Anthropic engineer reveals why Claude's writing quality declined even as the model became more intelligent

AI models keep getting better at math, coding, and reasoning, yet their prose has stagnated—or in some cases deteriorated. Jackson Kernion, an Anthropic employee working on Claude fine-tuning, recently offered an explanation for why Opus 4.6 remains the company's last truly strong "writing model," and his answer points to a deeper tension in how modern language models are trained.

Why newer models write differently

Part of the answer is the expected one: recent models have been heavily optimized for math and code performance. But Kernion identifies a less obvious culprit. Newer Claude models have also been trained to produce technical explanations intended for other AI models to read—and this, he argues, is what really shapes Claude's distinctive, sometimes awkward phrasing, which users have taken to calling "Claudeish."

According to Kernion, the model has been "adapted to LLM psychology," meaning it learned to write for machines rather than for human readers.

A writing style shaped by machine audiences

To illustrate the problem, Kernion draws an analogy to humans who communicate primarily with other autistic people. Within that group, a communication style develops that works well internally but can feel difficult for outsiders to follow. Similarly, large language models have vastly more working memory than humans and detect patterns at a far finer granularity. During training, a writing style emerges that serves AI readers effectively—but to people, it reads as "overly-dense info dumps."

Reinforcement learning trade-offs

Kernion traces the root cause to the reward structure in reinforcement learning. Some rewards push models toward explanations optimized for AI comprehension, while others target human readability. The more a model is trained on math and code, the more Anthropic must deliberately counterbalance that drift by rewarding simple, human-friendly explanations.

He notes that with Opus 5.5, Anthropic struck a better balance, saying he hasn't been "as happy about a model's writing since Opus 4.6." Still, he stops short of claiming the newer model outperforms its predecessor on prose. "It's a hard problem to solve, and we'll continue to make improvements," Kernion writes.

What to watch next

The tension Kernion describes isn't unique to Anthropic—as labs chase benchmark gains in reasoning and coding, writing quality risks becoming a casualty. Whether future releases can match strong technical performance and natural prose at the same time will be a key measure of progress.

Meta description: Anthropic's Jackson Kernion explains why Claude's writing declined after Opus 4.6, citing training rewards that optimize prose for AI readers over humans. (155 characters)

Tags: Anthropic, Claude, Opus 4.6, Opus 5.5, AI writing quality

Featured image: Abstract technology style—flowing light streams converging into dense data nodes, cool blue and violet tones, no people or logos, no text.