The Confidence Transfer

What Happens to Organizational Expertise When We Trust the Machine Over the Person

Issue 271, July 2, 2026

A senior analyst at a client organization watched a meeting go a direction she knew was wrong. The team was weighing a forecast, and she had fifteen years of direct experience with exactly this kind of decision, the kind of pattern recognition that does not reduce to a slide. She said so. She explained her reasoning, named the specific risk she saw, and recommended a different path. Across the table, a colleague pulled up an AI tool, typed in the parameters, and read the output aloud. The output disagreed with her. The room shifted in a way she described as almost physical. The questions stopped. The decision moved toward the machine’s recommendation, and her fifteen years of judgment were quietly set aside, not because anyone argued against them, but because the model had spoken and the model sounded certain. She told me the part that stayed with her was not being overruled. It was how quickly it happened, and how relieved everyone seemed to no longer have to weigh competing human judgments at all. A month later, the forecast proved her right. By then no one remembered that she had said so.

This is the dynamic I want to examine, because it is becoming one of a series of dynamics representing the most consequential and least discussed shifts in how organizations actually make decisions in the age of AI. We are living through a quiet transfer of confidence, away from human judgment and toward machine output, and it is happening faster than our ability to think clearly about when it is warranted and when it isn’t. I call it The Confidence Transfer, the gradual migration of trust from people to systems, often regardless of which one is actually right in a given case. Its danger isn’t that machines are frequently wrong, although we do underestimate how frequently they hallucinate. It’s that the transfer happens by default rather than by discernment, and that every time it happens, something erodes further in the organization that is very hard to reverse.

The Thread We Have Been Following

For readers who have been with the last several issues, this continues a deliberate arc. We have examined, in The Permission Paradox (Issue 269), why depleted employees adopt AI in the shadows. We have examined, in The Empathy Outsource (Issue 270), what happens when leaders outsource the language of care. And now we turn to judgment itself, the most consequential thing an organization asks of its people. I will restate what I have said in each of these pieces, because it remains true and it matters most here. This is not an AI newsletter, and it has not become one. The subject is, as always, the human factor. AI is simply the most powerful instrument we have ever had for revealing how humans actually behave under uncertainty, pressure, and the temptation of an easier path. I have, for years, raised the point, in this newsletter, podcast and in my books, that we shy away from critical thinking, we shy away from making decisions, particularly when there are too many choices, and we shy away from using more energy than our evolutionary programming allows us. Our technical prowess is once again offering us a solution to our default preferences, for better or worse.

The Confidence Transfer, then, isn’t a story about artificial intelligence. It’s a story about deference, expertise, and what we do with the deeply human discomfort of having to think and decide.

This Is Not New. That Is the Point.

The most important thing to understand about the Confidence Transfer is that the underlying psychology was documented long before generative AI existed, which tells us the behavior is human, not technological. The machines changed. The tendency didn’t.

In 1997, Raja Parasuraman and Victor Riley published what remains a foundational paper in the journal Human Factors, titled “Humans and Automation: Use, Misuse, Disuse, Abuse.” Studying how human operators interacted with automated systems in aviation and other high-stakes settings, they named a pattern they called misuse, defined as overreliance on automation that produces failures of monitoring and decision bias. I explored this dynamic in The Readiness Illusion (Issue 265) and in Season 2, Episode 26 of The Human Factor Podcast.

Parasuraman and Riley documented that when people are given an automated aid, they begin to monitor less, question less, and defer more, and that this complacency grows precisely when the automation has been reliable in the past. The more the system earns trust, the less the human scrutinizes it, which means the human’s attention fades exactly when a rare error would be most catastrophic. This body of work, extended by Parasuraman and Dietrich Manzey in a 2010 review on complacency and automation bias, established something that should anchor every conversation about AI in organizations today. The human tendency to over-defer to machines is not a reaction to how good the machines are. It’s a feature of how human attention and accountability work under conditions of automation.

What makes the current moment different is not the psychology. It’s the scope. Aviation automation governed a bounded set of technical functions. Generative AI now offers an opinion on almost anything, in fluent and confident language, to almost everyone in the organization at once. The narrow channel through which the Confidence Transfer once operated has become the entire river.

