How do you break free from AI dependency?

Why Outsourcing Every Decision to AI Costs You Judgment

Naval Ravikant argues judgment, not effort, creates value, and every decision you hand fully to Claude or ChatGPT instead of forming your own first is a rep of judgment you didn't build. Break the pattern with a simple rule: write your own one-line answer before you ask the model anything, then compare. Do that for two weeks and you'll see exactly where you've outsourced thinking versus augmented it.

A founder I coach admitted something last month that I think about constantly now: she couldn't remember the last time she'd made a pricing decision without running it past Claude first. Not because Claude was smarter than her, but because asking felt safer than deciding. That's the real shape of AI dependency for high performers. It's rarely "I can't function without it." It's "I don't trust my own first answer anymore," which is a quieter, more corrosive problem, and it's exactly what Naval Ravikant is talking about in The Almanack of Naval Ravikant when he says judgment, not hours worked, is what actually creates value.

Naval's argument, compressed: in a world where execution is cheap and getting cheaper — which describes 2026's AI tooling perfectly — the scarce resource is judgment, the ability to make correct calls with incomplete information. "Specific knowledge," in his framing, is built the way a kid learns to play: through direct, repeated, often uncomfortable contact with a domain, not by reading someone else's synthesis of it. Every time you skip straight to a model's answer instead of forming your own first, you skip the rep that builds specific knowledge. You get the output. You don't get the judgment. And judgment is the thing that was never going to be replaced — it's the thing you were supposed to be building this whole time.

The tell isn't usage volume. I use Claude's Fable 5 model and ChatGPT dozens of times a day and don't consider myself dependent in the way that matters. The tell is whether the model changes your answer or just makes you more confident repeating it. Those are opposite outcomes wearing the same UI. So here's the protocol I now run with clients who suspect they've drifted into the second one — I call it the Judgment Ledger, and it takes about ninety seconds a day:

1. Answer first, alone. Before you open Claude or ChatGPT on any real decision — pricing, a hire, a hard email — write one line with your own best guess. No research, no lookup. Thirty seconds, timestamped.

2. Ask, then diff. Now ask the model. Compare its answer to yours. Did it genuinely change your reasoning, or did it just restate your instinct with better vocabulary and more confidence? Note which.

3. Log the outcome, weekly. A week later, note what actually happened. Over a month you get a real map: the decision categories where your instinct beats the model (usually anything involving people you know well) and the categories where the model consistently adds real information you didn't have (usually anything involving unfamiliar markets or technical specifics).

Run this for two weeks and the pattern gets uncomfortable fast, in a useful way. Most people I've walked through this discover the model almost never changes their reasoning on people-decisions — it just launders the anxiety of deciding alone into something that feels externally validated. That's the dependency Naval's framework predicts: you're not gaining judgment, you're renting confidence, and confidence you rent has to be re-rented every single time.

None of this is an argument against using AI heavily. Naval himself built an entire wealth philosophy on leverage, and code and models are the most permissionless leverage that's ever existed. The argument is narrower: leverage multiplies judgment, it doesn't replace the need to build it. If you've never sat with your own uncomfortable first answer, there's nothing for the leverage to multiply. The founders I've watched get worse at their jobs while using AI heavily are the ones who quietly stopped forming opinions before they went looking for the model's. The ones who got sharper are the ones treating the model like Naval treats a mentor: useful for pressure-testing a view you already have the courage to hold first.

Sources: Eric Jorgenson, The Almanack of Naval Ravikant (Magrathea Publishing, 2020); Gallup, "AI Adoption and Productivity" (July 2026), on the gap between individual AI use and organizational judgment; claude.com/pricing, for current Claude model access.