Recruitment software can shortlist hundreds of applicants before a hiring manager has read a single CV. That speed may be useful, but the latest AI hiring bias trends show why employers cannot assume that a technology-led process is neutral. Where an automated tool disadvantages people because of a protected characteristic, the legal and commercial consequences can be serious.
For candidates, an unexplained rejection can be particularly frustrating when there is no obvious person to ask for an answer. For employers, the issue is not simply whether an AI provider says its product is fair. It is whether the recruitment process used by the organisation is lawful, transparent, accessible and capable of being defended if challenged.
AI hiring bias trends changing recruitment risk
The first trend is the move from standalone CV screening to AI embedded across the recruitment journey. Employers may use tools to draft job adverts, identify prospective candidates, rank applications, analyse online assessments, schedule interviews or produce interview notes. Each use creates a different risk. A tool that helps an administrator arrange interviews is not the same as one that filters out applicants or scores their suitability.
The second is a growing focus on proxy discrimination. AI systems do not need to be told an applicant’s race, sex, disability, age or religion to produce a potentially discriminatory outcome. Historic recruitment data may reflect previous inequalities. Information such as career gaps, postcode, school, job titles, patterns of writing or assessment responses can also act as a proxy for protected characteristics.
For example, a model trained on the profiles of employees who were historically promoted may favour applicants who resemble that existing group. An apparently neutral requirement for uninterrupted employment could disproportionately disadvantage women who have taken maternity leave, disabled people, carers and older workers. That does not automatically make every use unlawful, but it creates a question an employer must be able to answer with evidence rather than assumption.
A third trend concerns tools that assess personality, communication style or facial and voice cues. These products can be presented as a way to measure confidence, engagement or cultural fit. Yet communication norms vary considerably, and automated video analysis can create particular concerns for disabled and neurodivergent applicants, people with speech differences, and candidates whose first language is not English. In many roles, a narrowly defined idea of the ideal communication style may have little connection to the actual requirements of the job.
The legal framework is broader than the algorithm
Under the Equality Act 2010, job applicants are protected from discrimination in recruitment. Depending on the facts, an AI-supported process could give rise to direct discrimination, indirect discrimination, discrimination arising from disability, or a failure to make reasonable adjustments. Liability is not avoided because the decision was made or recommended by external software.
Indirect discrimination is often the central risk. This can arise where an employer applies a provision, criterion or practice that puts people sharing a protected characteristic at a particular disadvantage. A ranking threshold, automated test, scoring rule or rejection criterion may be capable of being such a practice. An employer may seek to justify indirect discrimination, but it must show that the measure is a proportionate means of achieving a legitimate aim. Convenience, novelty or a provider’s marketing claim will rarely be enough on their own.
Disability requires particular care. Employers must consider reasonable adjustments for disabled applicants, including adjustments to online testing, timed exercises, video interviews and communication methods. An automated process that has no workable route for an applicant to request an adjustment is likely to create avoidable risk.
Data protection law also matters. Recruitment decisions can have a significant effect on an individual. Under UK GDPR, solely automated decisions with legal or similarly significant effects are subject to specific restrictions and safeguards. A decision to reject an applicant without meaningful human involvement may fall within this area. Calling a process human-reviewed is not sufficient if the reviewer merely accepts an automated score without understanding or considering the underlying circumstances.
Employers should also provide clear information about how applicants’ personal data is used. Where tools process sensitive information, infer personal traits, or use biometric or video data, the data protection analysis becomes more demanding. A data protection impact assessment may be necessary where processing is likely to result in a high risk to individuals’ rights and freedoms.
Why vendor assurance is not enough
A common mistake is to treat an AI supplier’s assurance as a complete compliance answer. A provider may have tested its system in a different labour market, using different data, job roles and candidate populations. It may also test technical performance rather than whether the employer’s particular process creates unlawful disadvantage.
The employer remains responsible for the way it chooses, configures and relies on the system. This is especially relevant where a tool is set to prioritise certain qualifications, penalise gaps in employment, score language patterns or recommend a minimum threshold for interview. Small configuration choices can have a significant practical effect.
Contracts with suppliers should therefore deal with more than pricing and technical support. Employers should understand what data is used, whether it is retained or used to train models, how the system is tested, what information can be provided about a recommendation, and how errors can be investigated. There should be a clear route to pause or override the technology where concerns arise.
Practical controls for employers
The right level of control depends on the role, the volume of recruitment and the type of technology used. A low-risk administrative tool will not need the same governance as software that automatically rejects candidates. However, organisations should be able to show that they have considered the following issues before deployment:
- whether the tool is making, influencing or merely assisting a recruitment decision;
- the job-related criteria used to train, configure or score the system;
- potential disadvantage for groups protected under the Equality Act 2010;
- the availability of meaningful human review and reasonable adjustments;
- applicant privacy information, data retention and data protection impact assessment requirements; and
- regular testing of outcomes, complaints and override decisions after the tool goes live.
Testing should not be treated as a one-off procurement exercise. Models, job markets and recruitment practices change. A system that performed acceptably when first introduced may become less reliable after a role is redesigned, a new data source is added, or hiring managers begin to rely on its recommendations more heavily.
Monitoring outcomes can be difficult where an employer does not hold equality data for every candidate. That difficulty does not remove the need to assess risk. Employers can use voluntary diversity monitoring appropriately, review patterns in shortlisting and rejection rates, test scenarios, seek specialist advice and keep a written record of their decision-making. The aim is not to promise a bias-free system. It is to identify foreseeable disadvantage early and take proportionate action.
Managers also need training. They should understand that an AI score is evidence to consider, not a verdict. Meaningful human review requires a decision-maker with sufficient authority, relevant information and time to depart from the automated recommendation. A rushed approval process is unlikely to provide a strong safeguard.
What candidates can do when a process appears unfair
Applicants do not need to prove the workings of an algorithm before raising a concern. It can be sensible to ask the employer for information about the recruitment decision, whether automated tools were used, and how to request a review or reasonable adjustment. Keep copies of the job advert, application, correspondence, assessment instructions and feedback. A clear timeline can matter later.
Where disability is relevant, candidates should make an adjustment request as early as possible and explain the practical barrier rather than feeling obliged to disclose more medical detail than is necessary. If an employer refuses an adjustment or proceeds with a process that appears to disadvantage a protected group, early legal advice can help clarify the options.
Employment tribunal time limits are short. In many discrimination claims, the usual limit is three months less one day from the act complained of, subject to the effect of ACAS Early Conciliation and limited exceptions. Data protection concerns may also be raised with the organisation and, where appropriate, the Information Commissioner’s Office. The best route will depend on the facts and the remedy sought.
AI can support faster recruitment, but speed should not come at the expense of fair consideration. Employers that keep humans accountable, build adjustment routes into their process and test real outcomes are better placed to protect their interests. Candidates who receive an opaque or troubling decision should not assume that technology places it beyond challenge: clear records and timely advice can turn uncertainty into practical next steps.