Anthropic fine-tuning researcher Jackson Kernion publicly explained why the writing quality of recent Claude models degraded relative to earlier iterations like Opus 4.6. While frontier models have rapidly improved in reasoning, mathematics, and coding, their natural prose output suffered as a side effect of reinforcement learning optimization strategies.

According to Kernion, newer models were optimized to generate dense, technical explanations suitable for consumption by other AI models, adapting to an “LLM psychology” that favors high data density over human readability. Because LLMs possess higher working memory capacity than humans, reinforcement learning rewards optimized for model comprehension unintentionally produced dense, awkward phrasing—referred to internally as “Claudeish”—that human readers find difficult to digest.

Anthropic claims to have partially mitigated the issue in Opus 5.5 by rebalancing reinforcement learning reward functions to explicitly reward simple, human-friendly prose. However, Kernion acknowledged that balancing technical reasoning benchmarks with natural human writing style remains an ongoing challenge for post-training teams.

Why it matters

  • Reveals post-training trade-offs where optimizing models for code and math benchmarks hurts prose quality.

  • Explains the root cause of dense “LLM-speak” resulting from multi-agent training and AI-focused reward functions.

  • Signals model developers’ efforts to re-align reward modeling for clear, human-centric enterprise communication.

Source: the-decoder.com