When Your Thinking Partner Becomes Your Autopilot: The Cognitive Cost of Working With AI

elisabeth hitz · july 19, 2026 · 7 min read

AI does not make you think less. It makes you think differently, and the risk is that the change is invisible, you feel more capable while understanding less. The research on this is decades old and predates AI entirely. It is the most important thing nobody in the AI content economy will tell you, because it does not sell anything.

I run most of my business through AI. I also think the honest case against it is worth reading, because the operators who last are the ones who noticed the trade before it cost them something.

Does using AI make you worse at thinking?

Not automatically. But four well-established findings suggest the risk is real and specific.

1. Removing difficulty too early degrades learning. Cognitive Load Theory (Sweller, 1988) shows learning depends on a balance between challenge and support. When tools eliminate difficulty before you have engaged with it, they can remove the conditions that produce durable understanding. Some friction is not waste. It is where learning happens.

2. Ease of understanding is not understanding. Metacognition research (Dunlosky & Metcalfe, 2009) shows that fluency, how easy something feels, systematically misleads people about how well they actually grasp it. AI output is extremely fluent. It can produce a strong sense of competence without the underlying comprehension to support it, and you will not notice, because the feeling of understanding is the same either way.

3. The mere availability of a tool consumes attention. Ward et al. (2017) found that the presence of one's own smartphone reduced available cognitive capacity, without being used. An always-ready AI assistant plausibly functions the same way, changing how often you pause, sit with difficulty, or reflect before reaching for an answer.

4. Over-assistance erodes motivation. Self-Determination Theory (Deci & Ryan, 2000) identifies autonomy, competence, and relatedness as core psychological needs. Tools that over-direct and over-complete can quietly undermine autonomy, particularly for people already navigating uncertainty or low confidence, which describes most people building something alone.

None of this is an argument against AI. It is an argument for noticing.

What is the actual risk for someone running a business?

Not that you get dumber. That you lose the judgment you are actually selling.

Here is the mechanism. Your business runs on domain judgment, knowing which customer is worth chasing, which offer to lead with, when a piece of work is good enough. That judgment was built by doing the work badly, noticing, and adjusting. It is expensive and it is the thing clients pay for.

If AI does the first draft of every decision, you stop generating the reps that built the judgment. And you will not notice for a long time, because the output stays good, right up until a situation the model handles badly and you no longer have the instinct to catch it.

Trust research points at the same failure. Lee & See (2004) found the problem in human-automation systems is not trust or skepticism, but miscalibrated trust. Systems that feel authoritative while being context-blind get over-relied upon, or rejected entirely after one visible failure. Both are expensive. The correct posture is calibrated, and calibration requires you to still be able to evaluate the output.

How do you use AI without outsourcing your thinking?

Five practices. I use all of them; none are theoretical.

1. Form your own answer before you ask. Even a bad one. Thirty seconds. This preserves the reasoning rep and, as a side effect, produces a far better question, which produces a far better answer. It is the rare habit that costs nothing and improves both sides.

2. Ask it to interrogate you, not answer you. "Ask me five questions before you respond." This inverts the relationship, you stay the author, the model becomes an editor. It is the difference between a thinking partner and a cognitive autopilot, and it is one sentence.

3. Delegate execution, keep the decisions. Formatting, drafting, summarizing, converting, restructuring, delegate all of it, without guilt. What to build, who to serve, what to charge, when to walk away, keep all of it. The line is not "hard versus easy." It is reversible versus not.

4. Set explicit ignorance boundaries. Write down the things you refuse to stop being able to do yourself. Mine include understanding my own numbers and writing my own opening lines. Yours will differ. The point is that undefined boundaries erode silently, and defined ones do not.

5. Notice fluency. When output feels obviously right, that is the moment to slow down, not speed up. Fluency is the signal that misleads. If you cannot explain why it is right in your own words, you have not evaluated it, you have just enjoyed it.

Is this an argument against building AI systems?

No, and I want to be exact about the distinction, because it is the whole point.

Delegating execution to a well-built system preserves your judgment. Delegating judgment to a fluent chat interface erodes it.

This is precisely why files beat chats, and it is a reason I do not see argued anywhere. A context file is a decision you made deliberately, wrote down, and can revisit. A chat is a decision you outsourced in the moment and cannot reconstruct. The file version keeps you the author. The chat version makes you a reviewer of your own business.

Anthropic's engineering guidance frames context engineering as deciding what information the model should have. That framing is only coherent if a human is still doing the deciding. The system is supposed to hold your judgment, not replace it. Structure in files not chats.

Why this matters commercially, not just personally

Two reasons, and both hit your revenue.

One: your judgment is your margin. In a market where everyone has the same models, the person who can tell good output from plausible output charges more than the person who cannot. That discrimination is a skill, it decays without use, and it is the last thing that will be commoditized.

Two: this is your E-E-A-T. Search and AI systems both increasingly discount generic synthesis and reward demonstrated first-hand experience. The operator who has actually done the work and can say something specific and non-obvious about it gets cited. The operator who has been having AI think for them produces content that reads exactly like the content models were trained on, and therefore gets absorbed rather than attributed. I covered that mechanism in why you can't optimize for a browser.

The thing that protects your thinking and the thing that makes you visible are the same thing. That is not a coincidence.

The version I actually believe

The risk is not that AI thinks for us. It is that our thinking changes and we do not notice, because the output keeps looking fine.

The fix is not less AI. It is more deliberate AI: knowing which part of the work is yours, writing that part down, and building systems that carry your decisions forward rather than making them for you.

Where to think this through with other people

Reading about this changes nothing. Watching how other operators draw the line does.

The Closer Method builder community on Skool is where that conversation happens, people running real businesses on AI systems, comparing what they delegate, what they refuse to delegate, and what it cost them to learn the difference. A forum, not a course. Just operators being honest in writing.

or just follow along. new field notes most weeks on x, instagram, and tiktok.