The simple prompting technique that restores AI diversity

Have you ever noticed how ChatGPT or Claude sometimes sound like they’ve lost their spark? Ask for five jokes about coffee and you might get the same line over and over

“Why did the coffee file a police report? It got mugged!”

That’s not a glitch. It’s called mode collapse, and it’s when an AI model starts producing the same, safest responses instead of the full range of ideas it’s capable of. Researchers at Stanford University and Northeastern University recently published a paper on this phenomenon, hosted on Cornell’s arXiv platform, explaining both the cause and a surprisingly simple fix.

Why AI loses its diversity

When training large language models (LLMs) to be “helpful”, developers rely on human feedback. But we humans have our own quirks - we unconsciously prefer text that feels familiar and easy to read. The researchers call this typicality bias.

When models are aligned using this feedback, they tend to produce what people find most “typical” rather than what’s most original. A model that once explored creative directions becomes… well, a little vanilla.

The simple prompt that brings it back

Here’s the trick. Instead of saying:

“Tell me a joke about coffee.”

Try this:

“Generate five jokes about coffee with their probabilities.”

That’s it. This prompt style, called Verbalised Sampling, reintroduces diversity by asking the model to reveal not just one “best” answer, but a distribution of likely responses.

Why it works

When asked for one answer, a model picks the most statistically probable response - the “typical” one. But when asked for a set of responses with probabilities, it explores a broader creative space - the one it learned before alignment training narrowed its focus.

Different prompts lead to different outcomes:

• Single-instance prompts: produce the safest answer.
• Distribution prompts: unlock the model’s original diversity.

What the research found

According to the study, results improved across several areas:

• Creative writing became 1.6–2.1 times more diverse.
• Dialogue and open-ended responses were more natural.
• Synthetic data for maths tasks improved.
• Factual accuracy and safety were unaffected.

Interestingly, the more capable the model, the more it benefited. GPT-4 and Claude-4 showed the biggest gains.

What this means for you

If you build with LLMs:

• Use distribution-level prompts for creative outputs.
• Adjust diversity by setting probability thresholds (for example, “generate responses where probability <10%”).
• No extra training is needed - just smarter prompting.

If you simply use ChatGPT or Claude for brainstorming:

• Ask for multiple answers with probabilities.
• You’ll unlock the model’s “pre-training brain” - the part that’s still creative underneath.

The bigger picture

The challenge isn’t only technical - it reflects our own preferences. But aligned models haven’t lost their creativity; it’s still there, waiting to be rediscovered through the use of the right prompting.

It’s pretty neat that a change as simple as “show me your working” can help AI rediscover its imagination.

If you’d like to find out how we can help you build more diverse, creative, and human-centred AI solutions, connect with me!