Artificial intelligence is generating millions of new hypothetical materials, but many scientists are questioning whether these discoveries are truly groundbreaking or simply theoretical noise. When Google DeepMind announced it had used AI to predict 2.2 million new crystalline materials, it sparked excitement about a new era in materials science. The list included thousands of layered compounds like graphene, potential lithium-ion conductors, and other promising candidates. However, the enthusiasm was quickly tempered by criticism from experts who argue that many of the predicted materials are impractical, unoriginal, or based on flawed assumptions. Materials scientist Anthony Cheetham from the University of California, Santa Barbara, reviewed DeepMind’s list and found over 18,000 compounds containing rare, radioactive elements such as promethium and protactinium—elements so scarce and unstable that they are unlikely to be useful in real-world applications. “It’s one thing to discover a compound, and a totally different thing to discover a new functional material,” he said. Similar concerns arose with Meta’s AI-driven search for metal–organic frameworks (MOFs) capable of capturing carbon dioxide from the air. While the team identified over 100 promising candidates, computational chemist Berend Smit from EPFL argued that the AI model may have overlooked real-world constraints, leading to overly optimistic results. He warned that the excitement around the technology sometimes blinds researchers to practical limitations. The core issue lies in how these AI systems are trained. Most rely on data from density functional theory (DFT), a powerful but computationally expensive method for predicting material stability. DFT tends to favor highly ordered crystal structures—idealized forms that may not exist under normal conditions. In reality, many materials are disordered, with atoms swapping places or arrangements that DFT fails to capture. Johannes Margraf from the University of Bayreuth trained an AI model on experimentally confirmed crystal structures and found that 80–84% of the stable compounds predicted by DeepMind’s GNoME system would likely be disordered in practice. This means many of the “new” materials may not behave as expected—or even be synthesizable in their predicted form. The A-Lab project, which used AI to guide robotic synthesis of new materials, also drew scrutiny. Critics, including Robert Palgrave and Leslie Schoop, argued that many of the materials claimed to be newly synthesized were already known in disordered forms. They pointed out that the AI system was tasked with making ordered structures that don’t exist in nature, but the robots ended up producing disordered versions—something the team still considers a success, though others disagree. Despite these concerns, many researchers believe AI still holds transformative potential. The key is collaboration. Kristin Persson of the Materials Project at Lawrence Berkeley National Lab emphasizes that AI tools like GNoME and MatterGen are not replacements for traditional science but powerful accelerators. They help prioritize promising candidates, reducing the need for exhaustive trial-and-error. Ekin Dogus Cubuk, one of the lead authors of the GNoME paper, now at Periodic Labs, acknowledges that most predicted materials won’t be stable in their idealized forms. But he argues the real value is in using AI to guide further research—not to claim instant breakthroughs. DeepMind notes that more than 700 of the GNoME predictions have since been independently synthesized, and some new caesium-based compounds have been made with AI guidance. While the initial claims of “an order-of-magnitude expansion” in known materials may have been overstated, the underlying technology is proving useful. The takeaway? AI is not yet a magic wand. But when used thoughtfully—with humility, experimental validation, and a clear understanding of its limits—it can dramatically speed up the search for real, functional materials. The dream of AI-driven discovery is alive—but it’s still grounded in the messy, complex reality of chemistry.