AI systems are unlikely to make the scientific discoveries some leading labs are hoping for, Hugging Face’s top scientist says

Speaking to Fortune at Viva Technology in Paris, the Hugging Face cofounder said that while large language models (LLMs) have shown an impressive ability to find answers to questions, they fall short when trying to ask the right ones—something Wolf sees as the more complex part of true scientific progress.

“In science, asking the question is the hard part, it’s not finding the answer,” Wolf said. “Once the question is asked, often the answer is quite obvious, but the tough part is really asking the question, and models are very bad at asking great questions.”

Wolf said he came to this conclusion after reading a widely circulated blog post by Anthropic CEO Dario Amodei titled “Machines of Loving Grace.” In it, Amodei argues the world is about to see the 21st century “compressed” into a few years as AI accelerates science drastically.

Wolf said he initially found the piece inspiring but started to doubt Amodei’s idealistic vision of the future after the second read.

“It was saying AI is going to solve cancer, and it’s going to solve mental health problems—it’s going to even bring peace into the world. But then I read it again and realized there’s something that sounds very wrong about it, and I don’t believe that,” he said.

For Wolf, the problem isn’t that AI lacks knowledge but that it lacks the ability to challenge our existing frame of knowledge. AI models are trained to predict likely continuations—for example, the next word in a sentence—and while today’s models excel at mimicking human reasoning, they fall short of any real original thinking.

“Models are just trying to predict the most likely thing,” Wolf explained. “But in almost all big cases of discovery or art, it’s not really the most likely art piece you want to see, but it’s the most interesting one.”

Using the example of Go, a board game that became a milestone in AI history when DeepMind’s AlphaGo defeated world champions in 2016, Wolf argued that while mastering the rules of Go is impressive, the bigger challenge lies in inventing such a complex game in the first place. In science, he said, the equivalent of inventing the game is asking these truly original questions.

Wolf first suggested this idea in a blog post titled “The Einstein AI Model,” published earlier this year. In it, he wrote: “To create an Einstein in a data center, we don’t just need a system that knows all the answers, but rather one that can ask questions nobody else has thought of or dared to ask.”

He argues that what we have instead are models that behave like “yes-men on servers”—endlessly agreeable, but unlikely to challenge assumptions or rethink foundational ideas.