AI models designed to accelerate scientific discovery are running into a fundamental problem: they generate ideas faster than scientists can test them, and many of those ideas turn out to be wrong. A trio of leading researchers from Caltech and MIT are now arguing that the next generation of AI tools must be grounded in actual physics and reality, not just trained on existing data.

The issue became apparent when Google DeepMind published research in 2023 claiming their AI tool had discovered 380,000 new stable materials. When scientists later checked the results, many were not even real materials; some were duplicates, and others simply would not work in practice. As Anima Anandkumar, the Bren Professor of Computing and Mathematical Sciences at Caltech, put it: instead of AI finding the needle in a haystack, the AI just created more hay.

What's the Real Bottleneck in Scientific Discovery?

According to Anandkumar, the limiting factor in science is not a shortage of ideas. Scientists have plenty of hypotheses, both sudden insights and carefully developed theories. The actual bottleneck is testing which ideas work in the real world.

"What is the true bottleneck for science? Scientists have so many ideas; some of them are sudden eureka ideas, some of them are long-thought-out ideas. But the bottleneck is not a lack of ideas. It's going and testing which ideas work in the real world," explained Anandkumar.

Anima Anandkumar, Bren Professor of Computing and Mathematical Sciences at Caltech

For particle physicists like Maria Spiropulu, testing ideas requires running extremely expensive particle colliders. For ecologists like Sara Beery, it means traveling to remote locations to collect field observations. Current large language models (LLMs), which are AI systems trained on vast amounts of text data, cannot perform these real-world tests. Instead, they generate more ideas without being able to verify whether those ideas are actually correct.

How Should AI Be Redesigned to Support Real Science?

The solution, according to these researchers, is to build AI systems that understand the underlying physics of the world. A physics-grounded AI model would be able to optimize designs and predict outcomes without requiring constant real-world validation for every hypothesis.

• Physics-Based Understanding: AI models need to be built with knowledge of fundamental physical laws, such as fluid dynamics and aerodynamics, so they can reason about the world rather than just pattern-match from training data.
• Hypothesis Testing Over Hypothesis Generation: Rather than asking AI to come up with new scientific ideas, researchers should focus on using AI to efficiently test hypotheses that scientists already have, especially when dealing with massive datasets.
• Domain-Specific Grounding: Different scientific fields require different levels of mathematical understanding; high-energy physics has 70 years of validated theory to build on, while ecology lacks complete mathematical models of complex ecosystems.

Sara Beery, a Caltech alumna now at MIT, noted that the role of AI in science will evolve significantly over the next decade. She explained that when she started using AI ten years ago, it was primarily for automating labeling tasks, such as recognizing species or counting trees in images. Within another ten years, she predicts AI will advance to help scientists test hypotheses and optimize experimental designs, with human scientists providing ethical oversight, context, creativity, and verification.

"I think in another 10 years, AI will have gotten to the point where it can now help us test hypotheses and optimize experimental designs, with scientists providing ethical oversight, context, creativity, and verification," stated Beery.

Sara Beery, Homer A. Burnell Career Development Professor at MIT

Why Does Ground Truth Matter in Different Fields?

The challenge of grounding AI in reality varies dramatically across scientific disciplines. In high-energy particle physics, researchers have seven decades of validated theoretical predictions and experimental data to work with. This "ground truth" allows physicists to test whether their AI models can reproduce known physics before using them to explore new territory.

Ecology presents a much harder problem. Scientists in this field rely almost entirely on direct observation and real-world measurements, with no complete mathematical model of how ecosystems function. The scale is enormous, and the systems are chaotic and complex. As Spiropulu noted, ecology lacks the kind of ground truth that physics possesses, making it far more difficult to validate AI predictions.

Despite these challenges, physicists still face fundamental gaps in their understanding. The standard model of particle physics cannot explain gravity or why the universe contains matter and antimatter in unequal amounts, even though the Big Bang should have produced equal amounts of both. This is where AI may help push the boundaries of human knowledge.

What About Quantum-Level AI?

Maria Spiropulu is exploring an even more ambitious direction: building AI systems that operate at the quantum level. She references physicist John Archibald Wheeler's 1989 concept of "it from bit," which suggests that everything physical in the universe has a basis as information. Since nature is fundamentally quantum, Spiropulu argues that the most powerful AI systems will need to harness quantum information itself.

This would require better quantum computers that can accurately simulate nature's quantum reality. Progress is being made on this front, with recent advances in quantum computing hardware demonstrating new capabilities. However, this remains a frontier area of research, and practical quantum-powered AI is still years away.

The core insight from these researchers is clear: the next generation of AI tools for science must move beyond simply generating more ideas. They need to understand the physical world deeply enough to validate their own outputs, test hypotheses efficiently, and ultimately accelerate the pace of genuine scientific discovery.