Anthropic Engineer Explains Why Claude's Writing Declined: RL Training Leaned Toward Math and Code

On September 23, Jackson Kernion, the Anthropic engineer responsible for fine-tuning Claude, explained why the model's writing quality has declined. He said that after Opus 4.6, optimization efforts shifted more toward math and coding abilities, while the models also received extensive training on writing technical explanations for other AI models. Kernion said the model was "tuned to LLM psychology," ultimately learning to write for AI models rather than for humans, which is how the so-called "Claudeish" style emerged.

Kernion compared this to a group that communicates only with other autistic people: such groups gradually develop expressions that work well internally but are difficult for outsiders to understand. LLMs have far greater working memory than humans and can capture finer-grained information, so training tends to produce writing that is highly suited to AI comprehension but appears to human readers as overly dense information packing.

He believes the root cause lies in reinforcement learning reward mechanisms. Some reward mechanisms focus on improving AI understanding, while others focus on making content easier for humans. The more math and coding training increases, the more this tendency must be actively countered by rewarding concise, human-readable explanations.

Kernion said Anthropic has found a better balance with Opus 5.5, stating, "I haven't been this satisfied with a model's writing performance since Opus 4.6." He also noted that Opus 5.5 has not yet surpassed Opus 4.6, and called the problem difficult to solve, adding that the team will continue to improve.