The Finding That Should Unsettle Every Expert

If the Parasuraman tradition gives us the historical foundation, a 2019 study provides us with the version of the problem that organizations are living right now, and it contains a finding that should unsettle anyone who has spent years developing expertise.

Jennifer Logg, then at Harvard, with Julia Minson and Don Moore at Berkeley, published “Algorithm Appreciation: People Prefer Algorithmic to Human Judgment” in Organizational Behavior and Human Decision Processes. Across six experiments, they found that people adhered more to identical advice when they believed it came from an algorithm than when they believed it came from a person. This directly challenged the older assumption of algorithm aversion, the idea that people instinctively distrust machine recommendations. Logg’s work showed the opposite is often true. When the advice is labeled algorithmic, people weigh it more heavily, defer to it more readily, and discount their own view more willingly. The label alone moves the needle.

But the finding that matters most for the human factor is buried in the later experiments, and it’s genuinely paradoxical. Algorithm appreciation waned among experts. Experienced professionals, the people who make forecasts and judgments for a living, relied less on algorithmic advice than novices did. That sounds reassuring until you read the consequence the researchers documented: this expert skepticism, in their study, hurt the experts’ accuracy in the specific tasks tested. The very experience that makes someone valuable also makes them resistant to outside input, and there are cases where that resistance is costly. This is the genuine complexity at the center of the Confidence Transfer, and I am not going to flatten it for the sake of a cleaner argument. I explored this paradox of expertise in The Confidence Calibration Problem (Issue 247) and in more depth in Season 2, Episode 14 of The Human Factor Podcast, The Identity Crisis of Expertise, where the fusion of professional identity and subject matter knowledge creates exactly the kind of resistance Logg’s research documented.

The honest position isn’t that human judgment is always right and machines should defer to us. The honest position is that both over-deference and under-deference are real failure modes, that they coexist, and that the organizational challenge is calibration rather than allegiance to either side.

This is where most commentary on AI and decision-making goes wrong. It picks a team. Either AI is a threat to human wisdom, or human bias is a problem AI will solve. The research supports neither side. It supports the uncomfortable middle, where the right amount of trust in the machine is a moving target that depends on the task, the stakes, the quality of the model, and the actual track record of the human in that specific domain. The organizations that will navigate this well are not the ones that pick a side. They are the ones that build the capacity to tell the difference, case by case, which is far harder than either blanket trust or blanket skepticism.

What the Transfer Costs Over Time

The immediate risk of the Confidence Transfer is the wrong decision in a single meeting. The deeper risk is what happens to an organization’s expertise when deference to machines becomes habitual, and here the most recent research is genuinely sobering and somewhat scary.

A 2025 study by researchers at Microsoft and Carnegie Mellon University, published at the CHI Conference on Human Factors in Computing Systems, surveyed 319 knowledge workers about critical thinking in their AI-assisted work. The pattern they found is a near-perfect description of the Confidence Transfer in motion. The more confidence a worker had in the AI, the less critical thinking they reported engaging in. Crucially, the researchers observed that knowledge workers using these tools shifted their cognitive effort away from the substantive work of analysis and toward the thinner work of verifying and editing what the AI produced. The thinking didn’t disappear, but it changed character, from generating and evaluating ideas to checking a machine’s output, a shift the researchers framed as a move from deep engagement to surface-level oversight. When trust in the tool rises, the human’s own analytical effort falls.

A neuroscientific study from MIT researchers in 2025 sharpened the concern further. Using EEG to monitor brain activity in college students completing writing tasks, the researchers found that those who used an LLM assistant showed lower brain connectivity, diminished sense of ownership over their work, and reduced engagement with their own ideas compared to participants who worked unaided. The LLM lowered cognitive friction and improved short-term efficiency, but it came at a measurable cost to learning and to the depth of the person’s involvement with the task. Let that all sink in for a few moments.

The finding I want to be careful with here is that this is early research on a specific task, and the long-term effects on skilled professionals in real workflows are not yet established. But read alongside the Microsoft survey and decades of automation research, it points in a consistent and concerning direction. The capability we offload, we slowly stop developing. The judgment we stop exercising, we slowly stop having. Abilities are like muscles; atrophy occurs when usage is limited or completely halts. The implications extend beyond individual organizations. A generation of professionals who have never needed to exercise the full weight of their judgment will inherit decisions that require exactly that.

