You Can Outsource Your Thinking — But You Can’t Outsource Your Understanding

You can outsource your thinking, but you can’t outsource your understanding. Andrej Karpathy’s line keeps getting quoted online, and every time I see it, I think about the same moment: the answer arrived and felt great, and an hour later I couldn’t explain why any of it was true. That’s the whole meaning of the line, right there. Thinking can be delegated; understanding cannot.

Something still has to direct the thinking, and whatever’s doing the directing is capped by what you actually understand. If you’ve handed off the reasoning too, you’re not directing anything. You’re just forwarding.

So this isn’t a panic piece and it’s not a permission slip. It’s a custody argument. Thinking is delegable labor. Understanding is the thing that has to stay home and direct it. I kept noticing I couldn’t explain the answers I was accepting, and honestly, tracing why turned out to be way more interesting than the usual “AI is melting your brain” discourse.

What “outsourcing thinking” actually means

Here’s the confession beat first, because I did this to myself last week: I asked for an answer, got a great one, felt a small dopamine tick, closed the task. An hour later I couldn’t reconstruct why the conclusion held. The thinking happened. It just happened somewhere that wasn’t me.

And the dangerous part isn’t being wrong, it’s being confident with no retained reasoning behind it. It’s shipping code you can’t debug. You can state the what but not the because, and that gap is the failure mode that bites later. (I’ve unpacked the phrase in more depth in our piece on what outsourcing thinking actually means, but the short version is above.)

Why AI feels different from Googling a fact

Pasting a Googled fact leaves a visible seam. Everyone can see the borrowing. AI output is sneakier: it arrives as fluent first-person prose in your own register, and something in your brain files it as “I worked this out.” The seam disappears.

It’s like a skin that renders over the actual call stack. No stack trace, no borrowing visible, no way to audit where the reasoning came from without deliberately lifting the hood.

I want to be clear that this is first-hand observation, not a lab result. We don’t have good instincts for this yet. New tech, no muscle memory. Curiosity is the right response, not alarm. But no, it’s not the same as Googling, and the missing seam is exactly why.

Engagement isn’t understanding

Karpathy has also pointed out that a lot of educational YouTube gives the appearance of learning while actually being entertainment, which is convenient for everyone involved. Guilty. I watch those too. AI-assisted thinking has the same shape but more intimate: typing a prompt literally feels like cognition, the way watching feels like learning.

Only typing the question isn’t wrestling with it. Engagement isn’t understanding, full stop.

The 2,500-year argument: Feynman and Siddhartha

Karpathy traces this next idea to Feynman: if you can’t build it, you don’t understand it. Very maker-bench energy. The build is the test, because the micro-gaps in your knowledge stay invisible until you try to construct the thing yourself. Everything works fine until you attempt a rebuild from scratch, and then suddenly you discover which parts you only ever skimmed. AI makes it easier than ever to mistake the map for the territory, to feel like having the summary is the same as having the understanding.

Hesse got there first. In Siddhartha, the protagonist hears the Buddha’s teaching, judges it flawless and complete, and walks away anyway. Even the Illustrious One can’t hand over his liberation. Received knowledge isn’t earned knowledge, and the gap between them is exactly your unlived experience. It’s the reason the AI answer never quite fits your actual situation.

So there’s one continuous argument running for two and a half millennia. Feynman says build it. Hesse says live it. Karpathy says AI can do neither for you.

Which makes it a genuinely useful self-test: try to construct an AI answer’s reasoning yourself, and the micro-gaps that appear are the distance between possessing an answer and understanding it. (AI just gave the old argument a speedrun patch.)

What the MIT Media Lab study found about ChatGPT essays

Okay, so check this out. A preliminary MIT Media Lab study (Kosmyna et al., 2024) had students write essays, some with ChatGPT. The ChatGPT group retained less, showed reduced cognitive effort, and produced lower originality. And here’s the number that stops the scroll: 83% of the ChatGPT essay writers couldn’t quote a single sentence from their own submissions moments earlier. Moments. Earlier.

