“We keep a human in the loop” is the reflexive answer to every AI hiring risk. The research is blunt: drop a person on top of a biased model and they mostly absorb the bias rather than catch it.
There is a phrase that ends every nervous conversation about AI in hiring: “we keep a human in the loop.” Regulators recommend it. Vendors advertise it. Legal teams sign off on it. It has become the comfort blanket of automated hiring.
The research on whether it actually works points the other way. When you put a real person in front of a biased AI recommendation, they mostly do not correct it. They absorb it.
What people do with a biased recommendationThey follow it
The largest human-subjects experiment on this to date put 528 people through 1,526 resume-screening scenarios. The design was simple: give reviewers AI recommendations and watch what they pick.
Given a neutral AI, people selected candidate groups roughly evenly, the way you would hope. Given a biased one, they followed it, choosing the AI-favored group up to 90% of the time. The bias did not stay in the model. It traveled straight through the human and out the other side as a decision (No Thoughts Just AI, 2025).
Read that again, because it inverts the entire premise. The human was not a filter on the AI's bias. The human was a delivery mechanism for it.
And the bias is nearly invisible to the person watching for it
Maybe, you think, reviewers would catch obvious bias if they were paying attention. They were paying attention. They still missed it.
In a controlled study, reviewers were handed algorithmic recommendations that had been deliberately skewed, then asked about them directly. Around 60% did not notice the bias at all (Kaibel et al.). Not “noticed but deferred.” Did not register that anything was wrong.
This is the quiet problem with oversight as a checkbox. You cannot catch what you cannot see, and a recommendation arrives looking clean, confident, and quantified. There is nothing on the screen that says “this number is skewed.”
The reviewers who trust AI most are the worst check
Here is the finding that should reshape how you staff oversight. A study of 2,784 reviewers found that the people most favorable toward automation overrelied on the AI and adopted its errors, while only the skeptics reliably caught them. Financial incentives to get it right did not close the gap.
So the instinct to put your most enthusiastic AI adopter in charge of the AI is exactly backwards. Enthusiasm is the disqualifier. The useful reviewer is the one who assumes the model is wrong until the evidence says otherwise.
What “human in the loop” has to actually mean
The phrase is not wrong. It is just empty until you specify the conditions under which a human is a real check rather than a rubber stamp. Three of them, drawn straight from what the research says fails:
The human must be able to see the underlying evidence, not just the score, because invisible bias cannot be overruled. The human must be accountable for the decision, owning it rather than ratifying it. And the human must be rewarded for catching the model, not for moving fast in agreement with it, because deference is the default and you have to design against it.
So, does a human in the loop fix biased AI?
Only under conditions most teams have not built. Drop a person on top of a biased model and you usually get the model's decision with a human's name on it: bias followed up to 90% of the time, missed by 60% of reviewers, and worst among the reviewers who trust the tool most.
Oversight is not a checkbox. It is a design problem, and the design starts with whether the human can actually see what they are signing.
A human is only a check when they can see the evidence and overturn the call. Confeti binds every AI claim to a specific moment in the interview, with speaker, timestamp, and quote, human-editable and traceable. The reviewer sees what they are signing.
See how it worksCommon questions
Does keeping a human in the loop remove AI hiring bias?+
Not on its own. In the largest study to date, people followed a biased AI's recommendations up to 90% of the time. The human absorbed the bias rather than correcting it.
Why don't reviewers catch biased AI recommendations?+
Because the bias is invisible on screen. In one study about 60% of reviewers failed to notice deliberately biased recommendations even when asked about them directly.
Who makes the best human reviewer for AI decisions?+
Skeptics. A 2,784-person study found reviewers most favorable toward automation overrelied on the AI and adopted its errors, while only skeptical reviewers reliably caught them.
What makes human oversight actually work?+
Three conditions: the reviewer can see the underlying evidence, is accountable for the decision, and is rewarded for catching the model rather than agreeing quickly.
Should we remove the human from AI hiring decisions?+
No. A human who can inspect and overturn the decision is the only safeguard there is. The point is that the phrase 'human in the loop' is not a safeguard by itself.
References
- No Thoughts Just AI: Biased LLM Recommendations Limit Human Agency in Resume Screening (2025). 528 participants, 1,526 scenarios. arXiv. link
- Kaibel et al., Michael is better than Mehmet: selective adherence to biased algorithmic advice in hiring (n=260). NCBI/PMC. link
- Bias in the Loop: How Humans Evaluate AI-Generated Suggestions (2025), n=2,784. arXiv. link
- Shaping the fairness journey: human-in-the-loop perceived fairer than AI-only (2025). ScienceDirect. link