Is AI replacing writers? What actually happened, three years in

The 2023 consensus said yes and it has not held up. Joe Lazer's argument is that AI got far better at coding than at writing because coding has a defined rule structure and a clear definition of success, while good writing does not. He points to job listings for storytellers more than doubling, and to AI companies hiring human writers because their own models struggle to produce stories people want to read.

It is the cost you pay with the person reading, and Joe takes it from a recent meta-analysis. Human communication is not only the words. We read intent from how someone says something, so when an email sounds nothing like the coworker who sent it, the reaction is not that the writing is bad. It is that they seem to have been body snatched. Outsourcing your writing says something about you that you did not mean to say.

The Y Combinator founder wrote on X that he no longer reads an email from a founder when he can tell AI wrote it. Joe's reading of it: what the email communicates is that you do not trust yourself to put your own ideas into words. All you have demonstrated is that you can use AI, which is not impressive, and you have shown you would not stand behind your own voice.

Work slop is when someone takes a model's output, pastes it into a doc and sends it to you, handing you the job of working out whether it was any good. Joe cites research finding around 40% of workers had been on the receiving end of it. Message slop is the newer version, coined by Ryan Levesque: an email chain where three or four people are all writing in the same flat, interchangeable voice, and you cannot tell who is actually in the conversation.

Both Nick and Joe use AI daily and both stopped letting it write first. Nick's practice is to write the draft himself, typos and all, and then ask the model to clean it up, because he does not want it generating the ideas. Joe draws the line at what the reader expects to come from a person. An administrative memo is fine. Anything meant to carry your own perspective is not, and the research says people can tell.

Three reasons, in Joe's account. The training data is mostly not good prose. Human fine-tuners are instructed to steer models toward safer, more jargony, fluffier language. And the labs are making deliberate tradeoffs toward the work that pays, like coding and financial analysis. Under all of it sits a harder problem: defining good writing objectively is close to impossible, because voice is not comma placement. It is the sense that there is a person on the other side of the page.

Joe cites figures from late last year putting art history graduates at roughly 3.5% unemployment and computer science graduates above 8%. Part of it is fewer junior coding jobs. The rest, in his view, is that the learn-to-code push was never quite honest, because the companies making it were hiring out of a small number of elite schools the whole time.

Joe argues no, and he is direct about it. New York's planned AI high school was mocked and then shelved after parents objected. His point is that nobody can say what the AI systems of 2041 will look like, so an eleven-year-old learning today's chatbot is not building a job skill. What does transfer is reading, writing, critical thinking, and the ability to work out whether a source is true, which matters more now that a search gives you a summary instead of a list of links.

Joe taught in New York through Teach For America and is not dogmatic about it. He can see offline leveled reading apps helping a teacher who has been told to run four reading levels in one room of 28 kids. But one AI-first school network found the AI-led part of the day was the part that did not work, and put teachers back into small classrooms, which undercut the promise of cheap private education. Nick's version is a tutor who said nothing when a student claimed he had done the homework, just tilted his head, and the student folded. That accountability to another person is the part that does not transfer.

Relatability, ease, novelty and tension. Relatability is seeing yourself in it. Ease is not having to work to get in. Novelty is being shown something you did not know. Tension is the gap that keeps you there.

Tension, which is the only thing Joe would pass on if he could pass on one. It is the gap between what is and what could be, and storytelling is opening and closing that gap over and over. That is why the near-miss in a heist movie works, and why setting up conflict in your first line beats any hook formula.

Because what makes art meaningful is the choices, and the choices come from a life. Joe's test is how much of you went into it. Prompting a picture into existence is not meaningful in the way a film shot on a real Viking ship is, and we do not ask that of software, where we only want it to work. Used as a tool, where the story and the characters and the thousand small decisions are still yours, he thinks it can be genuinely interesting.