Mode Collapse

Mode collapse is when an AI keeps producing the same handful of phrases or patterns, even in situations where many different ways of expressing something are possible.

If you ask 100 people the same question, you get 100 different answers, each with its own tone and wording. But if you ask a chatbot similar questions repeatedly, you may notice familiar phrases and sentence structures coming up again and again. Even though the model is theoretically capable of writing as diverse a range of text as a human, in practice it narrows its answers down to just a few fixed patterns — that's mode collapse.

This isn't a sign that the model lacks ability. It happens because of the process, before release, of training the model on behavioral rules like conversational etiquette or refusing dangerous requests. A raw model without these rules applied can write in as wide a variety of styles as a person would. But once it's trained on these rules, probability starts piling up around certain apology phrases or formulaic transition patterns. Ironically, this concentration itself can become a clue for spotting AI-written text.

Articles point to 'mode collapse' as the reason ChatGPT's or Claude's writing keeps coming out in a similar tone. It's described as "a phenomenon where a model converges on one particular phrase instead of the range of expressions a human would use." What's easy to misunderstand is that this isn't caused by weaker model performance. The key point in such articles is that it's actually a side effect of the expressive range narrowing during training on safety measures like conversational etiquette or refusing risky statements.

1. Ask a chatbot the same question three or four times with slightly different wording. For example: "How's the weather today?", "Tell me today's temperature", "What does it look like outside?"
2. Check the answers for repeated phrases or sentence structures. See if it always starts with a similar opening line or apology expression.
3. Now try asking, "Rewrite the following sentence in 5 different styles." Compare how genuinely different the 5 versions are, or whether they're just similar sentences with slightly swapped vocabulary.
4. Comparing this to how differently multiple people would answer the same question helps you feel just how narrow the AI's range of expression really is.