“But we’ve got a human in the loop.”
This has been the line a lot of teams reach for when someone questions the use of AI within hiring. It usually describes something real: a human performs a check before a decision is made.
What often happens in practice is that someone glances and approves. For example, they might see an AI-ranked list, but not the reasons or logic behind it. This makes challenging it really difficult. Meaning disagreeing is rarely part of the process, unless an error is particularly noteworthy.
In these cases, what is meant to be ‘human in the loop hiring’ is just a click on the way to a decision that was already made. The behaviour is common enough to have earned its own name: rubber stamping.
Rubber stamping is problematic for several reasons, with increasingly sharp evidence building over the last few years.
This article is for general information, not legal advice. UK regulatory guidance on automated decision-making in recruitment remains under review.
Unfair or biased outcomes
In November 2024, ICO audits of AI recruitment tool providers found tools that let recruiters filter candidates out by protected characteristics, tools that inferred gender and ethnicity from names, and tools that collected far more data than they needed and kept it indefinitely to build databases without candidates’ knowledge.
None of that is visible to the person signing off the shortlist. The reviewer sees names and scores, but not the filtering that shaped the list. That side happened upstream. The potential for bias is huge.
But that’s only one part of the equation. Another part of it is how humans respond to potential bias.
In a 2025 study presented at the AAAI/ACM Conference on AI, Ethics and Society, researchers at the University of Washington asked 528 people to shortlist candidates for 16 jobs, from computer systems analyst to nurse practitioner to housekeeper. Each participant picked three of five candidates to interview. Four of the five were equally qualified.
With no AI recommendation, or with a neutral one, participants chose white and non-white candidates at roughly equal rates. The researchers then introduced AI recommendations with racial bias built in. With a moderately biased system, participants’ picks drifted to match its preferences. With a severely biased one, they followed it about 90% of the time.
“Our findings were stark,” said Kyra Wilson, the study’s lead author. “Unless bias is obvious, people were perfectly willing to accept the AI’s biases.”
One detail points at a fix. Bias dropped 13% among participants who took an implicit association test before they started. Being primed to look for bias helped. But being asked to review did not.
Human in the loop hiring is no longer a legal protection
Until recently, UK data protection law restricted decisions made entirely by machine. Keeping a person involved was part of how employers stayed on the right side of it.
From February 2026, solely automated decisions are permitted — including things like rejecting candidates — but there has to be safeguarding. The rules say that you must:
tell people when a significant decision is being made automatically
let people make representations about it
let people contest it
provide a route to human intervention.
Decisions based on special category data, such as health, ethnicity and religion (among others) stay restricted to explicit consent, contract necessity or a legal basis.
Human in the loop hiring is no longer what keeps you compliant, but rather a choice about how you want to hire. A half-made choice is the worst position available: if your reviewer cannot change the outcome, you are running automated decision-making and carrying every obligation that comes with it, without ever having decided to, and without the safeguards you would have built if you had.
Perhaps unsurprisingly, the lawsuits have begun too. Three separate US cases filed in 2026 challenge how algorithms are used to hire and to fire. One argues that applicant scoring amounts to an undisclosed dossier candidates never get to see or contest, one over a system alleged to have selected employees for layoffs after they took parental or medical leave, and one alleging age discrimination in screening.
These are US cases; the European picture is different. A candidate here challenges a decision through an employment tribunal or a complaint to the regulator rather than a class action. But from 2 December 2027, employers using ‘high-risk’ recruitment AI in the EU carry explicit duties of competent human oversight and record-keeping.
Equality law has not moved either. A screening criterion that disadvantages a protected group is unlawful unless it is a proportionate means of a legitimate aim, regardless of whoever or whatever applied it.
What real human in the loop hiring looks like
So how does a company actually adopt human in the loop hiring? Start with three questions about any AI-assisted stage:
Can the reviewer see the evidence, or only the result? A score with nothing behind it can only be accepted or refused on faith, and candidates are entitled to meaningful information about the logic involved. A team that cannot show its own reviewer how a result was reached cannot explain it to a candidate who asks.
