| TPI’s Research Roundup is our monthly compilation of recent outside research for tech policy nerds. Authors are not affiliated with TPI. Each paper is here because we find it interesting and the method holds up. We do not necessarily agree with everything, or even anything, in these papers. |
| Send candidates to [email protected]. |
| COCKTAIL PARTY MATERIAL | | 01 | Firms that hired AI specialists faster have recently pulled ahead in productivity, after a decade of AI hiring with no measurable payoff. The authors say firms needed that time to learn how to use the technology. | | 02 | When Peru’s environmental regulator let an algorithm replace managers in assigning cases to its staff lawyers, they finished fewer cases on time, about 5 percentage points below a 40 percent baseline. | | 03 | Ruwiki, a regime-aligned 2023 copy of Russian Wikipedia, deleted only 0.14 percent of its articles in 20 months and changed 27 percent. The edits justify Russia’s wars and suppress accounts of its domestic critics. | | |
|
| Canaries in the Gold Mine: Early Productivity Gains from Artificial Intelligence Creating Organization Capital |
| Tania Babina, Alex X. He, and Renhao Jiang. NBER Working Paper 35684, August 2026. |
| THEIR QUESTION | Do firms that invest in AI become more productive, and if so, why did the gains take a decade to appear? | | WHAT THEY FOUND | Firms whose share of AI-skilled workers grew one standard deviation faster saw 7.5 percentage points more growth in total factor productivity and 6.2 points more growth in sales per worker over six years. The relationship appears in recent years but not in the decade before. The authors attribute the lag to the time firms need to build organization capital, the firm-specific know-how that AI work creates. | | WHY IT MATTERS | Skeptics cite flat aggregate productivity statistics as evidence that AI’s payoff is hype. This paper finds a firm-level payoff and offers an explanation for why it arrived late. | | HOW THEY KNOW | The authors measure AI investment by hiring of AI-skilled workers, not AI use. They instrument it with firms’ pre-2010 hiring ties to universities strong in AI. The exclusion restriction holds only if those ties affect productivity solely through AI hiring, and they may also proxy for general technical sophistication. | | DISCLOSURES | The NBER page lists no funding sources or financial relationships. | |
| Read the paper |
|
| Automated Assignment and Bureaucrat Productivity: Evidence from a Field Experiment in Peru |
| Alipio Ferreira and Sandra Cermeño León. Working paper, June 2026. SSRN 7483405. |
| THEIR QUESTION | When an algorithm replaces managers in deciding which cases go to which staff, does a government agency get more work done? | | WHAT THEY FOUND | At OEFA, Peru’s environmental enforcement agency, analysts randomly given algorithm-assigned caseloads finished about 5 percentage points fewer cases on time than the 40 percent baseline for analysts whose managers assigned their cases. Fines and appeals did not change. Showing analysts suggested completion dates raised on-time completion by about 4 points, mostly for cases managers had assigned. | | WHY IT MATTERS | Governments are turning to automated case management to clear backlogs. Here the rule did what it was built to do, spreading work evenly, and the agency got less done. The authors’ preregistration predicted the opposite. Their evidence suggests managers had been steering complex cases to experienced analysts, a match the algorithm broke. | | HOW THEY KNOW | The authors randomized at two levels inside the agency, algorithm assignment across 95 analysts and suggested dates across 1,405 cases. The main estimate is significant only at the 10 percent level, with a randomization-inference p-value of 0.07. The algorithm is a workload-balancing optimization rule, not machine learning. | | DISCLOSURES | The paper lists no outside funding. Cermeño León works at Innovations for Poverty Action, which partnered with OEFA on the study. | |
| Read the paper |
|
| The Anatomy of Censorship and Propaganda: Evidence from Russian Wikipedias |
| Ekaterina Zhuravskaya, Ulrich Matter, Ruben Durante, and Vladimir Avetian. CEPR Discussion Paper 21698, July 2026, revised August 2026. |
| THEIR QUESTION | When an authoritarian government takes editorial control of an encyclopedia, does it delete inconvenient material or rewrite it, and what subjects does it touch? | | WHAT THEY FOUND | Twenty months after Ruwiki was created by copying all 1.9 million articles of Russian-language Wikipedia, 27 percent of its articles had changed at least somewhat, 2.8 percent were heavily edited, and 0.14 percent were deleted. The edits legitimize Russia’s wars and suppress accounts of the domestic opposition and human rights abuses. Rewritten sensitive articles became harder to read. | | WHY IT MATTERS | Debates over online information control focus on takedowns and blocking. Here control worked mainly through rewriting, which readers without the original cannot see, and it reached past politics into culture, family, gender, and law. | | HOW THEY KNOW | Because Ruwiki began as an exact copy, the authors can compare each edited article with its original, using ordinary Wikipedia edits over the same period as a baseline for normal drift. Large language models classify the edits, and the paper’s summary reports no human validation rate. Ruwiki’s editors appear to be paid professionals, so some differences may reflect who edits rather than what the regime wants. | | DISCLOSURES | No funding or financial relationships appear on the CEPR or SSRN pages. | |
| Read the paper |
|