Eric So’s framing of it sticks with me: the essays passed from the computer screen onto the homework assignment without ever entering their brain. That’s the invisible seam with receipts.

Field note: Preliminary study, not proof of rewiring — but the retention gap maps neatly onto skipping the slow System 2 effort.

To be fair to the reader, and to the study: it’s preliminary. Most alarmist coverage drops that caveat, and you deserve better. The researchers observed the pattern; they didn’t prove a permanent rewiring. Still, it maps neatly onto Kahneman.

System 1 is fast, efficient, and often error-prone. Real critical thinking is the slow, disciplined System 2 effort. AI answers default you to System 1. That’s exactly the effort you’re skipping.

AI gravity: why “just use AI responsibly” doesn’t work

Yes, we’re outsourcing more of our thinking to AI, and Eric So’s argument is that the pull is systemic, not a personal moral failing. He calls it “AI gravity”: the constant push-pull to outsource more thinking in pursuit of efficiency, presented at MIT Sloan’s “AI + X” speaker series. He names three drivers: AI’s power amplifying our instinct to conserve mental energy, societal pressure to succeed pushing us to mimic expert performance, and the simple difficulty of detecting when peers are using AI.

That last one is where teams get weird. Once one person’s AI-assisted speed becomes the visible baseline, “just keep up” quietly replaces “show your reasoning.” I’ve seen that pattern described often enough that I don’t think it’s rare.

So’s sharpest warnings are worth keeping intact. The danger isn’t the tool, it’s the mindset it promotes, one that turns users into “ventriloquist dummies.” And: If we can’t think without these machines, I would argue we are not thinking at all. Which leads to an uncomfortable structural conclusion: willpower-based advice can’t work against gravity. Custody has to be designed.

The organizational stakes: skills collapse and lost tacit knowledge

Yes, overreliance on AI genuinely risks skills collapse and the loss of critical thinking, especially wherever the skipped effort compounds, a tension educators and self-learners know well when deciding where assistance becomes a shortcut. So warns it undermines individual learning patterns and business goals, and the mechanism is boringly mechanical: the skill you skip is the skill you lose. Skill atrophy isn’t dramatic. It’s just quiet.

The two failure modes are distinct, and both are worth staring at. Skills collapse is personal: your critical thinking, problem-solving, and judgment degrade because you stopped loading the muscle. Tacit knowledge loss is organizational, and arguably worse. Tacit institutional knowledge is the stuff that never got written down: how a supplier actually behaves when a shipment slips, which manager needs the bad news in writing, which process breaks under load.

It’s built from experience, and it’s most at risk among younger workers who outsource the experience-building process itself. What never got written down is exactly what walks out the door. That’s a continuity threat for any organization that relies on people knowing things.

That third row deserves a beat. Non-augmented capabilities reveal themselves in offline moments: the client meeting, the job interview. The test is simple and slightly horrifying: can you perform unplugged?

Gaffney’s story makes the abstract risk personal. Chris Gaffney, Managing Director of Georgia Tech’s Supply Chain and Logistics Institute and a former executive at Frito-Lay, AJC International, and Coca-Cola, self-identifies as a strong critical thinker. He took an open-source critical thinking test and scored below his own expectations. Not a beginner.

Not a skeptic. An expert who thought he was fine and was not fine. His result prompted deeper inquiry, plus a worry worth carrying: two decades of social media have already shaped how we consume information, and this stacks on top of that.

So works in supply chain contexts where judgment matters: building a forecast, evaluating a supplier, responding to a disruption, modeling risk exposure. These are exactly the decisions with second- and third-order consequences, competing priorities, and interdependencies, where the AI answer runs out of road. His broader point: each act of cognitive outsourcing is small, but they aggregate into dramatic societal change. If you want the rabbit hole, his book The Collision: What AI Does to Us publishes October 2026.

