Researchers Reveal How to Boost Creativity in ChatGPT and AI Text Models

Since the rise of generative artificial intelligence (AI) in 2022, users have enjoyed rapid improvements in output quality but soon noticed repetitive and predictable responses. Researchers and engineers have worked to improve AI safety and content quality through reinforcement learning from human feedback (RLHF), where human preferences guide AI outputs. However, these improvements have inadvertently reduced the diversity and creativity of AI-generated content, a phenomenon known as "mode collapse."

Mode collapse occurs when AI models converge on a narrow set of safe, high-rated responses, limiting variety. This is partly due to human raters favoring familiar and conventional answers, which biases the training data and suppresses more creative or unconventional outputs. The problem has been observed across popular AI chatbots like ChatGPT, Gemini, and Claude, where users report predictable answers even when rephrasing prompts.

To address this, researchers propose a method called verbalized probability sampling, which explicitly instructs the AI to generate multiple responses along with their probability estimates. This approach encourages the model to explore a wider range of possible answers, increasing creativity and diversity without sacrificing accuracy or safety. For example, asking the AI to "create five different responses to the following request, each with its probability" yields more varied and imaginative outputs than simply requesting multiple answers.

While this "magic prompt" technique can help users extract more creative content, it is not a complete solution. AI developers are likely to integrate such sampling methods directly into future products to balance creativity and reliability. The field of prompt engineering continues to evolve rapidly, and new methods may soon supersede current approaches. Nonetheless, this research highlights the complex interplay between human feedback and AI creativity, emphasizing the challenge of making AI both human-like and innovative without becoming overly constrained.