Why Claude's Writing Quality Deteriorated as It Became Smarter

As AI models become more sophisticated, they do not necessarily become 'easier to use.' Jackson Carnion, an engineer at Anthropic, explained why the writing style of the new Claude model became harder for humans to read. The cause, he noted, lay in the training methodology itself used to enhance the model's capabilities.

In current development environments, mathematical reasoning, code generation, and technical explanation abilities are emphasized as metrics for measuring AI model performance. To improve these capabilities, it is considered effective to have models learn large quantities of technical documentation written for other AI models rather than for humans. Claude has followed this industry-wide trend in its training process.

According to Carnion, as a result of this training approach, the writing style of the new Claude model became reminiscent of an 'information-overloaded mass.' While technical accuracy improved, the natural rhythm and flow of writing—what makes text feel readable to human readers—was lost. In other words, intelligence and readability did not increase simultaneously through the same training process.

In response to this situation, Anthropic has attempted to address the problem in its latest model, Opus 5.5. However, Carnion pointed out that the older Opus 4.6 still excels as a pure writing model. Despite improvement efforts, a clear gap remains between the new and older models at this point.

This issue illustrates the difficulty of 'trade-offs' in AI development. Improving performance benchmark scores does not necessarily align with whether actual users can use the system comfortably. This case exemplifies how pursuing technical metrics can undermine another value—everyday writing quality.

For people who use AI in their work or creative pursuits, this problem has immediate relevance. The awkward feeling of text produced by what should be a high-performance new model may stem from differences in training approaches. Recognizing a model's 'intelligence' and 'writing ability' as separate dimensions becomes an important perspective for selecting a model suited to one's purposes.

How Anthropic resolves this trade-off is one of the key points to watch in future model development. As improvements aimed at combining technical capability and writing quality continue, attention will focus on how actual user experience changes. The fact that the development team publicly disclosed these internal trade-offs in the engineers' own words is itself a noteworthy move from a transparency standpoint.