Everyone had the same prediction. Writing would be the first casualty of the AI revolution. It was the obvious call: language models generate text, writers generate text, therefore writers go first. Clean logic. Wrong conclusion.
Andrej Karpathy, co-founder of OpenAI and one of the most credible voices on AI capability, published a thread on April 10 that explains precisely why this prediction failed. And his explanation is worth sitting with, because it reframes how we should think about where AI is actually winning.
The Perception Gap Nobody is Talking About
Karpathy’s first observation is about how unevenly AI capability is understood. A large group of people tried the free tier of ChatGPT at some point, saw it fumble basic questions, laughed at the hallucinations, and moved on. That experience formed their worldview on what AI can do.
A much smaller group uses frontier agentic tools like OpenAI Codex and Claude Code professionally, in technical domains, every day. They are watching these models restructure entire codebases, solve problems that would have taken weeks, and do it in an hour. For this second group, the experience borders on disorienting.
These two groups are not disagreeing about AI. They are talking about completely different products and calling them by the same name.
Why Coding Went First, Not Writing
The deeper point in Karpathy’s thread is structural. Coding has verifiable reward functions. When AI writes code, you can run a unit test. The test passes or it fails. That binary signal is exactly what reinforcement learning needs to improve. You can train a model on billions of correct and incorrect outcomes, and the gradient knows which direction to move.
Writing has no equivalent. There is no unit test for a good sentence. There is no pass/fail for whether a paragraph is compelling. Human judgment is the only signal, and human judgment is expensive, inconsistent, and impossible to scale the way automated tests are. RL cannot hillclimb on “this story resonated emotionally.” So it doesn’t.
Add to this the economic reality Karpathy points out: coding is where the B2B value is. The largest fraction of every AI lab’s engineering team is focused on improving coding performance because that is where enterprise revenue comes from. Writing improvements are slower, harder to measure, and lower on the priority list. Not because writing doesn’t matter, but because the incentive structure points elsewhere.
Also Read: How AI Is Replacing Personal Note-Taking
What This Means for Writers
The conclusion some will draw is that writing is safe. That’s the wrong takeaway. AI writing tools have gotten genuinely good at surface-level tasks: summarizing, drafting, formatting, rephrasing. Those tasks are being automated, and that pressure is real.
What is not being automated, and what is structurally difficult to automate, is the thing that makes writing worth reading: voice, narrative arc, the particular way a writer sees the world and chooses what to leave out. That is not surviving because it is artistically special. It is surviving because it cannot be scored. You cannot write a loss function for a great essay the way you can write a test suite for a working API.
This distinction matters. Writers who understand it will focus on developing exactly the capabilities that resist automation: original perspective, specific observation, structural judgment. Writers who miss it will compete against models on tasks the models are already winning.
The Real Lesson from Karpathy’s Thread
The gap in understanding AI capability is not just about technology. It’s about access. The people who have genuinely seen what AI can do are paying $200 a month and using it in highly technical environments. Most people are not. Their mental model of AI is outdated by roughly two to three product generations.
That gap is closing as agentic tools become more visible. But until it does, we will keep having the wrong conversation: debating whether AI is overhyped based on a version of the technology that no longer exists.
The capability is real. It is just unevenly distributed, and unevenly understood.
