"AI won't take your job, someone using AI will." But which someone?

"As a programmer you won't lose your job to AI, you'll lose it to a programmer who DOES use AI." It has been doing the rounds for a while, and is usually attributed to the economist Richard Baldwin, who said something similar at the World Economic Forum's Growth Summit in 2023, although Baldwin himself says he is not sure he came up with it (Business Insider via Yahoo).

I agree with it, and I don't often say that about a line that would fit so neatly on a tea towel. But it leaves something open. Everyone has a subscription by now (what such a subscription actually costs the vendor, I looked into earlier), and if using it were enough, everyone would have won already. So which programmer do we mean?

I still build software every week, and with AI it goes about 25 times faster than a few years ago, by my own estimate. I haven't measured that, it is a feeling, and further down you'll see why feeling is a poor yardstick here. But even if it is half that, something shifts. What the model writes in ten minutes, someone has to understand before it goes to production, and that someone is now the bottleneck.

I do that review with AI as well, but with a different model from another vendor entirely: Claude from Anthropic writes, a model from OpenAI reads along. Brand loyalty has nothing to do with it. A model checking its own work makes the same mistake twice, the way you only spot your own typos when a colleague reads your email. The second model finds things the first one misses, and sometimes things that aren't there. (Side note: that last bit is secretly the most useful, because then I have to explain why the code is right after all, and that is how I get to know it.) The decision stays with me.

How misleading that feeling of speed can be, METR showed in 2025. Sixteen experienced open-source developers did 246 real tasks in codebases they knew well, randomly with AI one time and without it the next. With AI they took 19% longer. Beforehand they expected to be 24% faster, and afterwards they still believed they had gained 20%. The study is narrow and early-2025 models are not today's, so I won't make a law of it. The gap between feeling and measuring I do take seriously, including for my own 25 times.

Take two programmers with the same subscription. Bas ships three thousand lines on Friday without knowing why half of them are there. Fatima ships four hundred, has run the other model over them, thrown out two of its five comments because they were wrong, and can explain to the client on Monday what changed. Fatima is the colleague in the saying. What sets her apart is judgement: seeing when something looks right and still isn't. I wrote about that before in Code is getting cheap. Engineering isn't., and every time writing gets faster it weighs more.

The strongest objection isn't about seniors. Brynjolfsson, Chandar and Chen at Stanford looked at payroll data for millions of American workers in 2025. In the occupations most exposed to AI, software development among them, employment for 22 to 25 year olds fell by 13% in relative terms, while for experienced people in the same occupations it held steady or grew. The researchers call that consistent with AI taking entry-level work, not proof. But if it holds, the junior gets exactly what the saying describes: the work goes to a senior who does it on the side with AI, where it used to go to the newcomer.

I have been here before. Around 2004 routine programming work went to India by the bucketload, and the mood was that there would be little left for beginners. A task force of the ACM, the American professional body for computing, looked into how much of that was true in 2006 (summary of the report). The routine work did go first. Job losses from offshoring came to about 2 to 3% of the IT workforce a year, small next to normal churn, and the US had more IT jobs than at the peak of the dot-com boom. The juniors of then are the ones doing the review now. I expect it to recover this time too, because seniors grow out of juniors and nowhere else. I'm not certain, and one difference with back then bothers me: the work went to beginners in Bangalore who also had to learn the trade, and now it goes to a model, which nobody learns anything from.

My prediction: in two years nobody will say "I use AI" any more, just as nobody says they use an editor with syntax colouring. Then half of the saying falls away and what remains is who can judge the result. A company that stops hiring juniors now because the senior with AI does the work on the side will find out in five years, when that senior leaves and nobody can take over the review. I would hire them precisely now and have them reviewing from their first week, with a second model beside them and a senior looking over their shoulder.

Someone who works like that inside your team, with AI and a serious review, is what we offer as staff augmentation. If you want to set up writing with one model and checking with another yourself, we help with that in our AI solutions work. Meanwhile I'm going to try to actually measure those 25 times. If it disappoints, I'll write that up too.