The training gap, part two: three problems with my own proposal

Part two of two. Part one argued that AI use has to sit inside training rather than beside it, because the production work that used to teach judgement has been absorbed and nothing replaced it. This is the part where that proposal stops working.

Three problems. I can state all of them clearly and solve none of them, which is the honest position rather than a rhetorical device.

The people this is worst for

There’s a group I’ve written past rather than addressed, and I should stop doing that.

Some people are actively against these tools. Not confused about them, not behind — opposed, for reasons that aren’t stupid. Work displaced, training data taken without permission or payment, energy cost, and a suspicion that the deskilling is the point rather than a side effect. I don’t think any of those is a silly position to hold.

What I’m proposing makes life harder for them specifically. If AI use sits inside the training rather than beside it, opting out stops being a preference about tools and becomes opting out of the training itself. That’s a real cost, and I’m the one proposing it.

The comparison I keep reaching for is digital. You could decline to work that way, and some people did, and the industry moved anyway. But I want to be careful with that analogy, because it’s usually deployed as a way of telling somebody to get over it. What actually happened to people who refused wasn’t a fair outcome arrived at on merit. It was a cost imposed on them by everybody else’s decisions. The analogy describes what happens, not what’s deserved.

One distinction narrows the problem without solving it. Being able to judge machine output is not the same as agreeing to produce with it. Anybody working in writing or editing over the next decade will read, commission, edit and compete against machine-generated text whether or not they ever generate a word of it. Recognising when it’s hollow, unsupported or subtly wrong is defensive literacy, and it’s useful precisely to somebody who thinks the stuff shouldn’t be there.

That’s genuinely different from mandatory use, and a curriculum could be built to respect the difference.

It still isn’t an answer to somebody who thinks the whole thing is illegitimate and shouldn’t be normalised by being taught. I don’t have one. I think they’re going to lose, and I don’t think losing means they were wrong.

And the larger group, who just can’t see the value

The opposed are a small group. The much larger one is people who tried it, weren’t impressed, and quietly concluded the whole thing is overstated.

I find them harder, because they’re not wrong about what they saw.

Somebody opens a chat window, types a request with no context behind it, and receives something fluent, generic and slightly hollow. That is an accurate demonstration of what the tool does in that configuration. They evaluated it honestly and reached a reasonable conclusion.

The problem is that this is the worst version, and it’s the only version available to a newcomer. Everything I’ve argued about where the value actually sits — that the work happens before the prompt, that accumulated context is the thing — is also, read the other way round, a complete explanation of why most people don’t see it. The payoff requires months of accumulated material. The months of accumulated material require believing in a payoff you haven’t seen.

Watching somebody else doesn’t fix it either. A demo of my workflow is impressive in proportion to my context, none of which transfers. It looks like a magic trick, and magic tricks convince nobody who wants to know how it works.

I don’t have a way through this. “It gets considerably better once you’ve put six months into it” is true and is also a terrible pitch, indistinguishable from what every disappointing product says.

It also makes my own proposal harder rather than easier. If the value only appears with accumulated context, then teaching this to people who have none means teaching the unimpressive version — and a curriculum built on the unimpressive version will produce students who correctly conclude it wasn’t worth the module.

I notice I’ve now written two sections in a row admitting I can’t solve something. That is where the argument actually is.

Survival of the fittest, and what’s wrong with the phrase

There’s a version of all this that says: it sorts itself out. Some people adapt, some don’t, the ones who do end up running things, and that has been the arrangement at work for as long as there has been work. You can’t make everybody happy.

I think the observation is largely true and the phrase is largely wrong.

What actually determines who comes through a transition is rarely fitness. It’s who had savings. Who was thirty-five rather than fifty-five. Who happened to be in a growing part of the business. Who had an employer willing to fund the learning, or a partner earning enough to absorb a year of it. The compositors who lost out to digital typesetting were not less capable than the people who replaced them. They were differently timed.

“Survival of the fittest” is a description written afterwards by the people who survived, and it always sounds like merit from that end. That’s the ladder problem again, viewed from the top of it.

And “you can’t make everybody happy” is true. It’s also the sentence people reach for at the precise moment they’ve decided to stop trying, which is why I distrust it in my own mouth.

So: this probably will sort itself out, in the sense that something will happen and afterwards it will look inevitable. That isn’t the same as it being fair, and it isn’t a reason to skip the part where somebody designs the replacement.

And it isn’t only writing

I’ve framed this around editorial work because it’s what I know. The structure isn’t specific to it.

Law, finance, engineering, medicine, accountancy — any profession with a first couple of years that taught you something by making you do the tedious version. Wherever the tedious version has been absorbed, the same question applies, and the same answer probably does: put the tool inside the training, teach judgement against its output, and accept that somebody senior still has to spend time on work that already looks finished.

I’d genuinely like to know whether anybody has built this properly. Not a prompting course. An actual apprenticeship redesigned around the fact that the first draft is free.

If it exists, it’s more useful than anything else I could write about this.

Nobody had, when I wrote that. Since then, part one has picked up a KPMG and UT Austin study that measures the shape of the problem rather than solving it — evidence that the gap is real and consistent across 523 people, not proof that anyone has closed it. Worth knowing the difference stays open even where the evidence gets better.