A model handed me a citation last week so convincing I almost shipped it. The paper had a journal, a volume, a page range, two authors with the right kind of names, a finding that fit my argument like it had been measured to order. I had the cursor over send. Then some small custodial instinct made me search for the title. The paper did not exist. Not "the link was stale." Did not exist. The model had assembled a perfectly shaped crate out of nothing and slid it down the belt toward me, and I, the inspector, had nearly stamped it through.
That moment is the whole story of the next decade, compressed into thirty seconds.
For most of the history of work, the bottleneck was production. Writing the memo was the hard part; reading it was free. Building the model was the labor; sanity-checking the output was a glance. We organized entire professions around the cost of making things — analysts who could produce, associates who could draft, juniors who could grind out the first version so a senior could react to it. The senior's edge was taste, but taste sat downstream of someone else's sweat.
That arrangement is now inverting, and faster than most org charts have noticed. Generation has gone nearly free. A reasoning model will produce a plausible market analysis, a coding agent will produce a plausible pull request, a research tool will produce a plausible literature review — all in the time it takes to make coffee. The crates arrive by the thousand. They are well-built. Most of them are even correct. And the only thing standing between a company and the ones that aren't is whoever opens them.
The inspector at the dock
Think of a customs checkpoint. Goods flow in from everywhere, declared as one thing, sometimes containing another. Anyone can wave a crate through — waving requires no skill, and on a good day the waver looks exactly as productive as the inspector. The value of the operation lives entirely in the person who can open the right crate quickly, know what should be inside it, and spot the one that's wrong before it reaches the floor.
That is the job now. Not producing the crate. Inspecting it.
We have spent two years marveling at how much these systems can make. We have spent almost no time training the muscle that decides whether what they made is real. And here is the uncomfortable part: those are different muscles. The person who writes a beautiful paragraph and the person who can tell, in four seconds, that a beautiful paragraph is quietly lying — these are not the same person, and we have built schools, careers, and hiring funnels almost entirely around the first one.
Why checking is harder than it looks
Verification feels like a lesser task. It sounds janitorial. It is, in fact, the more cognitively demanding of the two, because plausibility is precisely the thing models are now superhumanly good at manufacturing. A wrong answer used to announce itself — it was clumsy, it was off in tone, it had the texture of error. Now the wrong answer wears the same suit as the right one. It cites. It hedges in the right places. It has the cadence of competence.
To catch it you need something the model can fake and you cannot: a real grip on the territory. You have to know what a correct answer would actually look like before you read the one in front of you. You have to hold a prior. The citation almost got past me not because I'm careless but because I wanted that finding to be true, and a system that produces exactly what you want is the most dangerous kind of crate on the dock.
So the new literacy isn't prompting. Prompting is asking the dock for more crates. The literacy is the inspection — the trained, fast, slightly suspicious read that separates true from merely true-sounding. It rests on three things: domain knowledge deep enough to hold a prior, a reflex to check the load-bearing claim rather than the decorative ones, and the discipline to do it even when the crate is flattering.
What this does to the org chart
Run this forward and the shape of teams changes. The leverage stops being in how many people can produce and starts being in how few people you need to verify what's produced. One strong inspector with good tools can clear the output of a dozen tireless generators. That ratio is the actual economics of the next few years, whatever the slide decks say about headcount and automation.
It also reorders who is valuable. The junior whose only edge was that they'd grind out the draft has lost their edge — the machine grinds for free now. But the person who was never the fastest writer in the room, who was instead the one who'd quietly say wait, is that number right? — that person was undervalued for decades, and is about to be repriced. At my own company we've started hiring for it on purpose, weighting the interview toward "here's a plausible-looking output, find what's wrong" over "here's a blank page, fill it." The blank page is the cheap part now.
I think this generalizes well past software and analysts. Any field where AI can now produce the artifact — legal drafts, diagnoses, designs, code, financial models — splits into the same two roles: the generators, which collapse toward free, and the verifiers, which become the scarce, defensible, well-paid work. Toffler said the illiterate of the future wouldn't be those who couldn't read, but those who couldn't learn, unlearn, and relearn. The 2025 edition is narrower and sharper: the illiterate will be those who can't tell a true thing from a thing that merely sounds true, at the speed the crates now arrive.
The trainable part
Here's the part I find hopeful, and it's why I call this literacy rather than talent. You can teach it. Reading wasn't a gift either; it was a skill we decided was important enough to drill into every child until it became invisible. Verification is the same kind of thing — learnable, practicable, gradable. You build it the way you build any reflex: by being wrong in public a few times, by developing the small custodial instinct that made me search for that title, by training yourself to feel the specific discomfort of a claim you haven't actually opened.
The mistake would be to treat this as a defensive crouch, a tax we pay for trusting machines. It's the opposite. The person who can verify at speed gets to use generation at full throttle without getting hurt by it. They can let the dock run hot, accept the flood of crates, and stay safe — because they can open any one of them in seconds. Verification isn't the brake on AI leverage. It's the thing that lets you floor it.
So the question to ask about yourself, and about anyone you'd hire, is no longer what can you make? The machines made that question boring. The question is what can you catch? I almost shipped a paper that didn't exist. The skill that saved me wasn't writing. It was opening the crate.
Learn to open the crate.