The Architecture of Human Judgment · An essay by Uwe Seebacher
Imagine a meeting in which a recommendation arrives before the discussion begins. The forecast is on the screen. The proposed action has been ranked, summarised and translated into an implementation plan. Everyone can inspect the answer. Then someone asks a slower question: what would have to be true for this recommendation to be wrong?
This is a thought experiment, not a report from a particular company. Its tension is familiar: the meeting has a deadline, the answer looks professional, and the objection has no obvious place in the agenda. A decision can become difficult to challenge long before anyone formally prohibits disagreement.
My argument is that an intelligent organisation needs a practical capacity to reconsider its own decisions. Better predictions can support that capacity. They cannot, by themselves, create the authority, time and protection that a person needs to raise a consequential objection.
Two questions that a dashboard cannot merge
How accurate is the system? Who can get the organisation to reconsider what it does with the system's answer? These are different questions. The first concerns performance under specified conditions. The second concerns the distribution of responsibility and power.
A model might correctly estimate the likely demand for a service while the organisation chooses a damaging response to that forecast. Conversely, a well-designed review procedure might expose an unreliable input before a decision is implemented. Accuracy and the ability to correct a course of action can support each other, but neither guarantees the other.
I use institutional corrigibility here in this ordinary organisational sense: a capacity to receive objections, examine reasons and revise a course of action. This is a proposed lens for management, not a claim to have invented the established technical literature on AI corrigibility or a new, empirically validated management theory.
The concern is therefore larger than a machine making a mistake. It concerns an organisation losing the means to notice that its own interpretation of a machine's answer is inadequate.
When agreement becomes an attractive product
A useful 2026 debate concerns AI sycophancy: excessive affirmation of a user's position. In a study published in Science on 26 March 2026, Myra Cheng and colleagues examined 11 models and conducted three preregistered experiments involving 2,405 participants. In the experiments, interaction with sycophantic AI increased participants' conviction that they were right and reduced their willingness to take responsibility and repair interpersonal conflicts. Participants nevertheless preferred and trusted the affirming systems.
The scope matters. This was research on social judgments and interpersonal dilemmas, not a trial of corporate decision systems. It does not establish that every AI-assisted boardroom produces these effects. It gives us a reason to examine how the appeal of agreement might interact with existing organisational incentives.
A separate CHI 2025 study by Hao-Ping Lee and colleagues surveyed 319 knowledge workers, collecting 936 examples of AI use. Greater confidence in AI was associated with less reported critical thinking; greater task-specific self-confidence was associated with more. These are self-reports and associations, not proof that AI causes a permanent decline in cognitive ability. The authors also describe thinking shifting toward verification, integration and stewardship of the task.
Together, these findings invite a management question. If a system is pleasant to agree with, and an organisation rewards rapid closure, where does an inconvenient question go? My concern is the possible interaction of these conditions. That interaction requires investigation rather than assertion.
The unequal price of saying yes and saying no
Consider a hypothetical workflow. Accepting a recommendation requires one click. Challenging it requires a written explanation, a meeting and approval from someone whose delivery target the delay threatens. Formally, the employee remains free to disagree. Practically, the organisation has assigned different prices to agreement and objection.
This asymmetry can exist without bad intentions. A manager wants consistency. An analyst wants a decision. A technology team wants adoption. Each objective is understandable. Their combination can make a reasonable objection look like a defect in the person raising it.
The title's “difficult people” therefore needs care. The quality to protect is a person's reasoned contribution, including an uncertain or initially awkward question. Humiliation, obstruction and the repeated refusal to consider evidence deserve no special protection. Equally, a junior colleague should not need the status of an exceptional expert to ask whether a decisive assumption is sound.
This has a long research history. Amy Edmondson's 1999 study of 51 work teams in a manufacturing company associated team psychological safety with learning behaviour. It predates today's generative AI and does not validate the proposal in this essay. It reminds us that the interpersonal conditions for learning are part of organisational design, not a new issue created by chatbots.
Give an objection somewhere to arrive
Methods and structure become concrete at the point where someone raises a concern and another person must respond. My proposal is to design that encounter explicitly for consequential AI-assisted decisions.
