---
title: Anthropic employee explains why Claude has unlearned how to write for humans
url: https://www.dataloco.com/en/anthropic-employee-explains-why-claude-has-unlearned-how-to-write-for-humans
published: 2026-09-26T14:23:30+00:00
language: en
section: AI
source: https://the-decoder.de/anthropic-mitarbeiter-erklaert-warum-claude-verlernt-hat-fuer-menschen-zu-schreiben/
organizations: Anthropic
publisher: Dataloco
---

# Anthropic employee explains why Claude has unlearned how to write for humans

Artificial intelligence models are reportedly improving rapidly in mathematics, coding, and logical reasoning. However, in linguistic expression, they are stagnating or even becoming worse.

Jackson Kernion, an employee at Anthropic who works on Claude fine-tuning, explains that Opus 4.6 was the last effective "writing" model from Anthropic. Unsurprisingly, this is due to the optimization of newer models focused on mathematics, coding, and technical explanations for other AI models. This specifically leads to Claude's strange way of expressing himself, termed "Claudeish". The model has been adapted to "LLM psychology" and has learned to write for AI models, not for humans.

Kernion compares this to individuals who communicate exclusively with autistic people: they develop a style that works within that group but is challenging for outsiders to access. LLMs possess significantly more working memory than humans and perceive details on a finer level. Consequently, a writing style emerges during training that is logical and efficient for AI models but sounds too dense and unnatural for human readers.

The issue specifically relates to the reward structure in what is called reinforcement learning, which is a reward-based training method. Some rewards optimize the understanding of AI models, while others optimize for humans. The more one trains on math and coding, the more one must actively counteract this by rewarding simple explanations for human readers.

With Opus 5.5, Anthropic has reportedly found a better balance. This model is the first release with targeted improvements in sentence clarity and overloaded information blocks. Kernion has not been as satisfied with writing quality since Opus 4.6. However, this does not mean that Opus 5.5 surpasses the old model. "It is a difficult problem," Kernion writes. There remains much room for improvements.

## This story in other languages

- [Svenska](https://www.dataloco.com/sv/medarbetare-forklarar-varfor-claude-har-svart-att-skriva-for-manniskor)
- [Nederlands](https://www.dataloco.com/nl/waarom-claude-moeite-heeft-met-schrijven-voor-mensen)
- [Italiano](https://www.dataloco.com/it/i-modelli-di-claude-di-anthropic-hanno-perso-la-capacita-di-scrivere-per-gli-esseri-umani)
- [Español](https://www.dataloco.com/es/empleado-de-anthropic-explica-por-que-claude-ha-dejado-de-escribir-para-humanos)
- [Português (Brasil)](https://www.dataloco.com/pt-br/funcionario-da-anthropic-explica-por-que-claude-desaprendeu-a-escrever-para-humanos)