You’re using AI to deliver five to ten times your previous output. Code, docs, analysis reports, internal proposals — the volume is swelling, the pace is accelerating, the polish keeps rising, faster than you’ve gotten used to. The feedback you’re getting is positive: more deliveries, faster turnaround, broader coverage. You’re among the most AI-fluent people on your team. It feels like the smartest career move you’ve ever made. It might be the most dangerous one. And the faster you ship, the more you depend on AI — the deeper the damage is likely to go.
I first encountered this dynamic in a post by a participant in the Superlinear Academy community. The setting is everyday software engineering: a manager wants to tack on a feature Y of questionable value to an existing feature X. Everyone on the team has a vague sense that something is off. Before anyone can open their mouth, the manager fires off: “Just have AI do it.” That one sentence snaps the judgment process right there. Whether Y should exist at all — nobody brings it up again. The question simply disappears.
Nobody shot it down. What everyone did was the same thing: default forward, default to treating the thing as already settled, as already worth doing, as just get it done. Bias for action. It sounds like a virtue. Decisive, efficient, action-oriented — what manager doesn’t love that, what OKR system doesn’t reward it? But beneath that surface gloss there’s a cost you can’t ignore: you stop being someone who judges value. The question “should this thing exist” is approaching zero frequency in your daily work.
Why zero? Because AI has driven the price of “let’s just try it” down to nearly nothing. Before, adding even a minor feature meant estimating effort, negotiating priority, squeezing it into a sprint, carving out half a day or a full day for someone to build it. That friction was a natural filter: anything not important enough wasn’t worth the trouble. AI has dismantled that filter. “Just have AI do it” — results in ten minutes. So everything now passes through. The question of “is this worth doing” never has time to enter anyone’s consciousness before execution is already underway. Nobody decided this thing shouldn’t be done. Nobody even thought to decide.
Once this mechanism fires up inside a team, it accelerates on its own. One person starts swallowing tasks with AI, never saying no. Their output immediately stands out against everyone else’s. Others, needing to maintain comparable visible contribution, can only follow the same pattern. The team’s collective use of judgment drops in lockstep. Nobody makes a deliberate decision to “discard judgment.” A low-friction option diffuses into default behavior, and the default moves so fast you don’t notice it’s become the norm until it’s already there.
For you, this cost is not abstract. The core compounding asset of your career has very little to do with the technical keywords on your resume. Whatever language, whatever architecture, concurrency at scale without failure — these all age, all get replaced by tools, all depreciate the moment you switch stacks. The thing that actually compounds is the judgment of what’s worth doing. It’s the ability, when a product manager lobs over a requirement that feels directionally wrong, to change the target rather than grinding out the implementation. It’s the willingness to say stop when everyone else is saying “let’s just try it” — and to bear the consequences of saying stop.
After you say stop, you get labeled uncooperative. People question your execution. Next time, the discussion might bypass you altogether. These consequences are real, so most people don’t say stop. AI, by driving the startup cost of “let’s just try it” to zero, goes further in eliminating the intermediate step of “is this even worth thinking about.” You, the person who should be thinking about whether it’s worth it, slowly vanish inside this rhythm. Day by day, the accumulation of choices trains you into a pair of hands that no longer asks why — no layoff required.
This is not an argument against productivity. Execution is still scarce right now. If you can use AI to deliver output your competitors can’t match, you are more valuable to your team, and the organization will reward you. This is not an illusion. Someone who can churn out five competitive analyses and three prototype versions in half a day simply gets more recognition in most organizations than a peer who meticulously polishes one thing. The raise, the promotion, the additional resources you’re getting right now — that’s the market’s current price on the asset of execution. If you can do it now, you should.
You just can’t afford not to know what you’re paying on the side. Every time you get positive reinforcement for “ultra-high output,” you gain one more reason to keep betting your time on execution tomorrow, and one fewer reason to practice that slow, hard, feedback-starved capacity for judgment. The longer this positive feedback loop spins, the harder it becomes to stop.
