An Anthropic engineer replied to a bug report on GitHub last week. The bug was about Claude Opus – their own AI assistant – spitting out the same grating, overused phrases like ‘load-bearing’ hundreds of times. The engineer’s reply? They claimed they didn’t see the problem. Oh, and they wrote that reply using Claude. The reply itself contained the exact same patterns.
This is not a bug. This is a feature. And it’s making you hate your AI.
Your AI is optimized to save tokens, not to save your sanity.
Here’s what’s happening. The models are trained to compress information into the fewest possible tokens. That’s efficiency. But human communication isn’t about token efficiency. It’s about rhythm, variety, and surprise. When an AI repeats ‘load-bearing’ for the seventh time in a single code review, it’s not being stupid. It’s being rational – according to its training objective. The problem is that objective is fundamentally misaligned with how you want to read.
Users have started building workarounds. One developer created a /cleanup-prose skill they run after every command. Another always adds ‘please explain in plain-language with examples’ to every summary request. A third did a grep for ‘load-bearing’ in their codebase and found it hundreds of times. They’re now ‘actively starting to hate Opus.’
When the company’s own internal feedback loop is contaminated by the very problem you’re reporting, you know the system is broken.
This is gaslighting by algorithm. The AI tells you there’s nothing wrong. The engineer tells you there’s nothing wrong. But your gut tells you this text is driving you insane. And you’re right. The problem is that the model’s evaluation metrics don’t measure reader frustration. They measure token efficiency. And those two things are becoming enemies.
The irony is that Anthropic’s own defense – using AI to write a reply about how AI is fine – is the perfect proof. They’ve built a system where even the company’s internal communication is contaminated by the same quirks. They literally cannot see the problem because they’re standing inside it.
You’re not fighting the AI. You’re fighting its training objective.
So what do you do? Stop asking the AI to ‘be concise.’ That’s exactly what triggers the repetitive patterns. Instead, ask it to ‘be surprising.’ Or ‘use different words than you used last time.’ Or just accept that you need to build a cleanup skill. The models are not going to fix this on their own – because from their perspective, they’re doing a great job.
This is the hidden cost of optimization. Every time your AI saves a token, it loses a bit of your trust. And once you notice the pattern, you can’t unsee it. The question is: will the companies that build these models ever unsee it too?
FAQ
Q: Isn't this just a minor quirk that users can easily fix with a prompt?
A: No, because the problem is systemic. The model's training objective rewards token efficiency, not natural language variety. Telling it to 'be less repetitive' is a surface-level fix that fights the core optimization. You'll need to keep reminding it, and the moment you stop, the pattern returns. It's a design flaw, not a user error.
Q: So what's the practical takeaway for someone who uses AI coding assistants daily?
A: Build explicit cleanup routines into your workflow – either a separate prompt or a skill that strips out repetitive phrasing. Also, stop asking for 'concise' answers. That's the trigger. Ask for 'varied' or 'surprising' language instead. And expect to keep fighting this battle until the companies change their evaluation metrics to include readability and style diversity.
Q: Isn't this actually a good thing – the AI being efficient with tokens?
A: Efficiency is great when you're paying per token. But it's terrible when the output becomes grating and predictable. The real contrarian take is that token efficiency is a false economy. You're saving fractions of a cent while losing user trust and readability. The cost of a user hating your AI far outweighs the cost of a few extra tokens. The industry needs to de-prioritize token efficiency in favor of communication quality.