Can AI support rapid evidence synthesis for policy? An evaluation of Elicit: The AI Research Assistant
This study evaluates the AI tool Elicit for policy evidence synthesis, finding that while it effectively expedites preliminary topic scoping, it currently lacks the precision and reliability required for complex rapid reviews without significant human oversight.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of public health and government policy, decisions about what treatments or programs to fund rely on a foundation of evidence. Before a new rule is written, experts must gather all the existing research on a topic, read it carefully, and summarize what it says works and what does not. This process, known as evidence synthesis, is essential for making fair and effective choices, but it is also incredibly time-consuming. Traditionally, teams of researchers spend months searching through thousands of scientific articles, checking if each one fits the specific question at hand, and then extracting the data to build a clear picture. As the needs of policymakers become more urgent and complex, there is a growing desire to speed up this work without losing accuracy. This is where artificial intelligence enters the scene, promising to act as a powerful assistant that can read and sort information at a speed no human could match.
A team of researchers from universities in the United Kingdom decided to put one of these new AI tools to the test. They wanted to see if a software program called Elicit could truly help policy researchers do their jobs faster and better. The researchers chose to evaluate the tool in two different ways. First, they asked if the AI could act as a helpful guide to get a quick overview of a broad topic, a step often called "scoping" that happens before a full review begins. Second, they tested whether the AI could take the place of human researchers in conducting a full, rapid review of evidence on a complex subject. The subject they chose was a difficult one: finding out which programs successfully encourage physical activity among older adults who are facing social and economic disadvantages. This topic was chosen because it involves many moving parts, including different types of people, various kinds of activities, and specific measures of success, making it a rigorous test for any automated system.
The researchers began by asking the AI to perform a quick scan of the literature to help them understand the landscape of physical activity tools used in primary care. In this role, the software proved quite useful. It generated a summary of the available evidence in just six minutes, offering a broad view of the topic that experts found helpful for shaping their thinking. The output was not perfect, and it lacked the detailed explanation of how the information was found, but it served well as a starting point. It gave the team a sense of what was out there and where the gaps might be, effectively acting as a fast, preliminary map before they began the harder work of drawing the full territory.
However, when the researchers asked the AI to perform the full, rigorous task of a rapid review, the results were far less encouraging. They fed the same complex question about physical activity in disadvantaged older populations into the system and watched how it handled the job. The process took over three hours, during which the human researchers had to spend significant time refining the questions and correcting the AI's assumptions. When the final report was produced, it contained a list of forty studies that the AI claimed were relevant. Yet, when the human experts checked these studies against the strict rules of the review, they found that thirty-six of them did not actually belong. The AI had included studies that looked at the wrong outcomes, focused on the wrong groups of people, or examined interventions that did not fit the criteria. In fact, the software missed thirty-four studies that the human team had successfully found and included in their own review.
The core problem appeared to be that the AI struggled with the nuance of the language used in scientific papers. It could easily identify words like "older" or "physical activity," but it failed to understand the deeper meaning of whether a study was actually testing the right thing for the right group of people. For instance, it could not distinguish between a program designed to improve general health and one specifically designed to increase physical activity, even if the wording was similar. Because of these errors, the final summary produced by the AI was filled with information that did not answer the original question. It discussed outcomes that were not relevant and drew conclusions based on evidence that was not eligible for the review. The human-led review, by contrast, was able to filter out the noise and focus only on the studies that truly mattered, resulting in a much more accurate and trustworthy conclusion.
The researchers also looked at how the software handled the data it did find. While the AI was able to pull out basic details like the country where a study took place or the number of participants, it was inconsistent when it came to extracting specific results, such as how much an intervention improved a person's health. In some cases, it missed the data entirely; in others, it was unclear what it had found. Furthermore, the software did not have the ability to judge the quality of the studies, meaning it treated a weak, poorly designed experiment the same as a strong, well-conducted one. This lack of quality control is a significant risk, as it could lead policymakers to rely on flawed evidence.
Ultimately, the study suggests that while artificial intelligence can be a valuable tool for getting a quick, high-level overview of a topic, it is not yet ready to replace human researchers in the critical task of conducting a full evidence review, especially when the questions are complex. The AI can speed up the early stages of research, but it still makes too many mistakes in the later stages to be trusted on its own. The researchers concluded that for now, the best approach is to use the software as a helper that works alongside a human expert, who must check every decision the machine makes. The time saved by using the AI is currently unclear, as the human effort required to correct its errors might cancel out the initial speed gains. As these tools continue to evolve, the hope is that they will become better at understanding the subtle details of human language and research, but for the moment, the human mind remains essential for ensuring that the evidence guiding public policy is accurate and reliable.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.