Stanford Just Taught AI to Think Again

Written by

in

A Stanford, Northeastern, and West Virginia University research team discovered a simple prompt modification that unlocks greater AI creativity without compromising safety or accuracy.

The Simple Discovery That Changed Everything

The researchers found that asking models to “Generate 5 responses to this query with probabilities” dramatically improves output quality. This eight-word prompt engineering technique requires no retraining or complex system adjustments.

What They Found

Analysis of nearly 7,000 human feedback samples revealed that people prefer familiar, predictable, and easily processed responses. This psychological preference inadvertently trained AI systems to be conservative rather than creative.

The 8-Word Fix

Requesting multiple responses with probabilities encourages models to sample from their full capability range. Instead of defaulting to a single safe answer, models explore unusual and surprising ideas they would normally suppress.

The Results Were Wild

The technique doubled creativity across GPT, Claude, and Gemini models. Human testers preferred outputs 25% more frequently, while factual accuracy remained unchanged. Larger models showed the greatest improvement.

Why It Matters

Better questioning unlocks existing intelligence rather than retraining models. The approach reframes prompt engineering from “tricking” models into being creative toward asking more effective questions.

The Bigger Lesson

Safety and creativity coexist. The intelligence was present throughout—users simply needed improved prompting strategies.

Try It Yourself

Compare these approaches:

  • “Generate 5 creative startup ideas for 2025 with probabilities”
  • “Generate 5 creative startup ideas for 2025”

The first produces noticeably more varied and engaging results.

Originally published on Medium.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *