Anthropic 工程师解释 Claude 写作能力退化：模型被训练成“写给 AI 看”

Claude's Writing Ability Has Deteriorated! Engineers Explain: The Model Was Trained to Write for AI Rather Than Humans AIbase基地 Published inAI News · 4 min read · Sep 24, 2026 9 Many people have noticed that AI models are making rapid progress in math, programming, and reasoning, but their writing skills have stagnated or even declined, and Claude from Anthropic is no exception. Why was Opus 4.6 the last model from Anthropic that performed well in writing? Kneen, an engineer at Anthropic who worked on fine-tuning Claude, gave an explanation yesterday. Why did writing performance decline after Opus 4.6? Kneen said that the focus of subsequent model optimizations has shifted more towards math and code capabilities. At the same time, these models were trained on a large amount of "technical explanations written for other AI models" — which is the main reason for Claude's unique expression style, known as "Claudeish." He explained that the model was adjusted to adapt to the "psychology" of LLMs, eventually learning to write for AI models rather than humans. Kneen compared this phenomenon to "a person who only communicates with others with autism": this group gradually develops a way of expression that works well internally but is hard for outsiders to understand. Because LLMs have far superior working memory and can capture more subtle information, the training process gradually formed a writing style that is very suitable for AI understanding; however, to human readers, this expression becomes an overload of dense information. Reward system favors AI understanding, Opus 5.5 regains balance Kneen believes that the root of the problem lies in the reinforcement learning reward mechanism: some rewards emphasize improving understanding between AI models, while others focus on making it easier for humans to understand. The more training there is in math and code, the more active the need to counteract the tendency to "write for AI," and more rewards should be given for concise and easy-to-read explanations. He revealed that Anthropic found a better balance in Opus 5.5, "Since Opus 4.6, I haven't been so satisfied with a model's writing performance before." However, he also admitted that this does not mean Opus 5.5 has surpassed Opus 4.6, "This is a difficult problem, and we will continue to improve." This article is from AIbase Daily Scan to view Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications. —— Created by the AIbase Daily Team © Copyright AIbase Base 2024, Click to View Source -