When to let AI think for you, and when it backfires

The decisive difference is stakes: low-stakes factual queries can be delegated immediately; consequential decisions warrant independent thinking first. You can outsource your thinking, but you cannot outsource your understanding, handing off recipes from what’s in the fridge, reliable car models, career pay ranges, lesser-known travel spots, or which TV to buy is fine. Go nuts. That’s not thinking, that’s lookup.

Where it backfires is the original-thought list: where to live, having kids, which goals to set, business versus family time, what to study, a conflict with your boss. Defaulting to advice has two problems. It produces the average answer, which fits poorly on a genuinely unique decision. And it means you never build the decision-making skill, which compounds into worse outcomes over time. Technical debt for your brain.

The relatable tell: facing a city/suburbs/country move, someone’s first instinct is Googling “city vs country living reddit.” That’s a Reddit poster’s own confession, and I recognized myself instantly. Clarity only showed up after real reflection on past living experience, a 3-5 year outlook, and business and network effects. The answer was mostly already in the room.

Which is the reassuring part: you usually already know, or can figure it out. People outsource because introspection is uncomfortable, and I say that as someone who skips it too. The sequencing rule that reconciles everything: think independently first, consult AI or other people only after hitting diminishing returns.

How to stop relying on AI and think for yourself

The fix is keeping the reasoning in the loop. Concrete habits, most of them cheap: the prompt swap, documenting assumptions, summarizing output in your own words, and a thinking journal keep understanding in your head while you still use AI every day.

Think first, prompt second

The signature remedy is one prompt diff: “help me build the reasoning” versus “give me an answer.” Same tool, same question. The first outsources the thinking; the second can leave understanding in you. That’s the entire difference in custody, and it’s honestly kind of elegant.

Stay in the loop as a learner, not just a reviewer. Reviewing output is QA. Learning is being in the build.

Then the six tactics, each bench-tested in spirit:

• Pen and paper. Not nostalgia, threat modeling: the device is the temptation, and Reddit is one tab away. Paper removes the notification surface. There’s a longer argument for this in our piece on keeping AI as a sparring partner, but the short version is just you and what you know.
• First-principles probing questions. Debug your assumptions. On employee tension: is he actually causing it? What’s actually causing it?

  What’s behind that? Anything unconsidered? You may never reach exact fundamental truths, but each question gets you closer to what’s real.

• Ruthless objectivity. Same discipline as reviewing your own code. The check: is this true, or do I just want it to be? The honest example is the country-living downsides I initially neglected: distance from the action, harder relationship-building, pace mismatch, time-consuming maintenance.
• Second- and third-order consequences. What happens after what happens. The “play nice” talk resolves the first-order conflict, and the same problem recurs within weeks, because it was a patch, not a root cause. One caveat: don’t rabbit-hole into infinite chains.

  They’re unpredictable. Consider the likely ones and stop.

• Stick with hard problems. Sometimes 30 minutes, sometimes days, occasionally weeks. Take breaks, go for walks, let background processing work. Persistence beats intensity; don’t do one session and abandon it.
• Outsource only after diminishing returns. This flips AI to the last step instead of the first. Often you don’t need external input at all. When you’re genuinely stuck and need fresh insight, that’s the deliberate moment to borrow other people’s thinking.

Structured thinking before structured answers

There’s an A3-style checklist that works as an actual runnable process, not a methodology seminar: define the problem, clarify what you know and what’s missing, analyze root causes or implications, generate multiple options, set decision criteria, choose and test a path before launch, monitor and adjust. Practicing it on small decisions matters, because skipping step one is where AI answers go wrong fastest, and speed isn’t clarity.

AI hygiene that keeps you in the loop

Document your assumptions before prompting. Journal your intent in two lines. Request counterarguments, alternative views, and sources you can check independently. Look for what’s missing or oversimplified, because that’s where your thinking earns its keep.

