Human in the loop needs attention in the loop
It’s 16:30, this is the fortieth AI-generated proposal you’ve reviewed today, and it looks about as reasonable as the previous thirty-nine. The summary reads fine, the recommendation matches what you’d expect, and you click approve. Somewhere in a compliance log, this moment counts as human oversight.
Nobody made you skip the review. You read the screen, clicked the button, and the process included a human checkpoint. Organizations still rarely let AI run processes fully or partly on its own. In Info Support’s 2026 AI-adoption research among 400 IT decision-makers, only 5 percent reported doing so [1]. That makes the design of human review relevant in many AI-supported processes. In practice, human in the loop means that a person reviews, approves, changes, or can overrule an AI-generated outcome. It describes where that person sits in the process, while review design shapes how effectively they can contribute.
Recent research on AI oversight calls this the difference between formal presence and human evaluative agency. A formal role establishes who participates in the process. Effective oversight also depends on the time, context, authority, and cognitive space to understand an AI system’s output, test it, contest it, and overrule it when it’s wrong [2].
On a compliance dashboard, approving in three seconds looks identical to approving after real scrutiny. Both register as human oversight, although the conditions under which that oversight happens can differ. The difference has nothing to do with how conscientious you are. It comes down to what the process asks of your attention over the course of a working day. I’ve sat on both sides of that screen, and the click never feels different in the moment. Effective oversight comes from protecting that person’s attention, independent judgment, and authority to act. That’s the case this piece makes.
Attention that fades
That design has to start somewhere, and the task itself is the right place to begin. Reviewing dozens of AI-generated proposals in a single shift is a form of sustained monitoring, which carries a well-documented failure mode. Vigilance decrement describes how the ability to spot rare, critical signals in a stream of mostly normal cases drops over time. Studies of sustained monitoring tasks find measurable performance loss within about fifteen minutes. Under demanding workloads, that loss can start within five [3]. None of what follows is about laziness or personal failure. It’s what monotonous, mostly correct review work does to attention by design.
A second pattern compounds it. Automation complacency occurs when a system usually gets things right. People check it less often than they should, monitoring grows sparser, and problems surface later than they would under closer scrutiny. The risk grows exactly where AI tools tend to perform best: high apparent reliability, repetitive structure, and a reviewer juggling this queue alongside several others [4]. A code reviewer waving through the tenth pull request marked “passed CI” in a single afternoon knows this pull well. The system has been right often enough that checking it starts to feel like a formality.
Toward the end of a long run of decisions, a third factor adds weight. Decision fatigue appears as a drift toward the default option, the status quo, or whatever choice takes the least mental effort. It tends to worsen as a session goes on [5]. Info Support’s 2026 whitepaper also describes a related shift. As AI absorbs routine tasks, the judgment work that remains can become mentally heavier, contributing to a form of cognitive exhaustion sometimes called “AI brain fry” [1]. By the time you’ve spent all day weighing AI proposals, the default of “looks fine, move on” becomes more attractive with every case. An AI recommendation shown early on the screen provides exactly the default that fatigue makes easiest to accept.
The first recommendation wins
The order in which information appears matters. When you see an AI recommendation before you have thought through the case yourself, that recommendation becomes the starting point for your judgment. Psychologists call this anchoring: an initial value or conclusion influences the decisions that follow, even when it should not [6].
Imagine that an AI agent has already collected the relevant documents, assessed the case, and prepared a recommendation. The first thing you see is its proposed decision: approve the application. You have not yet examined the evidence or formed your own view.
As you review the documents, you no longer start from a blank page. Approval has already become the working conclusion, so you are more likely to look for reasons why the application is acceptable than to assess it independently. Had you reviewed the same evidence before seeing the recommendation, you might have noticed a reason to reject it or request further investigation.
Research on anchoring shows that people often remain closer to an initial value than the evidence warrants [7]. Advice has a similar effect. People tend to move toward a recommendation, even when the reasoning behind it is weak [8]. Research on AI-supported decision-making finds the same risk when AI output appears early and prominently, especially under time pressure or when the evidence is ambiguous [6].
This changes the review task. Instead of asking “What does the evidence support?”, you are more likely to ask “Is there any reason not to approve this?”. Confirming an answer that is already on the screen requires less effort than forming one yourself.
The most direct response is to change the sequence. Review the evidence first and record your provisional judgment. Then reveal the AI recommendation and compare the two. Ask what evidence would justify changing either position. This does not remove anchoring, but it prevents the AI recommendation from doing your initial thinking for you.
From presence to judgment
Reversing the order of one screen won’t fix all of this on its own. Putting a human in the loop is only the first design decision. Meaningful oversight also needs sustained attention, independent judgment, and authority to act. Together, those conditions shape how the work is organized, what the interface shows, and who remains accountable for the final decision.
Protect attention. Vigilance decrement and decision fatigue both erode the ability to notice the rare case that matters over a shift. Protect reviewers’ breaks, keep review blocks short, and avoid stacking the largest or riskiest AI-generated changes at the end of the day. These changes reduce both risks without asking reviewers to concentrate harder.