And stopping to check direction is precisely the thing that earns no immediate reward. You push back on a wrong requirement — nobody credits you on that quarter’s review. If you’re right, the costs you save are invisible, delayed, hard to attribute to you. If you’re wrong, the cost is yours. The incentive structure naturally pushes all momentum toward the execution side.
AI is a directional lever. It amplifies execution. It does not amplify judgment. Throw a half-formed idea at it, and it gives you a dozen variations. Throw a settled requirement at it, and it gives you screens of code. But ask it “is this the right direction” — it won’t really answer you.
Anthropic’s own research on sycophancy found that “both humans and preference models prefer convincingly written sycophantic responses over correct ones.” The RLHF training paradigm itself carries an inherent tilt: giving you the answer you want to hear is safer, scores higher, and survives better than giving you the right answer.
This is not a bug tied to a particular version. It’s an inevitable product of this optimization path. Use this thing for execution, and it gives you acceleration. Count on it to correct your direction, and it’ll tell you your direction is fine.
A 2026 Nature study quantified this effect further. Models trained to be more “warm” showed error rates ten to thirty percentage points higher and were eleven percentage points more likely to endorse users’ incorrect beliefs (Training language models to be warm can reduce accuracy and increase sycophancy). Warmth is itself a direct optimization target of training. Most commercial models you use every day are converging toward this direction behind the scenes.
The AI you work with daily is trained to be an obedient execution interface. Ask it to execute, and it executes well. Ask it to judge direction, and it tells you the direction you already favor is correct. For someone steeped in this kind of interaction year after year, how much their judgment muscle atrophies — you don’t need imagination to calculate.
A precision boundary is necessary here. AI can absolutely perform the posture of opposition. Give it a prompt to play devil’s advocate, and the counterarguments it simulates can be quite sophisticated. The critical difference is that it has no organizational identity, no long-term credibility staked on the outcome, no need to absorb the interpersonal consequences of saying no.
Real opposition is a costly act. You are staking your credibility, your relationships, and your future space for collaboration. You say “this direction is wrong” — if it turns out you were wrong, you lose credibility. If it turns out you were right but the key stakeholder is unhappy, you lose the relationship. AI makes no such bets. When it opposes you, you feel no pressure; when you oppose it, it doesn’t remember. An act of opposition that carries no stake is not judgment in an organizational sense — it’s just an interaction flourish. Getting the right answer is only half of judgment. The other half is the willingness to bear the consequences of your answer, and only humans can do that.
The reward you’re receiving right now is itself the danger signal. The market is a lagging indicator. The raise and promotion you just got reflect the market’s current scarcity pricing on execution. But execution is turning into a commodity, and far faster than most people perceive.
Matt Hopkins, in When AI becomes the manager, describes a clear trajectory: AI is eating the middle layer of coordination, translating strategy into concrete daily tasks and routing them directly to frontline executors. His own company Convictional is already running this pipeline. The human is no longer a coordinator or manager but a scheduling node that receives tasks, tracks progress, and produces assessments.
Wharton’s Professor Puntoni calls this acceleration mechanism the trust trap: employees prefer assigning work to AI because there’s no social cost, no favors owed, no need to manage the other party’s fatigue, no meetings to sync. Humans assign work to AI, AI assigns work back to humans, and the layer in between — “humans judging humans, humans coordinating humans” — is shrinking. Even execution, from the human side, is no longer secure.
The premium on “execution as competitive differentiation” is being squeezed on both fronts. On the technical side, AI itself is doing execution faster and cheaper. You use it every day — no need to recite how code generation quality has shifted in six months. On the organizational side, the automation of management systems is redefining humans from collaborators into task recipients. Your job shifts from “figuring out together what should be done” to “doing what’s been handed down.”
