Results from a massive research project investigating whether previously reported scientific results can be replicated raises questions about their reliability.gettyA major, new examination of the replicability and reliability of published social and behavioral science experiments has found that only about half of previously published results can be replicated by new studies.That’s one of the main conclusions from the Systematizing Confidence in Open Research and Evidence (SCORE) project, which involved an analysis of thousands of scientific studies published in 62 different journals between 2009 and 2018. The large-scale, international project was funded by the U.S. Defense Advanced Research Projects Agency (DARPA).The reanalyses focused on research in the fields of criminology, economics, educational science, health sciences, leadership, marketing, organizational behaviour, psychology, political science, public administration and sociology. The results are contained in three papers published this week in the highly respected journal Nature:Investigating the reproducibility of the social and behavioural sciences, Olivia Miske et al., Nature, online 1 April 2026, doi: 10.1038/s41586-026-10203-5.Investigating the analytical robustness of the social and behavioural sciences, Balazs Aczel et al., Nature, online 1 April 2026, doi: 10.1038/s41586-025-09844-9.MORE FOR YOUInvestigating the replicability of the social and behavioural sciences, Andrew Tyner et al., Nature, online 1 April 2026, doi: 10.1038/s41586-025-10078-y.The first paper assessed reproducibility, addressing the question of whether the same results of an experiment will be found when its original data are analyzed again by a different set of researchers. Using a random sample of 600 papers published between 2009 to 2018, the SCORE team was able to obtain the original data from 144 papers, and they reconstructed the data for 38 other studies.From those 182 datasets, they determined that in 53.6% of the cases, the results were “precisely reproducible;” and in 73.5% of them, the results were “approximately reproducible,” meaning that they fell within 15% of the original effects or within .05 of the original probability values. They found higher reproducibility for papers from political science and economics compared with other fields; and recent papers performed better than older papers, a reflection perhaps of the open-science movement picking up steam during the latter half of the sampled time period. Results were also more reproducible for papers that were published in journals requiring data sharing. The second paper examined the robustness of the original studies, a measure of the degree to which alternative ways of analyzing a study’s original data will yield the same results as the published experiment. Of the 100 papers the researchers reanalyzed, five different reanalysts independently confirmed the results reported in the original study about a third of the time. At least one of the reanalyses arrived at basically the same conclusion 74% of the time. But in 24% of the papers, the reanalyses failed to find the effects that the original paper had reported, and in 2% of the cases, a different analysis came to the opposite conclusion of the published research. The third paper examined replicability, which is the key question that’s often asked, especially about a novel research finding — will it be confirmed when other investigators repeat the experiment. The research team attempted replications of 274 claims of positive results from 164 papers published over the years in 54 different journals.They were able to replicate statistically significant results for 151 of the original 274 claims, a 55.1% replication rate. Replication success was found for 49.3% of the 164 papers. But beyond those numbers was another finding that should cause concern — the average correlational effect size was 0.25 for the original studies, but it shrank to 0.10 in the replication studies. That’s a big decline, and it occurred in every one of the disciplines. The replication results also differed by field, but in different directions than for reproducibility. Economics had the lowest replication rate, while Education, which did poorly on reproducibility, did better than the other fields in terms of replication. Predicting A Study’s Replicability Is DifficultThe so-called “replication crisis” is not confined to the social and behavioral sciences. It’s a long-recognized problem that has concerned investigators in several fields, including medicine. That’s led to a search for any indicators that could signal the likelihood that a study either might lack credibility or be particularly trustworthy. In their new research, SCORE investigators attempted to determine if certain characteristics of a study (e.g, whether it offereddata-sharing or how frequently it was cited by others) would allow them to predict whether its findings could be repeated. But the search for credibility indicators was largely unsuccessful. Of all the possible relationships, only one factor correlated highly with reproducibility, and that was whether a study’s data had been made publicly available.The lack of a easily available corrective means more will need to be learned about what constitutes best practices for improving the durable credibility of research experiments. “A lot more evidence is needed before we would be confident in a valid, scalable solution,” Brian Nosek, a leader of the SCORE project and director of the Center for Open Science at the University of Virginia, told Science Advisor.The relatively high percentage of replication failures has led some critics to allege that scientific research is plagued by sloppiness, or even worse, fraud. But the situation is much more complex than that. While dishonesty or carelessness may be a culprit in some un-replicated studies, other complicating factors involved in the scientific process are probably more often at play. For example, some of the variables studied by social scientists may have relatively small effects in the real world, making them difficult to detect consistently in experiments. And even small variations in experimental methods or differences among the subjects who participate in a study can affect the results. So, because “exact” replications are not possible, exactly duplicated results should not be expected.Recognizing the replication problem should lead to a healthy skepticism about initial results and to a greater priority being placed on replication research itself. Abel Brodeur, founder of the Institute for Replication at the University of Ottawa, told Nature that he’s always skeptical of newly published papers. “I’m going to wait years for lots of other studies, doing similar analyses, finding similar results before feeling confident in any given finding,’’ he said.“This is the world’s largest research project to date investigating the reliability of reported scientific results, and an example of how large-scale collaborations can address questions that no single research group could answer alone,” added Gustav Nilsonne, an associate professor of neuroscience at Sweden’s Karolinska Institutet, who was a co-author on all three of the Nature papers. “I hope we will see systematic replication attempts in more fields of research in the future.”