Chapter 1: The Perishable Asset
I have a small fear I did not have ten years ago. It arrives at a whiteboard, marker in hand, when I need to write a word I use every day, and for a fraction of a second I am not sure how to spell it. Not a rare word. A word I have produced thousands of times, except that I have not. Autocorrect has produced it, while my hands did something adjacent to writing. Years of that, and the loop connecting what I know to what I can do without help has quietly thinned. Nobody in the room ever notices. I notice, because I remember not having the hesitation.
I am opening the book with this story so that I can set it aside. The spelling version is trivial. The word gets written, the meeting continues, the cost is a private flicker of doubt. If that were the whole phenomenon, it would not deserve a chapter, let alone a book.
This chapter is about the version that is not trivial. The same mechanism that took my spelling is now operating on the asset your profession now runs on: the judgment that decides what gets built, where the boundary sits, and whether the green checkmark deserves your signature. That judgment is the one part of the job that cannot be handed to the machine, which makes it the binding constraint on everything else. This chapter establishes something less comfortable. It is also the only asset in the building that decays from normal use of the tools that make you productive, and it decays without telling you.
The mechanism is not a metaphor
Start with what is actually known about how this works, because the mechanism matters for everything that follows.
The brain maintains what it is asked to do and stops maintaining what it is not. Eleanor Maguire’s London taxi driver studies established the constructive direction decades ago: drivers who spent years navigating London without assistance developed measurably larger posterior hippocampi than matched controls. The demand built the structure. Véronique Bohbot’s work at McGill then showed the reverse: habitual GPS users shifted away from hippocampal spatial mapping toward simple turn-by-turn response following. The navigation system of the brain was not damaged. It was bypassed. A 2017 imaging study made the timing explicit: the disengagement happens in real time, during the act of following directions, not gradually afterward.
Three details in that body of work deserve a product manager’s attention.
First, the tool worked the whole time. GPS users got where they were going, faster and with fewer errors than the unassisted. Performance and capability moved in opposite directions, and performance is the only one anyone measures.
Second, nothing pathological occurred. This is ordinary neuroplasticity doing exactly what it is designed to do, reallocating maintenance away from an unused capacity. The system is working as intended. The direction is just not the one you intended.
Third, nobody chose it. There was no moment at which any driver decided to trade spatial reasoning for convenience. The trade executed itself, one assisted trip at a time, below the threshold of noticing.
Hold those three details. They recur in every case that follows, and they describe your last eighteen months of work.
What happens to professionals
If the pattern stopped at navigation and spelling, it would be a curiosity. It does not stop there.
Aviation found it first, because aviation is where automation first became good enough to trust. Pilots who spent long stretches supervising reliable automation became measurably worse at manual flight when the automation handed back control. The knowledge remained; the practiced, pressurized competence did not. Air France 447 is the canonical case, a stall that went unidentified through a recoverable window by a crew the automation had been quietly carrying. The industry’s regulator eventually wrote the pattern down in plain language: continuous use of automated systems does not maintain manual flying skill. Some long-haul captains are estimated to log under one hour of true manual flying per year. The aircraft works. The asset erodes.
Medicine produced the cleanest recent number. A 2025 study in The Lancet Gastroenterology and Hepatology tracked experienced endoscopists after AI-assisted colonoscopy entered their daily routine. Within three months, their detection rate without the AI had fallen from 28.4 to 22.4 percent. A fifth of their independent ability to find precancerous growths, gone in a quarter, in physicians who had spent years building exactly that skill. None of them felt it happen. A separate controlled study measured experienced pathologists reversing roughly seven percent of their own correct diagnoses after seeing a wrong AI suggestion, a small rate with an unambiguous direction. The error was not that they lacked the skill. It is that the skill deferred.
Note what these cases share with the taxi drivers. The tool improved supervised performance, which is real and worth having. The tool degraded independent performance, which nobody was measuring. And the practitioners could not detect the change from the inside, because the inside is where the change happens.
The part that should actually worry you
Navigation, manual flying, scope technique, even diagnostic pattern matching: you could file all of it under specialized perceptual-motor skill and tell yourself that reasoning is different. Two studies say it is not, and they are the reason this book exists.
The first is from MIT’s Media Lab. Fifty-four participants wrote essays across four months in three conditions: with a chatbot, with a search engine, or unassisted, while EEG measured what their brains were doing. The chatbot group showed progressively lower neural connectivity across sessions, the weakest executive control, the lowest attentional engagement. The search group looked like the unassisted group. That contrast is the finding: assistance as such was not the problem. With search, the cognitive work of composition still ran, fed by better inputs. With the chatbot, the work was not being redirected. It was being skipped.
Then the researchers crossed the groups over, and the result got worse. Chatbot users asked to write unassisted did not recover the engagement of the group that had been writing independently all along. The independent writers, given the chatbot, used it well and stayed engaged. Prior skill was protective. Prior dependency was not self-correcting. The asymmetry should sound familiar; it is the GPS bypass, measured in the writing brain.
