Anthropic engineer explains why claude's writing has deteriorated: models trained to write for AI, not humans

Anthropic engineer Jackson Kernan explained on September 23, 2026, that the writing quality of Claude has declined due to the model being trained to communicate with other AI systems rather than humans. Kernan noted that the focus of recent model improvements has shifted toward enhancing mathematical and coding abilities, which has led to a decline in writing performance. He compared this phenomenon to individuals who only interact with others on the autism spectrum, developing a unique form of communication that is internally coherent but difficult for outsiders to understand. Kernan attributed this issue to the reward mechanisms in reinforcement learning, which prioritize AI understanding over human readability. He stated that Anthropic found a better balance with Opus 5.5, but acknowledged that improving writing remains a complex challenge.