instead of a fixable policy gap. That audit is worth doing on its own merits, well before any tribunal makes it compulsory.