There’s a version of this that ends badly. Someone buys a GPS unit, uses it on local trails for a few months, gets comfortable, and decides they’re ready to go off trail in the Absarokas. Three miles from the last cairn, the battery dies. The signal drops. They’re standing in a drainage with no idea which ridge they came over or how far they are from the trailhead. The GPS didn’t fail them. They failed to develop the skill the GPS was supplementing.
This is the conversation we should be having about AI, and we’re mostly not having it.
The dominant anxiety is about replacement. AI writes, therefore writers are obsolete. AI analyzes contracts, therefore lawyers are obsolete. AI generates code, therefore developers are obsolete. It’s a tidy narrative and it’s wrong about what’s actually happening, because it confuses the tool with the expertise required to use it well.
The GPS is easy to operate. A child can enter a destination and follow the arrow. But a skilled wilderness navigator using the same device brings something the device can’t supply: a mental model of the terrain. They’re reading contour lines, anticipating water sources, understanding why a slope that looks gentle on the map is a different story after a week of rain. The GPS tells them where they are. The underlying skill tells them what that means and what to do next.
That asymmetry scales to every domain where AI is making inroads.
A physician using an AI diagnostic tool doesn’t just read the output. She knows which differential diagnoses are plausible given this patient’s history, which symptoms the model is probably overweighting, and when the confident-sounding recommendation is almost certainly wrong. A litigator using AI to research case law knows how to evaluate whether the cases are analogous, whether the jurisdiction matters, and whether the argument will land with this judge. A financial analyst using AI-generated models knows which assumptions to stress-test and which outputs are garbage dressed in a clean table. In each case, the expert with the tool gets supercharged. The novice with the same tool gets a plausible-sounding answer they have no basis to evaluate.
This is the thing worth worrying about, not that AI replaces expertise, but that it makes the absence of expertise harder to detect. A wrong answer that looks authoritative is more dangerous than a wrong answer that looks uncertain. AI is very good at looking authoritative.
There’s a generative asymmetry here too. The person with the underlying skill knows what questions to ask. They can prompt precisely because they know what a good answer looks like, what’s missing from a weak one, and when to push back. The person without that foundation is hoping the tool figures it out. Sometimes it does. Often enough that the feedback loop never closes and they never realize how much they’re missing.
For writing specifically, this matters in a way that goes beyond the professional stakes. Writing isn’t just output. It’s a forcing function for thought. The act of trying to articulate something, struggling with the structure, cutting the paragraph that felt important but turned out to be throat-clearing, discovering mid-draft that you don’t actually believe the argument you started with, all of that is where the thinking happens. AI can generate prose. It cannot do that work for you. And if you skip it, you don’t just produce worse writing. You develop less of whatever it is that produces good thinking in the first place.
The people who will thrive with these tools are the ones who built the underlying capability before the shortcut existed, or who are disciplined enough to build it anyway. Give a skilled wilderness navigator a GPS and it supercharges everything they already know. Give someone who’s never learned to read a map the same device and you’ve handed them a dangerous confidence in unfamiliar terrain.
The technology isn’t the variable. The foundation is.