Introducing Living Science

Papers have been the de facto unit of knowledge in science. But science is not static and a paper is not meant to be the final destination, especially in empirical research. A paper may propose a novel way to look at the world and offer a snapshot of what it finds. That snapshot is far from the whole story.

The world keeps changing after a paper is published. New data, new models, and new methods give us a chance to return to its questions: does the pattern persist, does the effect size remain the same, and what has changed? As AI agents become more capable, we have an opportunity to make revisiting important work a routine part of science.

Today, we are introducing Living Science, a project to keep influential research alive by revisiting its findings as new data, models, and methods become available. We are starting with economics, focusing on studies whose data sources are regularly updated. The goal is to not assess whether the paper was correct at the time of publishing but to provide resources and venues for continued discussions of influential research.

We use SAI's replication agent to attempt to reproduce each paper's key findings, documenting what matches and what does not, before extending the analysis to newer data where possible. The AI agent will first look for public replication packages if available and then identify relevant data sources for extending the analysis. As a result, we may not use exactly the same dataset and setup. The exercise is closer to writing a follow-up paper than strict replication. We sanity-check the results and reach out to the original authors for feedback. We also invite the community to identify issues and post comments, helping us improve the analyses over time.

We plan to add at least one paper per week. We welcome suggestions, especially nominations of papers you would like to see revisited. If you would like to help replicate studies, extend analyses, or discuss the findings, we would love to hear from you!

Our initial reports revisit questions about wages, unemployment, and regional economies. Here are a few findings:

• Higher minimum wages still raise pay without a clear overall drop in low-wage employment. Extending Cengiz and colleagues' (2019) analysis to 236 new state-level increases through 2024, we find affected workers' wages rise by 4.6%. The estimated employment change is −1.1%, statistically indistinguishable from zero. Read the report.

• Slower job finding still explains most unemployment fluctuations. Shimer (2012) attributed about 90% of unemployment fluctuations over 1987–2010 to changes in how quickly unemployed workers find jobs. Extending the analysis through June 2026 gives 89% when we exclude 2020. During the pandemic, job losses played a larger role, accounting for 38% of fluctuations over 2010–2026. Read the report.

• Local job losses still leave lasting scars, but fewer people move away in response. Extending Blanchard and Katz's (1992) analysis through 2025, we estimate that migration responses absorb 24-28% of a local job loss shock in the first year n the periods we study after 2008, down from 38% in 1978–1990 using comparable household employment data. Read the report.

In both periods, jobs disappear below the new minimum wage and increase at or above it. Across all 236 increases through 2024, affected wages rise by 4.6%, while the estimated employment change of −1.1% is statistically indistinguishable from zero. Read the report.

These examples show why it is worth returning to influential papers. Some findings hold up remarkably well; others change with the period we study. Sometimes, the data needed to revisit the original question are no longer available. Making these distinctions visible is part of keeping research useful.

Many researchers could already try some version of this on their own. Why do we still need a shared public record like Living Science?

First, researchers should be able to build on work already done. We make our replication repositories public on GitHub so that anyone can start from an existing replication, inspect its choices, and extend it.

Second, common knowledge matters. An analysis that stays on one person's computer does little to update our shared understanding. Others may continue citing an old result without knowing that newer evidence supports, qualifies, or changes it. A shared public record makes those updates visible and gives us a common reference for discussing what we know.

Third, science is a communal effort. Keeping research alive takes people who catch errors, question interpretations, contribute new analyses, and return as more data arrive. We hope Living Science will give these efforts a shared home, so that each update can become a starting point for the next.

We invite you to explore Living Science, check out the latest findings, and help build this shared record by nominating a paper you would like to see revisited.