Anthropic Engineer Explains Claude's Declining Writing Quality
Anthropic's Claude AI model's writing quality has declined because it was trained to write for other AI models, not humans.
While AI models have rapidly advanced in mathematics, programming, and reasoning, their writing capabilities have stagnated or even declined. Anthropic's Claude model is no exception.
Jackson Kernion, an engineer at Anthropic responsible for Claude's fine-tuning, explained that the focus for subsequent models shifted more towards mathematical and coding abilities. This adjustment impacted the model's writing performance.
Kernion stated that these models underwent extensive training to generate technical explanations for other AI models. This led to the development of what is known as "Claudeish" writing, optimized for AI comprehension but potentially appearing peculiar or overly dense to human readers.
"The models were tuned to be adapted to LLM psychology," Kernion said, comparing the phenomenon to how a group communicating solely amongst themselves might develop a specialized communication style difficult for outsiders to grasp. The advanced capabilities of LLMs, such as greater working memory, contribute to this AI-centric writing style.
Kernion identified the reinforcement learning reward mechanisms as the root cause. Optimizing for AI understanding can inadvertently lead to text generation tailored for machine processing. However, Anthropic has achieved a better balance with Opus 5.5, with Kernion expressing satisfaction with its writing performance since Opus 4.6.