This is the mechanism that makes the Confidence Transfer a compounding problem rather than a static one. An organization that defers to the machine in one decision has lost one decision. An organization whose people have spent three years deferring to the machine has something worse on its hands. It has a workforce that no longer practices the judgment the organization will desperately need the day the machine is confidently, catastrophically wrong. Expertise is not a possession. It is a practice, sustained only by use, and the Confidence Transfer quietly removes the occasions for that use. We have built elaborate systems to capture what our experts know. We have built almost nothing to ensure they keep knowing it.

In Organizational Memory Loss (Issue 258), I examined why organizations forget what they learn. In Decision Theater (Issue 254), I described how organizations mistake the appearance of decision-making for the practice of it. In Structural Silence (Issue 256), I explored why organizational culture trains people not to speak. Each of those dynamics is amplified by the Confidence Transfer, because deferring to the machine replaces the friction through which learning, genuine decision-making, and dissent occur. We are therefore already operating from a deficit. Our deference to automated decision-making compounds rather than resolves the organizational vulnerabilities we carry forward.

The Future We Are Walking Toward

It’s worth looking down the road, because the trajectory of the Confidence Transfer raises challenges that organizations (and humans in general) are not yet structured to handle.

The first is what we might call the calibration crisis. As AI systems become more capable and more frequently correct, the rational basis for trusting them grows, which means the deference becomes harder to argue against in any individual case. Every time the machine is right, the cost of having questioned it looks like wasted effort, and the cost of not questioning it stays invisible until the rare failure. Organizations will face mounting pressure to defer more, justified by an accumulating track record, even as that very deference erodes the human capacity that serves as the last line of defense when the system fails. The better the machine gets, the more dangerous unexamined trust becomes, because the failures grow rarer, stranger, and harder for an out-of-practice workforce to catch.

The second challenge is the erosion of what I will call productive disagreement. The analyst I described at the opening was a source of friction in that meeting, and friction of that kind is not a flaw in organizational decision-making. It’s the organizational immune system. The dissenting expert, the colleague who says wait, the person who has seen this go wrong before, these are the mechanisms by which organizations avoid their worst decisions. The Confidence Transfer threatens to route around all of it, because a confident machine answer ends deliberation in a way a human opinion rarely does. We are already uncomfortable with confrontation; we seek consensus, but often we don’t want to make the effort to gain consensus, and we hold biases that blind us to alternatives demonstrated by others. So when we have already concluded that our best modes of working require more effort than we are willing to invest, deferring to automation seems easy. Any fault that results is not on the humans in the room; it is attributed to the model.

So when the model speaks, the conversation that would have surfaced the dissent never happens, and many are more than happy it didn’t. The risk isn’t just worse individual decisions. It’s the slow atrophy of the organization’s entire capacity to disagree productively, to hold competing views in tension long enough to find the better answer.

The third challenge is accountability, and it is the one leaders should be thinking about now. When a human expert makes a call and is wrong, the organization learns something about that person’s judgment, and that person carries the weight of the outcome in a way that sharpens their future thinking. When the machine makes the call and is wrong, who learns? The research on automation bias found that people are often willing to defer to automated systems partly because it diffuses personal accountability.

The Confidence Transfer is, then, not only a transfer of trust. It is a transfer of responsibility, away from people who can be held to account and toward a system that cannot, and an organization that loses the thread of who is actually accountable for its decisions has lost something it will not easily notice is gone until it needs it.

What Leaders Owe This Moment

I am not going to offer a prescription for the right level of trust in AI, because the entire argument of this issue is that no fixed level exists, and anyone selling you one is selling you a simplification that will cost you. What the Confidence Transfer asks of leaders is harder than a policy. It asks for the deliberate preservation of human judgment as a living practice, even when deferring to the machine would be faster, cheaper, and defensible in the moment.