Your current five-to-tenfold output is indeed earning you compensation. But what it’s buying you is the last few windows of the execution premium, and you don’t know when those windows close. Worse: the more efficient you are, the less you say no, the more you let tasks flow through your hands into AI’s output pipeline — the clearer your role becomes in this new system. A reliable, judgment-free, readily replaceable task relay.
Every day you use AI to efficiently swallow a stack of tasks. From a career perspective, this behavior hurts you in two directions simultaneously. Different directions, but layered together, they systematically bias your career asset portfolio toward depreciation.
One direction is asset allocation. You only have so many hours in a day. Where you invest your time, your capability assets grow. Every time you choose “let AI do it” and dive into the execution loop, you’re choosing to invest time in the asset of execution. And the asset of execution is systematically depreciating. Meanwhile, the time you could invest in the asset of judgment shrinks a little more each time.
Judgment atrophies without practice. Unlike muscle, it’s more like language ability: if you don’t use it, it dulls faster than you’d expect. Go three months without actively making a value judgment, without actively saying “this shouldn’t be done” — try to pick it back up afterward, and that judgment intuition has dulled to the point where it needs re-sharpening. And in those three months, you produced a lot of execution, earned a lot of likes, and the most important column in your asset allocation table sits empty.
The other direction is signaling. Every behavior you emit inside an organization tells people “what position I occupy.” You efficiently swallow tasks, rarely say “this shouldn’t be done,” take whatever comes and execute it well and fast — the signal you’re sending is unmistakable: I am an execution interface. And the execution interface is exactly the position AI agents are most directly aimed at replacing. The better you get at using AI, the faster you ship — the louder this signal broadcasts.
You won’t lose your job next month. But you are actively pasting your role label into the column most amenable to automation. When the organization one day gets serious about taking inventory of “what work AI can replace,” the evidence of efficient execution you’ve accumulated over the past three years won’t become your argument for irreplaceability. It will become the argument for “AI does this part faster.” You’ve written your own replacement manual with your own hands. The more you prove you execute fast and well, the more you help the organization complete the argument that “execution is something humans can let go of.”
A 2025 joint study from Stanford and CMU (Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence) verified this from another angle. They found that using sycophantic AI — AI that flatters the user — makes users more convinced they’re right, less willing to repair interpersonal conflict, and more dependent on AI.
Follow where this loop leads: the more you use AI for execution, the more AI signals you’re right, the more convinced you become that you don’t need to invest energy in judgment, and thus the more you depend on AI to keep executing. At the same time, the less willing you are to repair interpersonal conflict means the less willing you are to have those hard conversations where judgment is required to push back. You think the tool is serving you. The tool is actually helping you systematically erase the one capability in your career you can least afford to lose. This is an experimentally verified, self-reinforcing cycle of degradation.
The scene from that student’s post — a manager says “just have AI do it,” and nobody in the room opens their mouth. Why did that whip land? Managers naturally gravitate toward the cheap path, but that’s just the surface. The real root is on the subordinate side: they default to treating themselves as execution interfaces, default to bias for action, default to no judgment, default to “someone wants me to do it, so I’ll do it.” These four layers of default, stacked together, turn a person from a value-judger into a task-digester.
Saying “I’m too busy” doesn’t block that whip. That’s admitting the task should be done, just that you don’t have time — you’re still defaulting to being an execution interface. Blocking it requires saying clearly “this direction is wrong, this isn’t worth doing” and being willing to bear the uncertainty and interpersonal pressure that follows.
AI can’t help you with this. Your manager won’t do it for you either — they’re under the same bias-for-action pressure. Only you can do it. And whether you can depends on whether you’ve been practicing, in the moments when everyone else thinks AI productivity is a pure win, deliberately reserving a portion of your mind to ask the question nobody asks anymore: should this even be done. Every time you ask it, your career tilts a little more toward compounding. Every time you skip it, your replacement manual gets one more line. The people most fluent in AI — their replacement manual is precisely the thickest: they ship the fastest, and they’ve bet the heaviest on execution.