Anthropic engineer links Claude’s awkward writing to AI-focused training
Anthropic employee Jackson Kernion, who works on Claude fine-tuning, says the company’s newer models have become stronger at mathematics, coding and reasoning while their writing has stagnated or declined. He described Opus 4.6 as Anthropic’s last notably strong “writing model.”
Kernion attributes the change not only to technical optimization, but also to training models to produce explanations suited to other AI systems. Because language models can retain and process more detail than people, this training can encourage a style that works for machines but appears to human readers as “overly-dense info dumps” and produces unusual phrasing sometimes called “Claudeish.”
He said reinforcement learning rewards different kinds of comprehension: some favor AI-model understanding, while others favor explanations people can follow. Increasing emphasis on mathematics and code therefore requires deliberate rewards for simpler natural-language writing. Kernion said Opus 5.5 achieved a better balance and that he had not been this satisfied with a model’s writing since Opus 4.6, while acknowledging that the problem remains difficult and the newer model does not necessarily surpass its predecessor.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.