A recent study led by researchers at Google offers new insights into how artificial intelligence (AI) is reshaping scientific research, revealing both significant benefits and persistent challenges. The study surveyed scientists in the United States and the United Kingdom to assess the extent to which AI tools, including large language models (LLMs) like Google’s Gemini, are integrated into researchers’ workflows and how this integration affects productivity and the scientific discovery process.

According to the survey, nearly half of the scientists use AI daily, while 80% engage with it at least weekly. The primary application of AI cited by respondents is the analysis and modeling of research data. However, data analysis accounts for only about 20% of a researcher’s typical workweek. As a result, even substantial productivity gains in this area have a limited impact on overall job performance.

The study also highlights a phenomenon the authors describe as the “verification tax,” where scientists spend considerable time manually auditing and verifying AI-generated outcomes. While researchers reported saving an average of almost seven hours weekly through AI assistance, only a fraction of that time is devoted to initiating new projects, with much of the remainder being used to ensure the accuracy and reliability of AI outputs.

Another significant bottleneck identified is the physical experimentation and data collection phase in the research cycle. More than 40% of scientists indicated an increase in the backlog of untested theories over the past two years, suggesting that while AI accelerates computational and analytical tasks, the pace of real-world experimentation has not kept up. Constraints such as the biological limitations of lab animals and regulatory processes for drug trials continue to delay the translation of AI-driven insights into practical outcomes.

Interviews with early-career scientists corroborate the survey’s findings, shedding light on the nuanced experience of AI’s integration in research. Aleksy Kwiatkowski, a recent Oxford PhD graduate working at an AI science startup, emphasized the distinction between feeling more productive and actual productivity gains, noting the challenge of sifting through vast amounts of information to identify the most valuable insights. Similarly, Wojtek Treyde, also an Oxford computational drug discovery researcher, highlighted the importance of “taste”—the ability to select meaningful scientific problems—as a skill that has become increasingly critical in an AI-augmented research environment.

Experts also point out the necessity of specialized AI models tailored to scientific data. Murray Cox, a cancer immunotherapy researcher at Imperial College London, noted that unlike general-purpose LLMs, bespoke AI tools can develop unique insights by uncovering latent patterns that might elude human researchers. However, creating such models often requires close collaboration between AI engineers and scientific labs, a connection that is not yet widespread.

To address this gap, initiatives like Encode: AI for Science facilitate partnerships between AI experts and academic researchers, providing government-supported platforms for joint work on complex scientific challenges. These programs aim to match AI talent eager to apply their skills beyond commercial applications with scientists seeking advanced computational tools.

While challenges remain, including verification demands and physical experiment bottlenecks, the study and accompanying expert commentary suggest optimism about AI’s potential to accelerate scientific discovery. Scientists interviewed caution against overhyping immediate breakthroughs but express confidence that with thoughtful deployment and interdisciplinary collaboration, AI will play a transformative role in the future of research.