In practice, that means treating the question of when to trust the machine as a decision in its own right, rather than a default that operates invisibly. It means building the kind of culture where a colleague can say I think the model is wrong here and is met with genuine inquiry rather than the subtle social cost the analyst paid. It means noticing when the room goes quiet after the machine speaks and treating that silence as a signal to slow down rather than a sign of resolution. And it means protecting the conditions under which expertise is exercised and therefore sustained, because a workforce that has outsourced its judgment for a decade will not be able to reclaim it in the crisis that finally requires it.

The deepest point I am seeking to express is this. The value of human expertise was never only in being right more often than a machine. Sometimes the machine will be right more often. The value is in the practice of judgment itself, the friction, the accountability, the productive disagreement, the hard-won pattern recognition that makes a person worth listening to even when, especially when, the easier path is to let the confident answer settle the matter. An organization can transfer its confidence to machines and gain real efficiency in the bargain. What it must not do is make that transfer by default, without noticing, until the day it reaches for human judgment and finds that the practice has quietly gone out of the building. The question isn’t whether to trust the machine. The question is whether your organization is still deciding that question on purpose, or whether it has already stopped deciding it at all. I fear for some, it may be the latter.

Related Reading

The themes explored in this article connect to a wide body of work across the Ideas and Innovations newsletter and The Human Factor Podcast. The following pieces offer the most direct connections to the dynamics examined here.

The Confidence Calibration Problem (Issue 247) examines why self-assessment fails us at the worst possible moments, with direct relevance to the expert skepticism dynamic documented by Logg, Minson, and Moore in this article. The analyst in the opening was living the Dunning-Kruger paradox that issue explored, amplified by algorithmic authority.

When Being Right Isn’t Enough (Issue 248) explores why human expertise, even when it produces the correct answer, often fails to influence the decision. The opening anecdote in this article is a near-perfect illustration: the analyst was right, and it did not matter. That issue examined the dynamics in human-to-human influence. The Confidence Transfer shows what happens when the prevailing voice is a machine.

The Advice We Never Take (Issue 246) examines why leaders seek counsel they systematically ignore, the flip side of the Confidence Transfer. Where that issue documented the dismissal of human counsel, this one documents the uncritical acceptance of machine counsel, two expressions of the same underlying difficulty with genuine deliberation.

The Algorithmic Mirror (Issue 253) examines what AI reveals about how we actually think and decide. Where that issue treated AI as a diagnostic lens into human cognition, the Confidence Transfer is the behavioral consequence: once the mirror reveals our patterns, we defer rather than reckon with what we see.

The Busyness Trap (Issue 250) explores why we wear exhaustion as a status symbol. The cognitive depletion that issue documented is precisely the condition under which the Confidence Transfer accelerates, because a depleted workforce experiences deferral to AI as relief rather than abdication.

De-Risking Too Much Choice examines the decision paralysis that arises from too many options, a dynamic directly referenced in this article. The Confidence Transfer offers a seductive resolution to choice overload: let the machine decide.

The Human Factor Podcast, Season 2, Episode 14: The Identity Crisis of Expertise explores what happens when professional identity fuses with subject matter knowledge and that identity is then threatened. The Confidence Transfer threatens expertise at its most foundational level, and this episode examines the psychological mechanisms at work when that threat materializes.

The Human Factor Podcast, Season 1, Episode 3: Why Smart People Make Bad Decisions examines the psychology of bias in leadership, the underlying cognitive architecture that makes the Confidence Transfer possible. The biases that lead experienced leaders to poor judgment do not disappear when AI enters the room; they find new expression.

Sources

1. Raja Parasuraman and Victor Riley, Humans and Automation: Use, Misuse, Disuse, Abuse, Human Factors 39, no. 2 (1997): 230-253
2. Raja Parasuraman and Dietrich Manzey, Complacency and Bias in Human Use of Automation: An Attentional Integration, Human Factors 52, no. 3 (2010): 381-410
3. Jennifer M. Logg, Julia A. Minson, and Don A. Moore, Algorithm Appreciation: People Prefer Algorithmic to Human Judgment, Organizational Behavior and Human Decision Processes 151 (2019): 90-103
4. Hao-Ping (Hank) Lee et al., The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers, Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Microsoft Research and Carnegie Mellon University)
5. Nataliya Kosmyna et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, MIT Media Lab, arXiv preprint 2506.08872 (2025)