Make It So. Make It Stop.
I’ve done this thing where I reached for Claude to handle something tedious: a block of boilerplate, a paragraph rewrite, some config file I didn’t want to manually edit. I told myself it’ll be fast. Thirty seconds of processing, paste it back in, keep moving.
Except it’s never just thirty seconds.
The AI produces something. Now I have to evaluate it. Is it right? Close enough? Better than what I would have written? I read it. I consider it. I make a micro-decision: accept, modify, reject. Then I paste it back into my work and keep going.
Thirty seconds of AI time. Four minutes of my attention in evaluation. But more importantly: I’ve created an incomplete cognitive transaction. I asked the AI a question, it gave me an answer, but I didn’t decide what to do. I just ratified what it did.
That distinction matters more than you’d think.
If I do it ten times in a session then something shifts. I’m no longer in flow. I’m in evaluation mode. I’m not creating; I’m supervising a machine. And supervision is exhausting in a way that actual work isn’t.
The worst part? I got more done. Faster. Output improved. All the metrics point up. But something essential eroded somewhere in the process. The sense that I’m in control of the work. The momentum. The authorship.
That’s when I realize I’ve stopped deciding and started ratifying.
You can sustain flow while using AI. But only if you stay in the driver’s seat.
The Steering Wheel and the Rubber-Stamp
There are two places you can be in any AI-assisted work session.
At one end: holding the steering wheel. You direct, the machine executes, you check the output and the check is cheap. A test passes. A number adds up. The function does what you said. You’re Jean-Luc Picard zip-zipping and bebopping through the galaxy: make it so, and it’s so. Evaluation takes seconds. Decide, delegate, move on.
At the other end: holding a rubber stamp. The machine produced something and you’re deciding whether it’s any good. The check is expensive: is this paragraph right? Does this argument land? Is this the voice I’d have chosen? And there’s no test for any of it. Just your taste, your time, your flagging attention. You’re Mickey in the third act of the Sorcerer’s Apprentice: brooms everywhere, water rising, and you’re not sure you remember the spell that started this.
The axis between them is one thing: how expensive the output is to check. Not how much the AI did. Not even who held authority in the abstract. You can dictate every parameter and still end up rubber-stamping if the only way to verify the result is with your gut.
Coding sits near the wheel end because tests verify: the function either works or it doesn’t, and the quality check is baked in. Prose sits near the stamp end because only taste can judge, and taste doesn’t run in a terminal. This is why AI and flow-states work better in technical domains than creative ones, and why knowledge-workers who try to match their developer friends’ AI workflows end up exhausted and confused. The tools are the same. They’re at different positions on the same axis.
Writers have it worse than confused. When a developer asks AI for boilerplate, the judgment work is still in the developer’s hands. When a writer asks AI for prose, the judgment work is the prose. There’s no separating them. You can’t hand off the friction and keep the craft.
The rub is where you choose your position. There are two kinds of work in any task:
- Friction work — boilerplate, formatting, the mechanics of construction. Verifiable: it either follows the spec or it doesn’t. Hand it off and the check is almost automatic.
- Judgment work — structure, voice, deciding what’s persuasive. Requires taste. Hand it off and the only way to evaluate the result is to bring your taste to someone else’s decisions. Still rubber-stamping, regardless of who typed the words.
Ask the AI to execute friction work: you’re near the wheel. Ask it to decide judgment work: you’ve started sliding toward rubber-stamping.
The axis has gravity.
Every chat session pulls you toward the rubber-stamp end, because that end feels like less work in the moment. You asked. It answered. You approved. Effortless. You started with the best intentions as Picard. By noon you’re Mickey.
You can’t wrestle this by intent alone. You hold the wheel by installing friction against the gravity: tight scope before you prompt the robots, a decision deadline before you brainstorm, simple plaintext output so there are fewer things to evaluate. Remove that friction and the slide resumes. The machine isn’t pulling you toward the stamp. The path of least resistance is.
Even technical work has limits. Scope matters as much as domain. Code that passes unit tests can still introduce bloat and security issues that only surface under system-level review. And that ain’t cheap — you’ll drift toward the stamp. The best moment to hold your non-negotiables is before you prompt.
