Why AI is speeding up scientific research but not lab experiments

A new Google, Google DeepMind, and MIT report reveals that while artificial intelligence is generating many new ideas for scientists, it has yet to significantly speed up laboratory experiments. Researchers surveyed 637 scientists and analyzed over 2,600 specialized models, finding that 44% of respondents reported shifting their main research bottleneck to later stages such as experimentation and data collection, with 41% experiencing a growing backlog of untested hypotheses.

However, only 89% of those who save time using AI spend more than a tenth of that time checking AI outputs, and 46% spend more than a quarter. This indicates that the time spent verifying AI results significantly impacts the usefulness of AI in different fields. While AI is effective for mathematical proofs, which can be checked against formal rules, its application to protein functions, which require lab testing and validation in living systems, faces significant challenges.

James Zou, a biomedical data scientist at Stanford, has developed methods for integrating AI with both large language models (LLMs) and specialized tools like AlphaFold, but extending this automation to real-world experiments poses substantial hurdles. Automated labs are limited in their ability to handle complex experiments such as those involving animals, and the cost of establishing and maintaining such facilities is prohibitive.

Moreover, the type of questions tackled by AI often relies on established ground truth, which is scarce in many medical fields. Clinical trials also move at a slow pace dictated by regulatory reviews and safety checks, making it difficult for AI to accelerate experimental procedures involving human subjects. Despite these challenges, researchers are still striving to overcome these obstacles and explore the potential of AI in drug discovery and experimental biological research.