Search

We tested whether large language models (LLMs) can help predict results from a complex behavioural science experiment. In study 1, we investigated the performance of the widely used LLMs GPT-3.5 and GPT-4 in forecasting the empirical findings of a large-scale experimental study of emotions, gender, and social perceptions. We found that GPT-4, but not GPT-3.5, matched the performance of a cohort of 119 human experts, with correlations of 0.89 (GPT-4), 0.07 (GPT-3.5) and 0.87 (human experts) between aggregated forecasts and realized effect sizes. In study 2, providing participants from a university subject pool the opportunity to query a GPT-4 powered chatbot significantly increased the accuracy of their forecasts. Results indicate promise for artificial intelligence (AI) to help anticipate—at scale and minimal cost—which claims about human behaviour will find empirical support and which ones will not. Our discussion focuses on avenues for human–AI collaboration in science.

Many-analysts studies explore how well an empirical claim withstands plausible alternative analyses of the same dataset by multiple, independent analysis teams. Conclusions from these studies typically rely on a single outcome metric (e.g. effect size) provided by each analysis team. Although informative about the range of plausible effects in a dataset, a single effect size from each team does not provide a complete, nuanced understanding of how analysis choices are related to the outcome. We used the Delphi consensus technique with input from 37 experts to develop an 18-item subjective evidence evaluation survey (SEES) to evaluate how each analysis team views the methodological appropriateness of the research design and the strength of evidence for the hypothesis. We illustrate the usefulness of the SEES in providing richer evidence assessment with pilot data from a previous many-analysts study.

Many publications on COVID-19 were released on preprint servers such as medRxiv and bioRxiv. It is unknown how reliable these preprints are, and which ones will eventually be published in scientific journals. In this study, we use crowdsourced human forecasts to predict publication outcomes and future citation counts for a sample of 400 preprints with high Altmetric score. Most of these preprints were published within 1 year of upload on a preprint server (70%), with a considerable fraction (45%) appearing in a high-impact journal with a journal impact factor of at least 10. On average, the preprints received 162 citations within the first year. We found that forecasters can predict if preprints will be published after 1 year and if the publishing journal has high impact. Forecasts are also informative with respect to Google Scholar citations within 1 year of upload on a preprint server. For both types of assessment, we found statistically significant positive correlations between forecasts and observed outcomes. While the forecasts can help to provide a preliminary assessment of preprints at a faster pace than traditional peer-review, it remains to be investigated if such an assessment is suited to identify methodological problems in preprints.

There is evidence that prediction markets are useful tools to aggregate information on researchers' beliefs about scientific results including the outcome of replications. In this study, we use prediction markets to forecast the results of novel experimental designs that test established theories. We set up prediction markets for hypotheses tested in the Defense Advanced Research Projects Agency's (DARPA) Next Generation Social Science (NGS2) programme. Researchers were invited to bet on whether 22 hypotheses would be supported or not. We define support as a test result in the same direction as hypothesized, with a Bayes factor of at least 10 (i.e. a likelihood of the observed data being consistent with the tested hypothesis that is at least 10 times greater compared with the null hypothesis). In addition to betting on this binary outcome, we asked participants to bet on the expected effect size (in Cohen's d) for each hypothesis. Our goal was to recruit at least 50 participants that signed up to participate in these markets. While this was the case, only 39 participants ended up actually trading. Participants also completed a survey on both the binary result and the effect size. We find that neither prediction markets nor surveys performed well in predicting outcomes for NGS2.

The Defense Advanced Research Projects Agency (DARPA) programme ‘Systematizing Confidence in Open Research and Evidence' (SCORE) aims to generate confidence scores for a large number of research claims from empirical studies in the social and behavioural sciences. The confidence scores will provide a quantitative assessment of how likely a claim will hold up in an independent replication. To create the scores, we follow earlier approaches and use prediction markets and surveys to forecast replication outcomes. Based on an initial set of forecasts for the overall replication rate in SCORE and its dependence on the academic discipline and the time of publication, we show that participants expect replication rates to increase over time. Moreover, they expect replication rates to differ between fields, with the highest replication rate in economics (average survey response 58%), and the lowest in psychology and in education (average survey response of 42% for both fields). These results reveal insights into the academic community's views of the replication crisis, including for research fields for which no large-scale replication studies have been undertaken yet.

The 2014 Research Excellence Framework (REF2014) was conducted to assess the quality of research carried out at higher education institutions in the UK over a 6 year period. However, the process was criticized for being expensive and bureaucratic, and it was argued that similar information could be obtained more simply from various existing metrics. We were interested in whether a prediction market on the outcome of REF2014 for 33 chemistry departments in the UK would provide information similar to that obtained during the REF2014 process. Prediction markets have become increasingly popular as a means of capturing what is colloquially known as the ‘wisdom of crowds’, and enable individuals to trade ‘bets’ on whether a specific outcome will occur or not. These have been shown to be successful at predicting various outcomes in a number of domains (e.g. sport, entertainment and politics), but have rarely been tested against outcomes based on expert judgements such as those that formed the basis of REF2014.