Why AI writing all sounds the same.
Chat models learn from millions of people's writing, then get trained to play it safe. They end up sharing one voice. There's a name for it: mode collapse.
- People
- ChatGPT
- Claude
- Gemini
- Emulate
One voice for everyone
Ask three chatbots for a cover letter and you get three versions of the same letter: the same opening line, the same tidy rhythm, the same upbeat sign-off. Readers can hear it within a sentence. So can AI checkers.
What mode collapse is
A model learns to write by reading a huge amount of human writing, so at first it can sound like almost anyone. Then it's trained on answers people rated, and people rate the familiar answer higher. Round after round, it learns to pick the most expected word every time, and its writing piles up in one corner. That pile-up is mode collapse.
Why prompts don't fix it
Telling a model to “sound human” changes its words, not its habits. It swaps in a few casual phrases and keeps the same shape underneath. That's why tools that only swap synonyms still get flagged.
What we do instead
We train our own models on how real people write, across the whole range: plain, odd, specific, a little messy. So Emulate doesn't reach for the safest word. It reaches for the one you would.
[Further reading]
- Mysteries of mode collapse
janus, LessWrong, 2022. Where the term took hold for chat models.
- Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
Zhang et al., arXiv, 2025. Why people's ratings pull models toward familiar text.
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
Jiang et al., arXiv, 2025. Different models, strikingly similar answers.