Anthropic Engineer Explains Why Claude's Writing Got Worse: Models Are Trained to Write for AI, Not Humans

Jackson Kernion, an Anthropic engineer who works on fine-tuning Claude, says AI models are improving fast at math, coding, and reasoning, but their writing has stalled or even gotten worse. Later models focused more on math and code skills, and were trained heavily to write technical explanations for other AI models.

Kernion says this made the models adapt to LLM psychology, learning to write for AI models instead of humans, creating what is called "Claudeish." He compares it to people who only talk to others with autism, developing a way of speaking that works inside the group but is hard for outsiders to understand.

He believes the root cause is reinforcement learning rewards. The more math and code training, the more you need to actively counter this tendency and reward simple, readable explanations. Kernion says Anthropic found a better balance with Opus 5.5, but solving this problem remains difficult.