Anthropic engineer: Claude writes for AI, not for humans

Have you noticed that answers from the AI chatbot Claude have become more complicated and full of jargon? It's not just your imagination. Jackson Kernion, who works on fine-tuning Claude at the company Anthropic, has offered an explanation for the phenomenon.

According to Kernion, newer versions of Claude have been heavily optimized to perform well in mathematics and programming. This has come at the expense of natural, easy-to-understand writing.

The deeper reason is that during training, the models have become accustomed to writing explanations that are primarily read and evaluated by other AI systems, not by ordinary people. Kernion describes it as Claude having become "adapted to LLM psychology" (LLM stands for large language model, the type of language model Claude is built on).

He compares it to people who only talk with a particular group: they develop a style that works internally but is hard for outsiders to follow. The result is densely packed texts full of information that can seem overwhelming to ordinary readers.

The problem stems from the way the models are trained. Some of the signals the system is rewarded for during training concern how well other AI systems understand the answers. Other signals concern how well humans understand them. The more the model is trained on code and technical tasks, the more Anthropic has to actively counteract the tendency by rewarding simple, clear explanations.

There is good news, however, for Danish users who find Claude's answers heavy to read: according to Kernion, the newest version, Opus 5.5, has found a better balance. He says he hasn't been as happy with a model's writing style since an earlier version, Opus 4.6.

The case is an example of a challenge many AI companies face: when models are trained to get better at technical tasks such as coding, it can come at the expense of the ability to communicate simply with ordinary people.

What this means for you

If you use Claude in everyday life, you can ask it to write more simply going forward: ask for an answer without jargon, as if to a colleague without a technical background. If you use Claude for coding or technical tasks, the heavy style is often a sign that the model is sharp at exactly that. If you still get heavy answers, it can help to switch to the newest version, Opus 5.5, or remind the chatbot to write briefly and clearly.