AI recruitment can decide which job applications reach a human before candidates have spoken to anyone. Employers may use automated tools to search CVs, compare applicants or organise assessments. The attraction is speed, but the information processed by those systems can be highly personal, and their decisions may be difficult for rejected applicants to understand.
Britain’s Information Commissioner’s Office examined recruitment technology and said some tools were collecting excessive information or allowing unfair filtering. In the regulator’s November 2024 report, the ICO explained that its audits led to nearly 300 recommendations to the providers examined. That does not mean every recruitment system discriminates, but it is direct evidence that real products required improvement rather than merely hypothetical future safeguards.
What AI recruitment may do before you get an interview
A recruitment process can involve numerous decisions that never appear on a candidate’s screen. A tool may collect applications, parse information from a CV, identify relevant experience, place people into groups or provide a ranking for a recruiter to review. Some products can also support interviews or assessments. These jobs sound administrative, yet they influence which people receive attention and which people quietly disappear from a shortlist.
The key distinction is between helping a recruiter organise information and giving software effective authority to reject someone. A human can remain technically involved while relying so heavily on a numerical ranking that little independent assessment takes place. Conversely, a well-designed tool could help a recruiter identify overlooked candidates if its data, controls and review procedures are sound. The impact depends on the process, not the presence of the letters AI in a sales brochure.
A CV is not a complete description of a person. It contains choices made by the applicant, constrained by experience, confidence, language, disability, education and the conventions of the industry. An automated system may treat those choices as useful signals even when they do not predict the ability to do the job. That creates a practical risk for people with unconventional career histories or those who express ordinary experience in less fashionable terminology.
LiveAIWire has examined the effects of CV presentation on automated hiring and the risk that employers using similar algorithms make similar decisions. Both themes matter because a rejected application may reflect the design of a screening process as much as a person’s actual suitability. Neither observation proves that any individual rejection was discriminatory.
The privacy issues found by the UK regulator
The ICO reported that some systems allowed filtering by protected characteristics, and that others inferred characteristics such as gender or ethnicity from a candidate’s name. Inferring such information is particularly troubling because guesses can be inaccurate as well as intrusive. The regulator instructed providers to address fairness and accuracy rather than treating the prediction as an innocuous feature.
Auditors also encountered excessive collection and prolonged storage of candidates’ personal information. A rejected application should not automatically become a permanent pool of data for an organisation to analyse for unrelated purposes. Employers and suppliers need a lawful basis for processing, clear reasons for collecting each field and appropriate retention periods. Candidates should understand what is happening without having to decipher an obscure technical document.
In a separate guide for organisations buying recruitment tools, the ICO urges employers to carry out a data protection impact assessment, examine potential bias and tell candidates how personal information is used. That guidance points to responsibility on the employer’s side: outsourcing the software does not remove the obligation to make sensible decisions about what data is processed and how.
It is worth separating the different organisations involved. A software supplier may build a model, a recruitment agency may configure it, and an employer may decide how much weight to give its output. A candidate usually sees only the employer’s application page. When a result is challenged, those layers can make it difficult to identify which decision or data field caused the problem. Good governance assigns responsibility in advance, including who must investigate a disputed outcome and who can correct inaccurate information across systems.
The regulator also says organisations should check suppliers’ assurances rather than accepting a vague claim of ethical AI. Useful questions include what information a system needs, whether a person can challenge or correct inaccurate records, how performance is tested and who can access the results. These are questions about the actual service a business purchases, not merely abstract principles.
Why disability can make automated screening especially difficult
Disability offers a clear example of why equal-looking procedures can have unequal effects. An assessment that requires a particular interaction or communication style may exclude someone who could perform the underlying job with a reasonable adjustment. A system might treat a gap in employment or a nonstandard response as poor performance without understanding its cause.
The US Equal Employment Opportunity Commission and Department of Justice warned in May 2022 that some software used in hiring and employment decisions can violate disability protections under the Americans with Disabilities Act. The warning concerns American law and should not be presented as a direct statement of British legal duties. It nevertheless illustrates a general design problem: a seemingly neutral automated measurement can disadvantage a person when it measures something unrelated to the work.
Employers need to think about what a test is really measuring, whether alternative ways of assessment are available and how applicants can request support. When a product gives a candidate no way to explain circumstances or obtain a human review, an efficiency gain for the employer can become a serious procedural problem for the person seeking work.
A practical adjustment is not necessarily complicated. It might mean offering another assessment format, allowing enough time for a person to explain a relevant need or ensuring that a software score is not interpreted outside the purpose for which it was tested. What counts as reasonable or legally required depends on the law and circumstances. The broader lesson is that an automated process should not make it harder to recognise individual circumstances that a careful human assessor would take into account.
It is equally important not to assume that every use of software creates illegal discrimination. A carefully validated tool, with clear boundaries and meaningful review, may improve consistency in some tasks. The burden is on those selecting and operating it to show that its benefits are relevant, that its limitations are understood and that people are not unfairly screened out.
What applicants and employers can do with this information
A candidate who encounters an unfamiliar screening step can ask what information is being collected, why it is necessary, how long it will be retained and whether automated processing influences a consequential decision. The answers may depend on the employer, jurisdiction and particular system. Asking does not guarantee an interview, but it helps distinguish an organised selection procedure from an opaque data-gathering exercise.
Candidates should also avoid believing they must write a CV exclusively for a machine. Clear, accurate descriptions of relevant achievements help both humans and software understand an application, but keyword games can obscure rather than improve the evidence of ability. The more significant reform is a process that tests job-related skills fairly and does not mistake a polished document for talent.
For employers, the practical test should begin with the intended outcome. Is the tool finding qualified people, or simply processing applications faster? Does it produce evidence of fairness across relevant applicant groups? Does someone regularly examine rejected cases? Could a candidate obtain an understandable explanation and correction if the system used bad information? Those questions deserve answers before deployment.
Automation can create genuine benefits, including faster responses and more manageable workflows. LiveAIWire has also reported on research into automated interviews and candidate outcomes, which is a separate line of evidence rather than proof that every recruitment tool improves hiring. The sensible ambition is to use technology to reduce unnecessary administration while keeping the human consequences visible.
Another useful test for employers is to follow an application all the way through the system. Can a manager see the source information behind a ranking? Can inaccurate education or employment details be corrected? Does an unsuccessful candidate receive appropriate information about retention of their data? If the tool cannot support ordinary accountability, its apparent speed may conceal future costs in complaints, distrust and poor hiring decisions. A successful recruitment system should make a defensible decision easier to explain, not merely faster to produce.
A job application is more than another data record. It concerns income, security, opportunity and a person’s next chapter. That is why transparency, a fair chance to be assessed and proportionate data use must be treated as the substance of responsible AI recruitment, rather than small-print additions after the system has already begun rejecting people.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.