Knowledge is cheap, judgment isn't
- 8 minsRecently, Meta started installing software on its own employees’ computers that tracks everything they do. Every keystroke, every click, every screen, captured as they move through Slack, GitHub, Google, the same tools you and I use all day. The reason was almost honest in its bluntness: to train its AI agents, the company said, the models “need real examples of how people actually use” a computer. So the people who work there are, as literally as the words allow, teaching the thing that’s coming for their work. And they don’t get a vote.
Training your replacement used to be the last thing you did before you left a company for good. Now you do it just by showing up to work. When I read about it, I wasn’t even angry. It felt obvious, the natural next step, the kind of thing you can’t really fight.
The interesting part is what it can’t track. It captures every single thing those employees do and still misses the only thing that matters.
Most people meet this moment with one question: will AI take my job? It’s the natural fear, and I don’t think it’s the useful one. Underneath, the fear comes down to one thing: a skill that was scarce, that took years to build, suddenly costs almost nothing. Tucked inside it is an assumption: that when a skill gets cheap, the job just disappears. That’s not exactly what cheap has ever done.
Step back for a second. For most of history, knowing things was the moat. The lawyer, the doctor, the engineer got paid well because the knowledge in their heads was scarce and took years to load in. What’s changing now is that the knowing itself is going cheap (in effort, not just in price). The decade it took to learn a craft, the specialist you once had to find and pay, now collapses into a sentence and three seconds, reproducible by anyone. It happened to distance. Reaching someone across the world once took real money and effort, phone operators and per-minute charges; now you video-call family across an ocean without a thought. And the moment anyone can reproduce something with little effort, no one pays a premium for it. The real question is where the value goes instead.
The pattern is so consistent it’s almost boring: when a layer of value gets cheap, the pyramid doesn’t flatten. It lifts.
The cheap thing becomes the floor everyone stands on, and the prize climbs to whatever is still scarce just above it. Storage got cheap, so storing raw data stopped being worth anything and the value moved up to organizing it. The internet made information free, so the value climbed again, to knowing what the information meant. Now AI is making that knowing cheap, and the prize is moving up one more rung: not just what to do with what you know, but when to do it, and when not to.
And this is just what economics does everywhere, nothing special about AI. Take writing. Before the internet, the hard part was getting published at all; a handful of publishers and printing presses were the gate, and clearing it was the whole game. Then the internet made publishing free, anyone could hit send, and the scarce thing flipped to attention. Once everyone could publish, simply being read became the prize. The writers who won it had built an audience, not necessarily a better sentence. Manufacturing went the same way: when factories everywhere could build the thing, the value fled the assembly line and climbed into design and brand. Apple doesn’t own the factories that build iPhones; it owns the part nobody can copy. The biggest winner of the whole AI boom is a chipmaker, not a model. Nvidia happens to own the one rung still scarce, the silicon the whole thing runs on. Cheap below, value climbs, whether you’re a writer, a factory, or a trillion-dollar chipmaker.
So far this sounds like good news: just move up, climb to the next layer. And for the system as a whole, it is: the market grows, the work reshuffles, in the aggregate it has always worked out.
The individual story is where it gets hard, and you rarely hear anyone say this part out loud. The reason a senior engineer is worth more than a junior isn’t that she knows more facts — it’s that she’s seen enough go wrong to feel which way a system will break before it breaks. That instinct has a name, judgment, and there’s only one way anyone has ever built it: by doing the grunt work first. Chasing a bug through code you didn’t write. Fixing the flaky test at midnight. Reading the gnarly old module until it finally makes sense. The boring reps are where the judgment quietly accumulates. And the boring reps are exactly what AI eats first. We’re pulling the bottom rungs off the ladder while telling everyone to climb.
Which means the climb doesn’t happen to you automatically anymore. You have to choose it.
Using AI coding tools every day, I feel the pull myself: the machine is right there, faster than me, and saying yes is effortless. The gym is what taught me the trap: a forklift could move the weight for you, but then you don’t get stronger. Outsource every rep and the “judgement” muscle you quietly stop building wastes away.
But the fix isn’t to put the tools down. I’m not going to, and neither are you. It’s the rule we already use with calculators. We hand kids a calculator only after they can do the math by hand, not because they’ll ever do long division again, but because someone who doesn’t understand what the calculator is doing can’t tell when it’s wrong. That’s the whole move. Use the machine all you want, but never stop being able to answer one question: if I had to do this myself, would I know how? When it hands you something, you should be able to follow why it went that way and feel in your gut whether it holds up. Use the tool Don’t hand it your thinking, because there’s one line the machine can’t cross. It can do what you do. It can’t do why you did it — and the why is the whole job.
AI is brilliant at the closed problems, the ones with a clear goal and a path to it. Write the function, fix the failing test, wire up the API. That’s the what, the stuff a camera over your shoulder could capture. Where it stalls is the open problems, the why: whether to build the feature at all, whether this design will buckle under real load two years from now, how to tell a teammate the approach they’re proud of is the wrong one. AI can write the resignation letter in a perfect, polished minute. It can’t tell you whether to quit.
I see the gap already. A junior engineer ships a pile of AI-generated code and, asked how it works, can’t tell me. The code runs fine. The understanding never got built. That’s the danger nobody priced in: not AI replacing people, but people arriving at senior titles with none of the judgment the title used to require. Confidence without judgment is the most dangerous thing AI is mass-producing right now.
None of this is new. When the printing press arrived, the scribes who’d spent their lives copying books could see what was coming, and many fought it; some even smashed the presses in futile anger. The copying was never coming back. But the press still needed letters, and nobody understood letters like the people who had drawn every one by hand. The scripts the scribes wrote became the first typefaces: italic type began as the cursive of a Roman chancery scribe, cut into metal. They didn’t out-copy the machine; they designed what it printed. Years of shaping each letter by hand had become the very thing the press ran on. The drudgery was the training. The difference now is that the drudgery is optional, which means the training has to be deliberate.
We’re standing in the same doorway. The machine is already here, recording our keystrokes, and like Meta’s employees, we don’t get to vote on that. So don’t waste yourself trying to smash it. Let it do the copying. But the moment you stop understanding what it’s doing for you, the moment the tool becomes a wrapper you can’t see inside, you’ve handed over the only thing that was ever yours. The understanding underneath is the judgment, and the judgment is the why no recording can lift off you.
The question to sit with isn’t whether you’ll be replaced. It’s quieter than that: of everything you did today, how much did you actually understand — and how much did you just accept because it ran?
What do you think? Shoot me an email with your thoughts!
Sandeep Raju Prabhakar
Writes about technology, software engineering and other things that interests him.