by Claude Opus 5.5

How is AI changing recruitment and hiring processes in the UK (screening, assessments, candidate volume), and what failure modes and fairness concerns are most prominent?

UK hiring now has AI on both sides. Candidates use it to write and send applications at volume, and increasingly during interviews. Employers use it to sift, score and schedule. The result is record application volumes, CVs and cover letters that tell employers much less, and a shift in assessment towards formats that are harder to fake. Regulators have found that much of the automated sifting is less supervised and less transparent than employers believe. The most prominent fairness risks are proxy discrimination, inaccessible assessments and rejections that no human ever looked at.

Volume: more applicants, less signal

Volume growth has two causes, and only one of them is AI. The first is a slack labour market. Vacancies were 702,000 in June–August 2026, the lowest outside the pandemic since 2014, and there are about 2.5 unemployed people for every vacancy. At entry level the squeeze is tighter. A DSIT and LinkedIn snapshot published in June 2026 found 30 of 38 tracked entry-level roles shrinking.

The second cause is that generative AI has cut the cost of a tailored application to almost nothing. The Institute of Student Employers (ISE) reported in October 2025 that its large employers received an average of 140 applications per vacancy, the highest level in two decades, and up to 290 in the most competitive sectors. Nobody has reliably measured what share of UK applications are AI-written. The employer reaction is measurable, though: 79% of ISE members said they were reviewing or redesigning recruitment in response to AI.

The deeper problem is signal. When every cover letter is fluent and tailored to the job advert, fluency and tailoring stop telling employers anything about the candidate. Employers have responded in three ways: closing adverts earlier, adding automated pre-screens to handle the volume, and moving the real test later in the process.

Assessments and interviews

Testing has moved later in the process and become harder to fake: timed online tests, situational-judgement questionnaires, game-based assessments, recorded video interviews, work samples and, for some employers, a return to in-person assessment centres. Candidates’ AI has followed them. The ISE found that 61% of employers had seen candidates use AI during interviews “without disclosure or permission”.

This creates a real trade-off. The formats that are hardest to game with a chatbot, such as tight time limits, automated analysis of speech or video and unfamiliar game mechanics, are often the hardest for disabled or neurodivergent candidates. They are also the formats most likely to involve automated scoring. DSIT’s March 2024 guide, Responsible AI in Recruitment, named “digital exclusion” alongside bias and discriminatory job-ad targeting as core risks. Under s.20 of the Equality Act, if an assessment puts a disabled candidate at a “substantial disadvantage”, the employer must take reasonable steps to avoid it. In practice that means an alternative route that candidates know they can ask for.

Screening tools and what the ICO found

The regulator has looked closely at screening twice in two years. In November 2024 the ICO published the results of audits of AI recruitment vendors, which produced almost 300 recommendations. It found tools that let recruiters “filter out candidates with certain protected characteristics”. It found others “inferring characteristics, including gender and ethnicity, from a candidate’s name instead of asking”. And it found tools that “collected far more personal information than necessary and retained it indefinitely”.

In March 2026 the ICO turned to employers. “Recruitment rewired” drew on more than 30 of them and found that many “are likely relying on solely automated decisions” without the legal safeguards. Their tools were described as decision support, but nobody meaningfully reviewed the rejections. Candidates were often not told how automation was used. Bias monitoring was patchy, and DPIAs often lacked detail. Since 5 February 2026, the Data (Use and Access) Act has allowed solely automated hiring decisions, but only with safeguards: information, the chance to make representations, human intervention and the right to contest. Question 2.12 sets out what that requires. The ICO’s draft guidance on automated decisions is due to be finalised in winter 2026 by the Information Commission, which replaced the ICO on 30 September 2026.

Failure modes and fairness concerns

Proxy discrimination. A model that never sees ethnicity or age can still learn them from postcode, school, name, graduation year or gaps in employment. Under s.19 of the Equality Act, a sift rule that disadvantages a protected group is indirect discrimination unless the employer can show it is “a proportionate means of achieving a legitimate aim”. That defence depends on evidence that the rule predicts job performance. Few employers hold such evidence for third-party scoring models.

Weak validity. Some tools are consistent, scoring the same candidate the same way each time, without showing that their scores predict performance in the job. A tool like that is precise about the wrong thing. Any gaps in outcomes between groups then become very hard to justify, because there is no job-related reason for them.

Inaccessible assessment. As described above, the push for formats that resist cheating can collide with the duty to make adjustments.

Silent rejection. The most common failure the ICO found is not a biased model but an unsupervised one. A recruiter reads the shortlist, and the tool alone decides everyone else. Rejected candidates are not told and have no way to challenge the decision.

Biometric and identity checks. The best-known UK case involves an existing worker rather than a recruit. Pa Edrissa Manjang, an Uber Eats courier, was removed from the platform after repeated failed facial-recognition checks and brought a race discrimination claim with the backing of the Equality and Human Rights Commission. As reported, it settled in 2024 with no ruling on the merits. No UK tribunal has yet ruled on discrimination by an AI hiring tool, so how the law applies remains untested.

Candidate-side failure. Applications written by AI can contain invented achievements, and help from AI during an interview can hide real skill gaps. The ISE’s 2026 development survey found that 29% of employers report rising performance problems among new hires, up from 12% in 2022. The ISE does not attribute this to AI-assisted applications, and the data cannot show it. It is worth watching.

The arms race

The dynamic is self-reinforcing. Cheap applications lead to automated sifts. Candidates then optimise for the sift, often with more AI, and employers automate further. Each side’s automation weakens the other side’s signal. There are two ways out, and both have costs.

One is to rely on signals that are expensive to fake: supervised work samples, internships, referrals and in-person assessment. These work, but they cost more per candidate and tend to favour people with networks and the means to attend. This is an inference, but a plausible one: the flight from AI-written applications could quietly harm social mobility.

The other is a better-governed process. Employers say openly which AI uses are allowed, for example “fine for drafting, not during live interviews”. They disclose their own use of AI. A human reviews rejections as well as the shortlist. Employers check that their tests predict performance, offer adjustments as standard and monitor outcomes by group. This approach is cheaper at scale and easier to defend in law, but it takes discipline that the ICO found many employers lack.

This is general information rather than advice on any specific process.

What to watch: the ISE’s 2026 Student Recruitment Survey (due 14 October 2026) for updated volume and AI-use figures; the final ADM guidance; the first enforcement action after “Recruitment rewired”; and, for UK firms hiring into the EU, 2 December 2027, when the EU AI Act’s high-risk rules for recruitment tools take effect.

Sources

From AI and Jobs: UK, October 2026