AI recruitment advice produced a surprising contradiction in one experiment: recruiters said they trusted algorithmic hiring advice less than advice from a human expert, yet their actual decisions were influenced more strongly by the algorithm. The contradiction emerged from a 2026 study of 1,208 recruitment professionals by Alain Lacroux and Christelle Martin Lacroux. It raises an awkward possibility for employers: asking a person to approve an AI-assisted hiring decision does not necessarily mean the person can resist the computer’s recommendation.
The researchers did not show that every recruiter behaves this way in real interviews. Their participants made controlled choices between candidates after receiving manipulated advice. Yet the result matters because companies increasingly use software to sort applications, screen candidates and recommend whom to interview. Human oversight is often presented as the answer to automation risk. This research asks whether oversight works when the human disagrees with a system but follows it anyway.
How AI recruitment advice was tested against human advice
The experiment gave each professional two candidate profiles and asked for an initial hiring decision. Participants then saw a recommendation from either a human expert or an algorithmic recruitment system. Some advice was accurate and some inaccurate, and the number of advisers also varied. The important feature was conflict: the researchers were interested in what happened when external advice challenged the recruiter’s first judgement.
People do not usually abandon an opinion simply because a second opinion arrives. The study found broad evidence of what researchers call egocentric advice discounting: participants tended to give considerable weight to their original choices. But the source of advice still made a difference. Despite expressing less trust in the algorithm, they were more affected by its recommendations than by those attributed to human experts. That gap between what people say and what they do is the centre of the finding.
It is not proof that computers are uniquely persuasive in every situation. The study was designed around particular candidate comparisons and advice conditions. Participants did not face all the uncertainties of a workplace, such as discussions with colleagues, legal review, references or a live interview. Nevertheless, a controlled setting can reveal a behaviour that would be difficult to identify from ordinary recruitment records: the moments when somebody quietly changes their decision after reading a recommendation.
Trust and reliance are two different things
Someone can distrust a system in principle yet still regard its output as useful in a particular moment. An AI score may appear consistent, efficient and untroubled by human distractions. It may also seem to carry statistical authority even when its method is poorly understood. Such impressions can encourage a recruiter to reconsider an initial choice. Reconsideration is not inherently bad; expert human advice can also expose a mistake. The risk appears when a recommendation acquires more authority than its evidence deserves. This is also why public trust in AI-assisted decisions cannot be treated as the same thing as reliable performance: confidence is a perception, while a decision still needs to be tested against facts.
The study’s authors describe a trust-reliance paradox. Trust predicted some advice-taking behaviour, but it did not neatly explain the stronger influence of algorithmic recommendations. The researchers also found that when advisers agreed, their recommendations carried more weight. Exploratory analyses suggested inaccurate advice could damage the quality of decisions more than accurate advice improved them, although that aspect requires careful interpretation beyond the main experiment.
For a job applicant, this may sound unfair: why should a recruiter let software affect the verdict after claiming not to trust it? But the more useful question is whether the employer knows how the recommendation was generated, whether it is accurate for the relevant role and whether staff have an effective route to challenge it. Blind obedience is not oversight. Equally, refusing every machine suggestion would discard useful information. The objective is informed, proportionate judgement.
Human involvement cannot be a box to tick
The UK government has addressed this practical problem in its Responsible AI in Recruitment guidance. It recommends that employers examine how staff will interact with a system, maintain effective human oversight, test performance and make the role of AI clear to applicants. That guidance concerns procurement and workplace practice; it is not an endorsement of the experiment’s particular findings.
A human reviewer needs more than a button marked accept or reject. They need to understand the limits of the system, know what evidence supports the recommendation and have enough time and authority to disagree. An organisation could make oversight nominal by rewarding speed or requiring staff to explain every disagreement while accepting AI recommendations without scrutiny. Conversely, meaningful review can use the machine’s suggestions as prompts to check overlooked qualifications or apparent inconsistencies.
LiveAIWire previously examined the privacy and fairness issues surrounding automated recruitment. That wider debate often focuses on which data employers collect and how tools may disadvantage particular candidates. The new experiment adds a different concern: the psychology of the person expected to supervise the system. Even a technically sophisticated tool can be misused if its outputs become an unquestioned final word.
An algorithmic recommendation is not a neutral fact
Recruitment systems may produce rankings or scores based on historical data, candidate descriptions or patterns in other decisions. Such outputs can be useful for prioritising review, but a score is not the same as a direct measure of someone’s ability to do a particular job. The employer must decide which abilities matter, verify that the system measures something relevant and ensure that candidates are compared fairly. A precise number can conceal a surprisingly subjective choice about what the employer values.
The experimental researchers manipulated advice accuracy to distinguish following good guidance from following poor guidance. Real organisations often have a harder task: they must establish accuracy themselves, using outcomes that can take time to observe. If a recruitment tool recommends someone who never gets interviewed, it may be impossible to know whether they would have succeeded. Feedback loops can reinforce the choices a system already favours, especially where organisations analyse only the people they hire.
This problem is not solved by attaching a disclaimer to every AI result. A warning may be read and forgotten, whereas the recommended candidate remains prominent on the screen. Employers need processes that make uncertainty visible at the point of decision. If the underlying evidence is thin, the software should not present an unqualified verdict. If a recommendation is contested, the reviewer should know what independent evidence could resolve the disagreement.
Candidates need to know who actually assessed them
For jobseekers, transparency can be as important as the technology’s accuracy. Applicants may spend considerable time tailoring a CV or preparing a supporting statement without knowing whether software will screen it first. Where automated decisions or personal information are involved, employers have obligations to explain relevant processing and protect people’s rights. A vague assurance that a human remains in the loop tells applicants little about how their case was evaluated.
Recruiters also need training that is more practical than a warning that AI can be biased. Training can present examples where the system is wrong, ask reviewers to document why they agree or disagree and test whether they notice changes in candidate evidence. Reviewing outcomes across groups can reveal problems that no single recruiter sees. Such measures will not guarantee fairness, but they create a better chance of detecting overreliance before it becomes an organisational habit.
The question connects with LiveAIWire’s earlier coverage of how AI-mediated job interviews are changing recruitment. Whether an algorithm screens applications, summarises an interview or proposes a shortlist, its greatest influence may come from the moment a person treats the output as an authoritative starting point. A recommendation can steer attention even when nobody explicitly delegates the final decision.
What the experiment cannot tell us
The 1,208 participants were recruitment professionals, which gives the study more workplace relevance than a classroom survey of students. Even so, the decisions were experimental. The result does not tell us how frequently AI changes real hiring outcomes, whether a particular commercial system is biased or how different recruitment laws alter behaviour. It also cannot establish that employers should always reject algorithmic advice. Those conclusions require additional evidence.
What the experiment does show is a measurable separation between stated confidence and behavioural influence. That is an important distinction for anyone promising that humans will retain control of consequential AI decisions. Control is not established by naming a human approver. It has to be demonstrated in the process by which advice is questioned, checked and sometimes refused.
A sensible employer should ask two questions: how well does this tool perform, and how do our people behave when it speaks? The first is about the technology. The second is about the workplace. Ignoring either would make AI-assisted recruitment harder to trust, even if the process looked carefully supervised from the outside.
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.