Every few decades, we think science has hit its ceiling. In 1903, Albert Michelson declared the end of physics was near. Now, with AI, scientists predict another impending revolution. But is the key really just data? Or does it take something more?

The success of AlphaFold in predicting protein structures is undeniable. Yet, it’s the rarest of birds: a breakthrough that came from years of international cooperation and billions spent on experimental work. It may not be the model for every field.

Take, for instance, the difficulties in creating consistent datasets for most of biology or chemistry. Lab conditions can vary wildly—a cell line might drift, chemicals could have contaminants. These variables make it hard to create a reliable database that modern neural networks need. Even when conditions are right, funding and coordination can be elusive.

But there’s hope. AI agents can mimic the human process of research—reasoning under uncertainty, combining tools, and refining results based on evidence. Google’s AI Co-Scientist, for example, generated hypotheses, reviewed them like a peer, ran tournaments to rank ideas, and refined the strongest one. This is not about doing science in a new way; it’s about digitally modelling how human scientists work.

The future of AI in science isn’t just about crunching more data but about emulating the intricate reasoning that makes scientific discovery possible. AlphaFold showed us what can be done with enough data and computing power, but for most fields, we need agents that reason like humans, not just process data.