AI Brainstorming is a truck-stop pee-break, not a destination. The moment you ask for options, you build an evaluation queue for your brain. Each option without immediate commitment pushes you toward rubber-stamping. You feel like you’re generating; you’re accumulating. Just set a decision deadline before you start. 30 minutes or something. If you haven’t committed by then, close the session. Forced constraint is the tool.
The Cognitive Debt Loop
When you operate at the rubber-stamp end long enough, three compounding harms emerge simultaneously. They feed each other.
Skill atrophy.
When you ratify rather than decide, your judgment tends to get shakier under pressure. The skill doesn’t disappear; it gets weaker. Like a muscle you stopped training. What makes this insidious: your brain registers task completion as a competence win. The work got done, faster than before. You feel productive. You’re not exercising your brain. The cognitive heavy lifting was outsourced. False self-efficacy masks the decay. When you need the brain meat skill, you’ll find it slower than you left it.
Don’t believe me? Try reconstructing the reasoning behind something you delivered with AI-assistance last week. Don’t go back and look at it. The gap between what you can rebuild in your head and what you put your name on is cognitive debt.
Judgment substitution becomes structural.
At first, you’ll be ratifying individual outputs. Your organization will scale this. They will build workflows around AI output. Decision-making becomes AI-first, human-second. You’re no longer the decision-maker; you’re the validator. You’re where outputs get ratified, not where decisions are made. That’s a different job, with lower authority and harder recovery. And once it’s structural, no individual can unilaterally opt back to the steering wheel. The corporate machinery isn’t incentivized to let them.
Attention fragmentation.
Constant evaluation fractures your cognitive capacity. Every context switch costs real recovery time before you can re-enter focus. And the robots will hand you a fresh evaluation task every few minutes. Your ability to sustain productive thinking doesn’t just feel reduced. The interruption pattern makes it harder to re-enter deep focus. Mickey doesn’t get to concentrate. He’s too busy with buckets and brooms.
How the loop closes: Atrophy feeds substitution feeds fragmentation feeds more atrophy. The loop doesn’t break on its own. And nobody notices from the outside. Your output looks great.
How this feels: You start the day with a clear head. By 10 AM, you’ve “saved time” on three emails, a slide deck, and a code snippet: all AI-assisted. By noon, you’re wrecked. Not because the work was hard, but because you spent two hours deciding if the AI was right. Your brain feels like it’s pushing a boulder uphill. The worst part? You’ll do it again tomorrow. The metrics say you’re “more productive.” You can’t argue with metrics.
None of this is new. Everything that’s ever actually helped you focus – music, plaintext, deep work, a stripped-down phone – works the same way: constraint enables focus.
But AI workflows introduce a new kind of choice: not just what to work on, but whether to accept AI suggestions, modify them, or reject them. Every interaction spawns a decision. This decision overhead taxes the attention budget in ways most people don’t see until it’s too late.
If AI operates at the steering wheel end, executing on your direction and cheap to verify, constraint is preserved. If AI operates at the rubber-stamp end, generating output that only taste can judge, constraint collapses into decision fatigue.
AI integration is probably inevitable. Focus is already scarce and getting scarcer. We live in a blizzard of interruptions and app-switching and the godforsaken For You algorithmic feed. The kicker: your organization will optimize for the rubber-stamp by default because it looks like productivity. Higher task completion. Quantifiable metrics. Someone visibly using the tool.
What they won’t see: the soul-depleting triad of suck compounding beneath the surface. Skill atrophy accumulating. Judgment offloading becoming structural. Attention fragmentation making meaningful work impossible. The damage at the individual level is real. The organizational consequences are inevitable. They’re just not visible yet.
We have a choice: design AI collaboration around the steering wheel end where the human directs, AI executes, and evaluation is fast. Where flow is preserved and judgment stays sharp. Or default to the rubber-stamp end, watch task completion soar, and pretend not to notice the people who stayed late to finish work they didn’t actually do.
Will you grip the wheel with all your might or drift because it feels frictionless in the moment, stamping away until there’s nothing left to decide?
Hold the wheel. The moment you hand off judgment to the bots, the brooms are already multiplying.
Judgment is the only thing that can’t be automated. It’s the only reason you remain necessary. Protect it or lose it. There’s no in-between.
