Verification effort: a critical limiting factor for the efficiency gains of using generative AI in research

The use of generative artificial intelligence (genAI) tools in science has become ubiquitous. Despite considerable debate about when and how such tools should (or should not) be used, there is widespread agreement on the need for verifying the AI’s outputs and assuming responsibility for them. Yet, most discussions about applying genAI tools to common scientific tasks center around these tools’ capabilities, with little attention given the effort required for verifying their outputs. Such verification effort varies significantly between tasks – while in some cases it may require considerably less effort to verify the AI’s output than to perform the task from scratch (e.g., when coding), in other cases properly verifying the output essentially requires performing the task oneself (e.g., when peer reviewing a paper or synthesizing published evidence). Thus, I argue that unless we are willing to trust the output of genAI tools blindly, some of their most touted applications may deliver very little gains in efficiency, remaining impractical even as the models improve, simply because the verification effort is close to the effort of performing the task from scratch. Importantly, such costs may only increase in the longer term, as the acquisition and maintenance of the skills necessary for verifying AI’s output is hindered by researchers’ and trainees’ increasing reliance on these tools.