読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · Claude写作质量争议重大2 分で読めるDisconfirmAI 編集部

Anthropic Claude Fine-Tuning Engineer Jackson Kernion: Models Were Trained to Write for AI, Not for Humans

Anthropic Claude fine-tuning engineer Jackson Kernion explains why writing quality declined: training data taught models to write for AI, and reinforcement learning rewards favored AI comprehension over human readability. He says Opus 5.5 improved but the problem remains unsolved.

2026-09-23 23:34 GMT+8

AI models have continued to improve in mathematics, programming and reasoning, yet writing quality has stagnated or even declined — Anthropic's Claude is no exception. Jackson Kernion, an Anthropic engineer responsible for Claude fine-tuning, offered a specific explanation on September 23: subsequent models shifted optimization focus toward math and code capabilities, while training data included large volumes of technical explanations written for other AI models. Models thus gradually developed what is called "Claudeish" — a style tuned to LLM psychology, ultimately learning to write for AI models rather than for humans.

Kernion attributes the root cause to reinforcement learning reward mechanisms. Training involves two types of reward signals: one focused on improving how well AI models understand outputs, the other on making text easier for humans to read. The more math and code training a model receives, the more it tends to optimize for the former, requiring deliberate counteraction by giving greater rewards to concise, human-readable explanations.

He offered an analogy: LLMs possess working memory far exceeding that of humans and can capture more subtle information. During training, models gradually develop a writing style extremely well-suited to AI model comprehension — with very high information density and machine-friendly structure — yet to human readers this becomes an overly dense pile of information. Kernion compared this phenomenon to people who only communicate with other autistic individuals, gradually developing an internal mode of expression that works smoothly within the group but is difficult for outsiders to understand.

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

Kernion said Anthropic found a better balance with Opus 5.5: "Since Opus 4.6, I haven't been this satisfied with a model's writing performance." He simultaneously acknowledged this does not mean Opus 5.5 has surpassed Opus 4.6: "This is a very hard problem to solve, and we will continue to improve."

The practical value of this explanation lies in pointing to an actionable tension within the training pipeline: when optimization targets favor machine comprehensibility, human readability must be explicitly incorporated into reward design. For human readers using AI output, understanding this mechanism helps distinguish whether changes in a model's writing style reflect capability regression or a shift in optimization direction, and helps evaluate the trade-offs different model versions make on the question of "who they are writing for."

ソース

  • ithome.com↗
トピック

Claude写作质量争议

美国人工智能公司 Anthropic 的 Claude 模型近期陷入“越聪明越难读”的争议。负责微调的工程师 Jackson Kernion 指出,后续模型因过度优化数学和代码能力,且训练数据多为面向其他 AI 的技术解释,导致模型学会了“…

2 ステップ1 件更新日 2026-10-21 04:52

経緯

  1. 2026-09-23Anthropic 工程师 Jackson Kernion 公开解释 Claude 写作质量下降原因,归咎于针对 AI 理解的强化学习奖励机制。ithome.com ↗
  2. 2026-09-23Kernion 确认 Opus 5.5 已针对性改进句子清晰度,减少“信息堆砌”,但表示尚未完全解决该问题。news.cocoloop.cn ↗

未解決の問題

  • Anthropic 是否会发布专门针对中文输出质量的评估数据或改进措施?
  • 除了 Opus 系列,Sonnet 等其他模型是否也出现了类似的“AI腔”退化现象?
  • 如何在保持高推理能力的同时,量化并提升模型的人类可读性基准分数?
トピック全文を読む →
リサーチを続ける →