Do reviewers have the time, training and authority to overrule a decision? Many processes fail at least one. A reviewer working a long list at speed, with no mandate to disagree, usually agrees.
Is anyone checking whether review improves decisions? Record overrides and review the patterns each quarter. If reviewers agree with the tool 95% of the time, treat that as a finding rather than reassurance. It might mean the tool is good, or it might mean the review is cosmetic, and you cannot tell which without looking.
Honest answers to those three usually point at the same gaps. The following steps then help to close them, running roughly from what to do before you buy a tool to what to check once it is running.
Map each AI-assisted stage. Write down what the reviewer sees, what they can change, and who decides when they disagree. The gaps are often obvious once they are on paper.
Run a data protection impact assessment before you buy AI-assisted recruitment tools. It should cover whether the approach is necessary and proportionate, and how bias and errors will be managed.
Ask suppliers about training data, bias testing and fairness testing. Keep the answers with the DPIA. Vendor assurance does not discharge your obligation, but its absence tells you something.
Give candidates a real route in. They must be able to ask for human intervention, give their point of view and contest the decision. A no-reply address meets none of the three.
Check reasonable adjustments. You must make them for disabled candidates, and an AI tool a candidate cannot use on equal terms is a barrier whatever its accuracy.
Have panels score independently before discussing. CIPD notes that discussing first sways people, including towards the most senior person in the room. The anchoring that makes a biased AI persuasive makes a confident colleague persuasive too.
Log overrides and read them quarterly. Keep UK records for at least a year after the decision: the limitation period is now six months, plus up to twelve weeks of early conciliation, plus a margin.
How Screenloop supports human in the loop hiring
We continue to build Screenloop around a straightforward belief: technology should remove admin, but humans should stay firmly in the decision seat. Our product follows this logic. AI transcribes and summarises, while your team interviews, scores and makes decisions.
Screenloop’s Interview Intelligence keeps what was actually asked and answered in each interview visible, so reviewers and panel members work from evidence rather than what they remember days later.
Processes are run the same way for everyone. Every step is structured and every decision is evidenced, so candidates for a role are asked comparable questions and judged against the same criteria, rather than against whatever each interviewer thought to ask that day.
Data and analytics show where the process works and where it does not, with talk ratio, trends and benchmarks in one dashboard. Candidate Pulse and Hiring Team Pulse cover the other side, measuring candidate experience and cNPS at each stage rather than guessing at it.
Our AI Notetaker writes the notes, but has no say in the verdict. Decisions stay with trained interviewers working to shared criteria.
Frequently asked questions
Is a human sign-off enough on its own?
Not if that person cannot change the outcome. A reviewer who can only approve is not providing oversight, they are providing cover. The test is whether they can disagree and have the result actually change.
Can you reject a candidate using AI alone in the UK?
Since 5 February 2026, yes, with safeguards. You must tell the candidate, let them make representations and contest the decision, and give them access to human intervention. Decisions using special category data (health, ethnicity, religion and others) remain restricted. Equality law applies to the outcome either way.
Do we need a data protection impact assessment?
A DPIA is expected for AI-assisted recruitment. It should cover whether the approach is necessary and proportionate, and how bias and errors will be managed.
Does the EU AI Act apply to you if you are UK-based?
It can if you hire into the EU. High-risk recruitment duties start on 2 December 2027, but the disclosure requirement for AI chatbots and interviewers has applied since 2 August 2026.
What actually counts as a rubber stamp?
There is no single test, but three signals together are a strong indicator: the reviewer sees a result without the reasoning, has less time than the decision needs, and has never been told that disagreeing is part of their job. Add a near-100% agreement rate and the question answers itself.
Should you log every override?
Logging all of them is simplest, and the EU will require six months of logs for high-risk systems from December 2027. The value is in reading them rather than storing them — an override log nobody opens is the same failure as a review nobody acts on.
This article is general information, not legal advice. UK regulatory guidance on automated decision-making in recruitment is under review, so check the current version before you act.
