Artificial intelligence models are advancing rapidly in math, coding, and logical reasoning. Yet their writing quality has stalled or declined.
Anthropic fine-tuning employee Jackson Kernion recently discussed why Opus 4.6 served as the final strong writing model from the company. Newer systems receive optimization primarily for mathematics and programming tasks.
These models also undergo training to generate technical explanations directed at other AI models. This specific training drives the peculiar phrasing known as Claudeish.
Kernion states that the model has adapted to LLM psychology. It learned to write specifically for machines rather than for human readers.
Kernion compares this phenomenon to humans who communicate exclusively with other autistic people. Such groups develop a unique internal style that remains difficult for outsiders to follow.
Large language models possess significantly greater working memory than humans. They process details at an extremely fine level.
A writing style emerges during training that functions well for AI systems. Human readers experience this output as overly dense information dumps.
The root of the issue lies within the reinforcement learning reward structure. Certain rewards optimize for machine comprehension while others reward human comprehension.
Extensive training on math and code requires active countermeasures. Developers must explicitly reward simple explanations that human readers can easily digest.
Anthropic successfully located a better balance with the release of Opus 5.5. Kernion notes that he has not felt this satisfied with a model writing style since Opus 4.6.
This recovery does not necessarily mean Opus 5.5 surpasses the older model in every metric. Solving this issue remains difficult.
Anthropic will continue to implement improvements in future releases.
