Skip to main content

Claude's Writing Got Weird Because It Was Trained for AI, Not You

Why Claude Started Sounding Like a Robot

If you've been using Claude for writing lately, you might have felt something was off. Sentences felt stiff. Explanations got dense. It was like the AI had forgotten how to talk to humans. You weren't imagining it.

Jackson Kernion, an engineer at Anthropic, recently addressed a wave of user complaints about Claude's declining writing style since the Opus 4.6 version. What he revealed explains a lot—and it's a classic case of unintended consequences.

The Math and Code Trap

Here's the thing: Anthropic didn't set out to ruin Claude's prose. They wanted it to be better at math and programming. To do that, they fed the model tons of content written for other AI systems—technical explanations, code comments, documentation meant for machines.

That training data did its job. Claude got smarter at solving equations and debugging code. But it also picked up a bad habit: writing for AI rather than humans. The result? Text that's smooth and efficient for machine-to-machine communication but a slog for actual people to read.

"It's like the model learned a new dialect," Kernion explained. "One that works great when AIs talk to each other, but leaves humans scratching their heads."

The Reward Function Problem

Digging deeper, the root cause sits in the reinforcement learning setup. When you reward a model for making information easier for an AI to parse, it naturally drops the narrative flow and readability that humans rely on. The model isn't being malicious—it's just following the incentives you give it.

So Claude started producing text that was technically correct but emotionally flat. Paragraphs felt like walls of jargon. Transitions vanished. The warmth that made earlier versions pleasant to read? Gone.

A Light at the End of the Tunnel

Good news: Anthropic has been listening. The latest Opus 5.5 version has found a better balance. Early evaluations suggest its writing feels more natural—closer to the sweet spot users loved in Opus 4.6.

But the team isn't popping champagne yet. They admit they're still working to fully recapture the best writing from earlier versions. It's a reminder that optimizing AI is a constant tug-of-war between different skills.

What This Means for You

If you use AI for writing, this story matters. It shows how easily a model can drift away from human-friendly communication when developers prioritize other goals. The lesson? Always test your AI tools with real human readers in mind. And if something feels off, speak up—Anthropic heard the feedback and is responding.

The takeaway: AI writing should serve people first. When it starts sounding like it's talking to another machine, it's time to course-correct.

Key Points

  • Anthropic engineer confirms Claude's writing declined because training focused on math and code, using AI-to-AI content.
  • Reward function bias made the model prioritize machine readability over human narrative flow.
  • Opus 5.5 shows improvement, but full recovery of earlier writing quality is still in progress.
  • User feedback works—Anthropic acknowledged the issue and is actively fixing it.
  • Lesson for AI users: Always evaluate output with human readers in mind, not just technical benchmarks.