| 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 | Patent lawyers with an AI assistant wrote better drafts. On an unaided task at the end of the study, senior lawyers scored 0.45 standard deviations higher while juniors showed no average gain. | | 02 | In the second year after the GDPR, EU firms stored about 26 percent less data and processed about 15 percent less than comparable US firms, which the authors’ model turns into a 22 percent higher cost of data. | | 03 | Online job postings, the data behind most claims that employers have too much market power, overstate labor market concentration by about 3x because small and mid-sized employers post far fewer of their openings online. Findings about which skills employers want, including the shift toward tech skills, survive the correction. | | |
|
| Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting |
| David Autor, Tanya Rodchenko, Josh Martin, Zanna Iscenko, Scott Strand, David Pearl, and Melissa Ferere. NBER Working Paper 35720, September 2026. |
| THEIR QUESTION | Does working with an AI assistant build professional skill or hollow it out, and does the answer differ for senior and junior practitioners? | | WHAT THEY FOUND | Patent lawyers given an AI assistant produced drafts rated 0.34 standard deviations better after ten days and 0.38 after ninety. On an end-of-study editing task participants were instructed to complete without AI, the treated group scored 0.32 standard deviations higher, but the advantage was confined to senior lawyers. | | WHY IT MATTERS | Doing better work with AI and being better without it are different things. This is the first test of that distinction among working lawyers, and the unaided advantage shows up only among the experienced ones. | | HOW THEY KNOW | A registered randomized trial with 133 lawyers at eleven firms, blinded scoring, and an unaided end-of-study task. Only 91 lawyers finished that task, 24 of them juniors, and the no-AI rule could not be enforced. | | DISCLOSURES | Google funded the study, built the tool, and employs or contracts six of the seven authors. Autor acknowledges support from Google’s visiting fellows program. | |
| Read the paper |
|
| Data, Privacy Laws and Firm Production: Evidence from the GDPR |
| Mert Demirer, Diego J. Jiménez Hernández, Dean Li, and Sida Peng. NBER Working Paper 32146, revised February 2026. Forthcoming, Journal of Political Economy. |
| THEIR QUESTION | How much did the GDPR change what firms do with data, and what did that change cost them? | | WHAT THEY FOUND | In the second year after the GDPR took effect, EU firms stored about 26 percent less data and did about 15 percent less data processing than comparable US firms on the same cloud provider. A model of how firms combine data and computation implies the regulation raised the cost of data by about 22 percent, and variable production costs by 0.62 percent for software firms, 0.16 percent for services, and 0.08 percent for manufacturing. | | WHY IT MATTERS | Privacy rules get argued as either free or ruinous. This paper puts numbers on one side of the ledger, what firms stopped doing with data and what that cost. | | HOW THEY KNOW | Difference-in-differences on account-level usage for roughly 37,000 domestic firms at one cloud provider, half in the EU, through March 2020. The usage results are well identified; the cost figures rest on an imposed production function and exclude fixed compliance costs, privacy benefits, innovation, and misallocation. Excludes multinationals. | | DISCLOSURES | All four authors report current or former Microsoft ties. | |
| Read the paper |
|
| Do Online Job Postings Capture Job Vacancies? |
| Michael Dalton, Lisa B. Kahn, and Andreas I. Mueller. NBER Working Paper 35697, September 2026. |
| THEIR QUESTION | Do online job postings, a widely used source in labor market research, actually represent all job openings? | | WHAT THEY FOUND | Online postings overstate labor market concentration by about 3x because small and mid-sized employers post far fewer of their openings online. Rankings of markets by concentration, and measures of skill demand and of how skill requirements change over the business cycle, hold up after reweighting to the government’s vacancy survey. The gap has narrowed since 2007. | | WHY IT MATTERS | Postings data are how most people now measure what employers want. This paper says the data make local labor markets look about three times more concentrated than they are, mostly by missing small employers, while which markets are most concentrated, and most results about skill demand, survive the correction. | | HOW THEY KNOW | The authors treat the BLS’s JOLTS survey as a representative benchmark and reweight the commercial data to match it. Reweighting corrects only on establishment characteristics the data record, so the threefold figure carries real uncertainty. | | DISCLOSURES | The authors report no funding or relevant financial relationships. Dalton is a BLS economist. | |
| Read the paper |
|