AI is giving scientists more ideas than they can test (www.scientificamerican.com)

Recent developments in AI have sparked a debate in the scientific community about its effectiveness in accelerating research, particularly in fields like drug discovery. A new report by Google, Google DeepMind, and MIT quantifies the challenges scientists face, revealing that 44% of surveyed researchers now find their main bottlenecks in later stages, such as experimentation and data collection, rather than hypothesis generation. This shift has led to a growing backlog of untested hypotheses, with more than 40% of scientists reporting an increase in unmet research opportunities. Despite significant advancements in AI, many researchers still find themselves spending substantial time verifying AI outputs, limiting the immediate utility of these technologies in applied research. The report highlights key barriers to AI’s broader application in scientific discovery, including the complexities of real-world experimentation and regulatory constraints, especially in fields like biomedicine. Experts point out that while large language models and automated labs show promise, they struggle with tasks that involve complex, living systems due to safety and practical limitations. Ongoing efforts, like those from Northwestern University's DREAM Lab, aim to enhance automation in protein engineering, but the path forward remains gradual. As researchers continue to navigate these challenges, the potential for AI to transform science hangs in the balance, emphasizing the need for an integrated approach to both AI capabilities and experimental methodologies.