Summarize output in your own words; it’s the fastest self-test for whether anything stuck. Track how AI influenced your decisions over time, which is basically a retro for your own cognition. Treat AI as a research assistant, not a strategist: it extends your reach, it doesn’t replace your reasoning.

What Eric So tells teams and organizations to do

So’s research covers AI, behavioral economics, human-computer interaction, and regulatory policy. His four moves apply at both the individual and organizational levels:

1. Value the struggle. Cognitive friction builds skill; removing it causes atrophy and can devalue training and degrees.
2. Value who you are without AI. Communication, reasoning, and negotiating matter in client meetings and job interviews.
3. Reinvest your cognitive surplus. Time saved should fund skill development or new processes, not just faster completion.
4. Make AI a cognitive trainer. Prompt ChatGPT or Claude to problem-solve alongside you rather than hand over answers; drive adoption through awareness, training, and leading by example.

Custody can be designed: the Stanford understanding-checks case

Students can use AI without losing critical thinking when accountability is designed rather than willed, and more than 1,200 Stanford first-years across 70+ sections just demonstrated it, in the 2025-26 academic year, the year the COLLEGE pilot phase came to an end. In COLLEGE 102, they wrote policy memos on how Stanford should handle AI, published May 6, 2026, in an assignment built by Keith Winstein and Emilee Chapman on a tradition running back to a 1921 student petition that produced the Honor Code. Students have been governing themselves for a century. That’s delightful.

The process was a pipeline any engineer would respect: small-group 350-word drafts, section discussion, distillation, cross-section vote. The winning proposal, pitched to Provost Jenny Martinez on May 4 by Neel Ahuja, Peter Vu, Ariadne Vidalakis, and Esme Zeineh, was “understanding checks”: periodic individual assessments through the quarter, graded oral check-ins, supervised problem-solving. Accountability as feedback loop, not gatekeeping. The rejected list is as interesting as the winner: an outright ban (unrealistic), custom per-class models (implementation, adoption, and capability problems), and AI tutors, killed after an LLM asked what an atom was in an advanced science class.

I love that. The demo failed exactly where it mattered most. They also negotiated a workload-neutrality tradeoff and suggested more hands-on, in-person work, which tracks: labs and builds are naturally AI-resistant.

Ariadne Vidalakis put the core plainly: you have to work to develop your own voice and opinions, you can’t outsource critical thinking, and if all the answers look the same, you lose diversity of thought. Martinez praised it as a thoughtful approach to arriving at a smart, nuanced recommendation. Worth noting the students knew their proposal wouldn’t become official policy. The process was the point.

The speaker isn’t anti-AI, and neither is this article

Yes, AI tutoring works when it problem-solves alongside you, and the quote’s originator is the proof this argument isn’t anti-AI. Karpathy has spent the last year building a school premised on AI tutors expanding what people can learn. His own practice is a builder’s use, not a crutch: personal knowledge bases and wikis built from articles, LLMs used for synthetic data generation over fixed data. His framing inverts the whole debate: LLMs are tools to enhance understanding precisely because LLMs don’t excel at understanding. Use them for what they’re good at; keep the understanding yourself.

Cutting through marketing hype is GeekExtreme’s whole editorial practice, and this argument does that for both the AI hype and the AI panic. The tool is fine. The default posture is the problem.

Understanding is a residue that settles after genuine struggle: confusion, dead ends, hard-won clarity. Nobody can hand it to you, which is why the AI answer never quite fits. And here’s the modern twist with a slight chill: Siddhartha knew he was walking away. Invisible-seam AI removes the feeling of borrowing itself.

Two decades of echo chambers and short-form content preconditioned this. AI accelerates it.

So here’s my friendly dare, six things I’d actually try: write out how you form opinions, on paper. Practice structured thinking on small problems weekly. Use AI with intention and never outsource judgment. Teach someone how you reached a conclusion.

Ask “what if I’m wrong?” regularly. Keep a thinking journal for 30 days. Reflective, rigorous, consistent. That’s the edge AI can’t replicate.

Frequently Asked Questions