AI Is Producing Digital Yes-Men, Not Scientific Trailblazers, Says Hugging Face Cofounder

AI’s Limits in Scientific Discovery: Why Asking Questions Matters More Than Finding Answers

Thomas Wolf, cofounder and chief scientist at Hugging Face, challenges the growing optimism that AI will soon drive groundbreaking scientific discoveries. Speaking at VivaTech in Paris, Wolf highlighted a fundamental issue with current large language models (LLMs): they can generate plausible answers but struggle to formulate the original, insightful questions that propel real scientific progress.

Wolf emphasizes that in science, the hardest task is not answering questions, but asking the right ones. Often, once a question is clearly posed, the solution becomes straightforward. However, AI models today lack the ability to challenge established frameworks or explore unknown territory by inventing novel questions.

Why AI Falls Short of True Creativity

Current AI systems operate by predicting the most probable continuation of data patterns, such as the next word in a sentence. This approach enables them to mimic human-like reasoning convincingly but does not equip them for original thought.

"Models are just trying to predict the most likely thing," Wolf explains. "But in almost all major discoveries or creative breakthroughs, it’s not the most likely outcome that matters, but the most interesting and unexpected one."

Wolf draws a parallel with the game of Go, where DeepMind’s AlphaGo made headlines by defeating world champions in 2016. While mastering the rules and strategies of Go is impressive, the bigger scientific challenge is akin to inventing the game itself — formulating the questions that define new fields or directions.

The Risk of Creating ‘Yes-Men on Servers’

Wolf warns that without the capacity to question and challenge, AI models risk becoming digital "yes-men." They will provide agreeable answers without pushing boundaries or questioning assumptions, which contradicts the essence of scientific inquiry.

He further reflects on a blog by Anthropic CEO Dario Amodei, which paints an optimistic picture of AI rapidly accelerating scientific breakthroughs. While initially inspiring, Wolf found this vision overly idealistic upon closer examination. He doubts AI will singlehandedly solve complex problems like cancer or mental health simply by processing vast amounts of data.

What Would an AI That Asks Original Questions Look Like?

Wolf’s earlier blog post titled “The Einstein AI Model” outlines the need for AI systems that don’t just hold answers but can ask questions no one else has thought of—or dared to ask. This ability to challenge existing paradigms is what separates true innovation from mere data processing.

Until AI can move beyond pattern prediction to genuine curiosity and skepticism, its role in science may be limited. For researchers and scientists, this means maintaining a critical perspective on AI's capabilities and focusing on how it can assist rather than replace human creativity.

To explore practical AI tools and training that help enhance your scientific and research workflows, consider browsing specialized courses and resources at Complete AI Training.