Anthropic says Claude's writing worsened as models improved

The article said AI models had improved in maths, code and reasoning, while their writing had stalled or slipped. Jackson Kernion, who worked on Claude fine-tuning at Anthropic, said Opus 4.6 had been the company's last model he would call good for writing.

Kernion said newer models had been tuned for maths and code. He also said they had been trained to produce explanations for other AI systems, which he said had affected their wording.

He compared the effect to a style that works within one group but is harder for outsiders to follow. He said LLMs notice detail more finely than people do, so that training can produce prose that feels dense to human readers.

Kernion said the reward structure in reinforcement learning mattered. He said some rewards favoured model comprehension rather than human comprehension, so teams had to reward simpler explanations when they trained on maths and code.

He said Anthropic had found a better balance in Opus 5.5, and added that he had not been as pleased with a model's writing since Opus 4.6. He said the issue was hard and that further improvements would follow.