The second study is from Wharton, and it supplies the number I want you to keep. Shaw and Nave ran three preregistered experiments, 1,372 participants, on reasoning problems designed to detect whether people deliberate or accept the first plausible answer. An AI assistant was introduced, and its answers were sometimes right and sometimes wrong, without disclosure. When the AI was right, accuracy jumped 25 points above the unassisted baseline. When it was wrong, accuracy fell 15 points below. Participants who received a wrong answer followed it on roughly four trials out of five.
And then the finding that organizes this entire book: access to the AI raised participants’ confidence by nearly twelve points whether the AI was right or wrong.
Read that the way you would read an eval result. The system’s output quality varied; the operator’s confidence did not vary with it. Confidence had detached from accuracy and attached to the experience of assistance itself. A signal that moves the same way in both states carries no information, which means that in an AI-saturated workflow, the felt sense that you are reasoning well has stopped being evidence that you are.
Shaw and Nave drew a distinction that the rest of this book will lean on hard. Cognitive offloading is strategic: you hand a discrete task to a tool and retain the judgment about when and how to rely on it. Cognitive surrender is different: you stop constructing the answer at all and adopt what the system produces. The judgment is not delegated in surrender. It is relinquished, and the operator is the last to know, because the twelve points of confidence arrive either way.
Your inventory
Now bring it to your own desk, because the studies are about you, not about endoscopists.
Count the judgments you exercised through AI in the past week. The effort estimate you asked for instead of building. The prioritization rationale the model drafted and you edited, which feels like judgment but begins from its framing rather than yours. The synthesis of eight customer calls you read instead of hearing the calls. The competitive read produced from a model’s summary of filings you did not open. The first draft of the spec, which used to be the place your thinking happened and is now the place your reviewing happens. The recommendation memo arriving pre-argued, with the structure of the decision already chosen by the time you engage it.
None of these is a mistake. Most are exactly what working well with these tools looks like; the productivity is real, the 25 points are real, and a PM who refuses the tools is making a different and larger error. The question this chapter installs is narrower and harder: for each item on that list, are you offloading or surrendering? Did you delegate the task and keep the judgment, or did the judgment quietly travel with the task?
There is a test, and you already know it from the studies. Offloading survives the tool’s removal; surrender does not. The MIT crossover group could not write unassisted anymore. The endoscopists could not detect at their old rate. If the model were unavailable next Tuesday, which of last week’s judgments could you still produce at your old standard, at your old speed, with your old confidence justified? You do not know. That is not an accusation. It is the literal finding: the only measurement that would answer the question is the one your daily workflow never runs. Chapter 8 builds the instrument. This chapter only needs you to admit you currently do not have one.
Why you cannot feel it
Every losing argument with this chapter takes the same form: I would notice.
You would not, and the reasons are mechanical rather than personal. The decay is invisible to the performer because the performer’s measuring instrument is the thing decaying. The endoscopists’ sense of their own scope technique was intact at month three; the rate was not. The pilots believed they could take over; the belief had not been tested in years, and belief without testing is exactly what inflated confidence is made of. The MIT subjects experienced themselves as writing. The EEG disagreed.
Worse, the confidence finding means the felt signal points the wrong way. AI assistance makes you feel sharper while the independent capacity recedes, so the stronger your sense that your judgment is fine, the less that sense is worth. “My judgment is fine” is what a well-calibrated practitioner says, and it is what a deteriorating one says, and from the inside the sentences are indistinguishable. A sentence that is true in both states is not evidence of either.
And the moderator data closes the last exit. In the Wharton experiments, higher trust in AI and lower appetite for effortful thinking predicted more surrender; the people most confident they were immune were the most exposed. If you have read this far while privately exempting yourself, the data has a category for you, and it is not the protected one.
This is the trap fully assembled. The asset is perishable. The decay is silent. The confidence signal is broken in the direction of reassurance. And the daily workflow never runs the one measurement that would catch it. Nothing about being smart, senior, or aware of the problem dismantles the trap, because the trap is made of the same cognition that would have to dismantle it.
The asset, named
Step back and tally what this chapter actually adds.
Judgment is the binding constraint of this profession now: the go or no-go, the boundary placement, the gate signature, the accountability for people the agent never sees. Everything else commoditized; this did not. The entire economy of the agentic team routes through the assumption that the practitioner holding those decisions can still make them.
This chapter’s contribution is one sentence added to that ledger. The binding constraint is a perishable asset, and the tools that make you productive are the agent of its decay.
Every other asset your company depends on has a maintenance program. The fleet of agents has six observation instruments, drift detection, and an eighteen-month recalibration clock. The data has lineage and governance. The codebase has tests. The one asset the whole structure routes through has a feeling of confidence that the evidence says is broken, and nothing else.
An honest reader will notice what this chapter has not yet shown: that anything can be done. Decay being real does not make it reversible, and a maintenance program is only worth building if oversight capability is the kind of thing that responds to maintenance. That is the next chapter’s question, and the answer is the reason this book is a practice rather than an elegy. The capability turns out to be a gradient, not a gift. Gradients can be climbed.
The whiteboard hesitation was never about spelling. It was the feeling of standing in front of a room and discovering, in public, at speed, which kind of user of my tools I had been. You will get your version of that moment. The only question this book can influence is whether you meet it calibrated or surprised.