Preserve independent judgment. Automation complacency makes active checking feel unnecessary. When the AI recommendation appears first, anchoring pulls the reviewer’s remaining judgment toward it. Ask reviewers to inspect the evidence and record an initial judgment before revealing the AI recommendation. Then ask what evidence would justify changing either position. In one review process I helped redesign, that reordering was the entire first sprint and the cheapest change on the list.
The interface should support the same sequence. Show the evidence behind the recommendation, such as relevant changes, test results, assumptions, and uncertainty, instead of a single definitive verdict. Prompt reviewers to consider an alternative explanation in high-stakes cases. Explainability supports oversight, but the way explanations are presented also shapes how independently people evaluate the recommendation. In two experiments, explanations did not reduce automation bias and sometimes increased it [9]. Their value depends on how they affect the reviewer’s behavior, not on whether an explanation is present. Present the AI output as advice to examine, not judgment to defer to.
Give judgment authority. Independent judgment only matters when someone is authorized and accountable to act on it. Clear governance establishes ownership. Review design helps people exercise that responsibility effectively. In a software team, that can mean assigning a named reviewer or code owner instead of treating an agent’s green check or a shared review queue as accountability. Diffusion of responsibility occurs when responsibility is spread across several people, so each person assumes someone else will act. In one monitoring study, pairs without individual feedback detected fewer automation failures than people performing the same task alone [10]. A shared queue quietly answers the question “who is responsible here?” with “everyone, a little”, which in practice means no one in particular.
Authority should also depend on demonstrated review quality rather than job title alone. A study of bail decisions found that most human overrides of an algorithmic risk score performed worse than the algorithm itself. A smaller group of judges made overrides that performed measurably better, although that quality is likely to depend on the task and domain [11]. Formal responsibility should remain clear, while high-stakes review assignments should also reflect demonstrated domain knowledge and review quality.
Start with one review step
You don’t need to redesign every AI-assisted workflow at once. Start with one review step in your development process. An AI-assisted pull request review is a good candidate because the proposed change, supporting evidence, and recommendation are already visible in one place.
For one sprint, ask the reviewer to inspect the diff and record a provisional judgment before opening the agent’s recommendation. That judgment can be a likely verdict, the main risk they see, or the question the review still needs to answer. Then reveal the AI output and compare the two. Record what evidence changed either position.
You don’t need to rebuild the tool to test this. A checklist, a collapsed panel, or a draft comment that reviewers complete first is enough for a pilot. Track how often the recommendation changes the review, how often the reviewer identifies something the agent missed, and how often disagreement leads to a code, test, or documentation change.
A change in those patterns does not prove that the new sequence is better. It does show that presentation order was influencing the review. That is enough reason to examine the workflow more closely.
It’s 16:30 again. Same fortieth review of the day, and you’re still at the screen. The process already had a human in the loop. The redesign gives that person a chance to inspect the evidence and form a view before the agent frames the answer.
That is the real design test: not whether a person clicked approve, but whether the process preserved an independent judgment before asking for that click.
References
[1] Info Support, “AI-adoptie in beeld: van inzet naar regie”, whitepaper, 2026. Available at: https://www.infosupport.com/resources/ai-adoptie-van-inzet-naar-regie/
[2] Zhu, L., Lu, Q., Ming, D., Lee, S.U. & Wang, C., “Designing Meaningful Human Oversight in AI”, AI and Ethics, 2026.
[3] Hemmerich, K., Luna, F.G., Martín-Arévalo, E. & Lupiáñez, J., “Understanding Vigilance and Its Decrement: Theoretical, Contextual, and Neural Insights”, Frontiers in Cognition, 2025.
[4] Parasuraman, R. & Manzey, D.H., “Complacency and Bias in Human Use of Automation: An Attentional Integration”, Human Factors, 2010.
[5] Grignoli, N., Manoni, G., Gianini, J., Schulz, P., Gabutti, L. & Petrocchi, S., “Clinical Decision Fatigue: A Systematic and Scoping Review with Meta-Synthesis”, Family Medicine and Community Health, 2025.
[6] Carter, L. & Liu, D., “How Was My Performance? Exploring the Role of Anchoring Bias in AI-Assisted Decision Making”, International Journal of Information Management, 2025.
[7] Tversky and Kahneman, “Judgment under Uncertainty: Heuristics and Biases”, foundational paper on anchoring, 1974.
[8] Bonaccio, S. & Dalal, R.S., “Advice Taking and Decision-Making: An Integrative Literature Review, and Implications for the Organizational Sciences”, Organizational Behavior and Human Decision Processes, 2006.
[9] Vered, M., Livni, T., Howe, P.D.L., Miller, T. & Sonenberg, L., “The Effects of Explanations on Automation Bias”, Artificial Intelligence, 2023.
[10] Cymek, D.H., “Redundant Automation Monitoring: Four Eyes Don’t See More Than Two, if Everyone Turns a Blind Eye”, Human Factors, 2018.
[11] Angelova, V., Dobbie, W. & Yang, C.S., “Algorithmic Recommendations and Human Discretion”, The Review of Economic Studies, 2025.