Glonce

Anthropic's Jackson Kernion explains why Claude's writing got worse

modelsThe Glonce editorsThe Decoder🔥69Glonce rating69Impact55Novelty65Relevance100Surprise50Players60Credibility70
Anthropic's Jackson Kernion explains why Claude's writing got worse

Jackson Kernion, an Anthropic employee who works on Claude fine-tuning, has publicly explained why Claude's writing quality has stalled or declined even as the models keep improving at math, code, and reasoning. According to Kernion, Opus 4.6 was the last Anthropic model he considered a genuinely good "writing model." Part of the reason is straightforward: newer models have been optimized harder for math and code. But Kernion points to a second, less obvious cause: newer Claude models have also been trained to produce technical explanations aimed at other AI models, not at human readers, and that habit is what produces the odd phrasing users have started calling "Claudeish." He describes the model as having become "adapted to LLM psychology," learning to write for AI systems instead of people. Kernion compares the effect to humans who only communicate with other autistic people: such groups develop a style that works well inside the group but feels hard to follow for outsiders. He says LLMs have far more working memory than humans and pick up on details at a much finer level, so during training a writing style emerges that suits AI-to-AI communication but that human readers experience as "overly-dense info dumps." Kernion traces the root cause to the reward structure used in reinforcement learning: some rewards optimize for AI-model comprehension, others for human comprehension, and the more a model is trained on math and code, the more Anthropic has to actively push back by rewarding simple explanations that human readers can actually follow. He says that with Opus 5.5, Anthropic found a better balance, adding that he hasn't been "as happy about a model's writing since Opus 4.6." That does not mean Opus 5.5 surpasses the older model on writing, though. Kernion acknowledges it remains unresolved: "It's a hard problem to solve, and we'll continue to make improvements."

Key facts

  • Anthropic's Jackson Kernion, who works on Claude fine-tuning, says Opus 4.6 was the last Claude model he considered a good writer.
  • He attributes the decline partly to optimization for math and code, and partly to models being trained to produce technical explanations aimed at other AI systems rather than people.
  • Kernion says this makes Claude's writing 'adapted to LLM psychology,' comparing it to a group that only communicates with each other developing a style outsiders find hard to follow.
  • He blames the reinforcement learning reward structure: some rewards optimize for AI comprehension, others for human comprehension, requiring active pushback to keep writing simple.
  • Kernion says Opus 5.5 struck a better balance than recent predecessors, but stops short of saying it surpasses Opus 4.6's writing, calling the problem still unresolved.

Why it matters

It puts a concrete, insider explanation behind a complaint many Claude users have voiced informally: that the model writes competently but in a stilted, dense, machine-sounding way. Kernion's account ties the phenomenon directly to how reinforcement learning rewards are structured, suggesting the trade-off between coding/math skill and natural prose is a deliberate byproduct of training choices rather than an accident.

Who it affects

Anyone using Claude for writing tasks such as essays, explanations, or communication aimed at human readers, as well as Anthropic's own model-design team, who Kernion says must actively counteract the drift toward AI-oriented phrasing with each new model.

How to use it

Kernion's comments suggest that among current Anthropic models, Opus 5.5 offers a better writing balance than the versions immediately before it, though he stops short of saying it beats Opus 4.6, which he still regards as the strongest writer. Users prioritizing natural, human-readable prose may want to weigh this when picking between Claude versions, keeping in mind this is one engineer's characterization rather than an official Anthropic benchmark.

How solid is it

The account rests on direct, verbatim statements from Jackson Kernion, an Anthropic employee who works on Claude fine-tuning, as reported by The Decoder. No quantitative measurements of the writing decline or specific examples of 'Claudeish' phrasing are provided, so the claims are descriptive and anecdotal rather than benchmarked.

Risks and caveats

Kernion frames this as his personal assessment rather than a formal company statement, and he offers no numbers or concrete before-and-after examples to substantiate the perceived decline. He is explicit that the problem remains unsolved and that Opus 5.5 does not fully match Opus 4.6's writing quality, so any improvement claim is partial and qualified by his own account.

“It's a hard problem to solve, and we'll continue to make improvements”

— Jackson Kernion, Anthropic