High Average Accuracy Coexists with Substantial Between-Setting Variability: A Systematic Review and Meta-analysis of Large Language Models for Data Extraction in Evidence Synthesis

Every time a team of researchers sets out to understand a medical treatment, they must first gather the facts from dozens, sometimes hundreds, of individual studies. This process, known as a systematic review, is the bedrock of modern medical evidence. It requires scientists to read through dense reports and pull out specific numbers: how many patients were involved, what dosage they received, and what the outcomes were. If a single number is copied incorrectly, the entire conclusion of the review can shift, potentially leading doctors to recommend the wrong treatment. For decades, this work has been done by human hands, a slow and meticulous task where two people often read the same document independently to catch mistakes. It is the gold standard for accuracy, but it is also incredibly time-consuming and expensive.

In recent years, a new tool has entered this arena: large language models. These are powerful computer programs trained on vast amounts of text that can read a scientific paper and attempt to pull out the same numbers a human would. The hope has been that these artificial intelligence systems could do the heavy lifting, speeding up the process and freeing up human experts for more complex thinking. But as these tools have become more common, a critical question has emerged: can we trust them to be as accurate as humans, and does their performance stay consistent no matter what kind of information they are asked to find?

A new study by Carlos Darcy Alves Bersot and his colleagues takes a deep, statistical look at this question. They did not simply ask if the machines work; they gathered every available report that tested these models against a human-made standard and combined the results to see the full picture. What they found is a story of two very different realities coexisting. On one hand, the average performance of these artificial intelligence systems is surprisingly high. On the other hand, that average hides a chaotic reality where performance can swing wildly depending on the specific task at hand.

The researchers collected data from forty-six different reports, representing forty-five independent investigations. They focused specifically on the moment before a human ever looks at the machine's work. They wanted to know: if a large language model is asked to extract a piece of information, how often does it get it right on the first try? When they pooled the results from eight key studies that met their strict criteria, the answer was encouraging. The models got the right answer about 86.4 percent of the time. This number suggests that, in a general sense, these tools are quite capable.

However, the story does not end with that average number. The researchers looked deeper and found that this high average was built on a foundation of extreme inconsistency. They calculated a range that shows what kind of performance one might expect in a new, unseen situation. This range stretched from a low of 19.2 percent to a high of 99.4 percent. To put this in perspective, imagine a student taking a test who gets an average score of 86 percent. If that student's scores on individual tests ranged from failing to perfect, you would not feel confident sending them into a classroom to grade papers. The same logic applies here. The fact that the average is high does not guarantee that the machine will perform well in a specific, real-world scenario.

The study reveals that the type of information being extracted makes a massive difference. When the task involves simple, clearly stated facts, the models perform very well. But when the task requires finding numbers buried in complex tables, interpreting ambiguous text, or pulling out data that is missing or poorly reported, the accuracy drops significantly. In one specific case where the model had to find open-ended information that required recalling details from memory rather than just copying them, the accuracy plummeted to just 17 percent. This single study, with its very different task, pulled the overall average down and widened the range of possible outcomes. When the researchers removed this outlier to look at more similar, structured tasks, the average accuracy rose to 91.2 percent, but even then, the performance could still vary widely from one setting to another.

The researchers also examined how these tools handle numerical data, which is the most critical part of a medical review. Here, the results were mixed. In some highly structured tests, the models achieved near-perfect accuracy, getting 99 percent of the numbers right. In other studies with different types of data, the accuracy fell to around 63 percent. This inconsistency means that a researcher cannot assume that because a model is good at one thing, it will be good at everything. A model that excels at finding a drug name might fail miserably at finding the exact number of patients who experienced a side effect.

Another crucial finding concerns the workflow. Many studies looked at how these tools perform when humans are involved in the process, such as when a human checks the machine's work or when the machine acts as a second reader. These human-assisted workflows often show better results and save time. However, the study makes a clear distinction: the speed and accuracy of a human working with a machine are not the same as the machine working alone. The time savings reported in some studies often come from the machine doing the initial draft, but the human still has to spend time verifying and correcting the work. The machine does not eliminate the need for human oversight; it merely changes the nature of the work.

The researchers also looked at the quality of the studies they reviewed and found that many had methodological flaws. Some studies did not clearly explain how they tested the models, or they used the same data to both train the model and test it, which can make the results look better than they truly are. Because of these limitations and the small number of truly independent studies available, the researchers cannot say with absolute certainty that these tools will work perfectly in every future situation. The evidence suggests that while the tools are powerful, they are not yet reliable enough to replace human reviewers entirely.

The conclusion of this work is not that artificial intelligence is a failure, but that it is a tool that requires careful handling. The high average accuracy is real, but it is a misleading benchmark if taken out of context. It tells us that the technology has potential, but it also warns us that performance is fragile. A system that works well for one type of data extraction might fail completely for another. The researchers argue that the best path forward is not to trust the machines blindly, but to use them as part of a risk-adapted strategy. For simple, clear tasks, they can be used with confidence. For complex, numerical, or ambiguous data, human verification remains essential.

Ultimately, this study shifts the conversation from "Can the machine do it?" to "Under what conditions can the machine do it?" The answer is that the machine can do it well, but only when the conditions are right. The high average score is a promise of capability, but the wide variability is a warning of unpredictability. For the scientists who rely on these reviews to guide medical practice, the lesson is clear: the machine is a helpful assistant, but it is not yet a replacement for the careful, critical eye of a human expert. The future of evidence synthesis will likely involve a partnership where the machine handles the volume and the human ensures the validity, with the understanding that the machine's performance is never a guarantee, but a variable that must be constantly checked.