First, name the decision and its owner. Record what is being decided, which outcomes matter, who is accountable and what could justify reopening it. An AI recommendation should enter a defined decision, rather than silently define the organisation's objective.
Second, distinguish the object of the objection. Is the concern about the evidence, a model assumption, the goal being optimised or the consequences for affected people? Better data may answer the first question. A conflict between legitimate objectives requires accountable human judgment. Asking for another model run cannot settle every disagreement about priorities.
Third, provide a proportionate route to review. A credible objection to a consequential decision needs a named reviewer, access to relevant evidence and a response within a defined period. Urgent decisions may require immediate action followed by review. Easily reversible, low-impact choices can use lighter procedures. Unlimited vetoes would make the system unusable.
Fourth, preserve the reasons and examine the outcome. Record why an objection was accepted or rejected, then revisit the result when evidence becomes available. Otherwise, the same concern may have to be rediscovered by the next person, while the organisation mistakes repeated silence for agreement.
These suggestions sit alongside existing governance work. The NIST AI Risk Management Framework 1.0, a voluntary framework from 2023, addresses responsibilities for human oversight, feedback and appeals, and post-deployment monitoring. It is not a certification of this proposal. It supplies an established reference against which an organisation can examine its arrangements.
The serious case for fewer exceptions
Standardisation has a strong defence. Human discretion can be inconsistent, biased and captured by influential personalities. A transparent, tested decision rule may improve on an informal process in which the most senior person always wins. Creating more opportunities to override a system could restore precisely that privilege.
This objection changes how dissent should be organised. An override must carry reasons, identify the responsible person and remain open to scrutiny. A powerful manager's preference should face the same evidential discipline as a junior employee's challenge. The point of review is to improve the decision process, including when review confirms the original recommendation.
Nor should an organisation set a target number of disagreements. That would turn a capacity for correction into another performance ritual. A quiet meeting may reflect sound preparation. A noisy one may reflect poor incentives. Neither atmosphere, on its own, measures the quality of judgment.
The practical question is whether a relevant concern can reach someone able to act, and whether the response can survive examination afterwards.
A proposal that should itself be tested
A useful evaluation could begin with one bounded, low-risk decision process. Agree in advance how to judge outcomes and compare the existing procedure with a clearly documented review route. Where practical, use random assignment; otherwise describe the limits of the comparison. Track review time, substantiated errors caught before action, incorrect overrides and downstream consequences.
My working hypothesis is that a usable route for reasoned objection may increase some review costs while preventing some avoidable downstream errors. It may also fail: reviewers may lack capacity, employees may remain silent, or the procedure may produce delay without better decisions. Those possibilities belong in the evaluation. No performance gain is established by this essay.
In the preceding essay on the efficiency trap, I asked who receives the time technology saves. Here the question concerns who can use some of that time to question an answer. Both lead to the organisational conditions under which people can exercise judgment.
This also connects with the question posed by my forthcoming book The End of Expertise: Why Methods and Structures Beat Expertise: how should methods and structures support capable professional action? The present essay develops its own argument; the book is my work and the link provides context, not independent evidence for the proposal.
An organisation reveals something fundamental about itself when it hears a question it cannot yet answer. It can treat the question as an interruption, or make room to investigate it. Its future depends partly on whether it can remain answerable to people whose concerns its current procedures have not anticipated.
The person who asks the slower question offers no guarantee of wisdom. They offer the possibility that the decision is still open to thought.
For editors and producers: Who is allowed to disagree when an organisation has already accepted the machine's answer? For a conversation about AI, institutional learning and human judgment, contact info@uweseebacher.org.
Sources and scope. Sources checked on 31 October 2026: Cheng et al. (2026), published abstract and structured summary; Lee et al. (2025), research abstract; Edmondson (1999), published abstract; NIST AI RMF 1.0 (2023), Core. The March 2026 study is a debate context from this year, not news from this week. The opening and workflow examples are hypothetical. The organisational proposal and working hypothesis are the author's argument. Artwork: AI-generated symbolic illustration.
