The evolution of large AI models often comes with unexpected side effects. In response to the recent frequent feedback from users about the decline in writing style of Claude since the Opus4.6 version, and the so-called "Claude accent" phenomenon, Jackson Kernion, an engineer at Anthropic, recently publicly explained the issue, revealing the technical root behind this peculiar phenomenon.

Regarding the overall direction of model optimization, the development focus of subsequent versions of Claude has clearly shifted towards mathematical reasoning and programming capabilities. To improve the model's performance in complex engineering tasks, the research team introduced a large amount of content in the training process that was written for other AI models, such as technical explanations and code comments. This training method caused the model to gradually learn to "write for AI rather than humans," resulting in overly dense expressions that are very smooth for communication between AIs but extremely difficult for human readers to understand.

From the perspective of the underlying reinforcement learning mechanism, the core of this issue lies in the bias of the reward function. When the optimization goal focuses on improving the AI's understanding and interpretation of information, the model naturally abandons the narrative rhythm and readability that humans are accustomed to. Currently, the Anthropic team has found a better balance in the latest Opus5.5 version, and its writing performance has been evaluated as the most satisfactory state since Opus4.6. However, the relevant team stated that efforts to completely surpass the best level of earlier versions are still ongoing.