Jeff Macher on Generative AI and the Future of Global Researc

Episode Summary

Executive Summary: Scott Walston speaks with Jeff Bacher about research showing generative AI is rapidly spreading in global science, especially in countries farther from English, and is making non-U.S. scientific writing more linguistically similar to U.S. papers. They discuss how this may democratize participation, raise editorial challenges, affect journal selection, and potentially reshape innovation and U.S. competitiveness.

Main Topics: Generative AI as a linguistic equalizer (Priority: 5/5): The paper finds AI uptake rose sharply after ChatGPT and was strongest in countries linguistically distant from English, suggesting AI helps non-native English speakers participate more fully in science. Convergence in scientific writing style (Priority: 5/5): As AI use rises, non-U.S. papers increasingly resemble U.S. scientific papers in wording and style, especially in non-English-speaking countries without native English coauthors. Top-tier vs lower-tier journal effects (Priority: 4/5): The AI effect is weaker in high-impact journals, likely because those authors already have stronger English support and editorial assistance; lower-tier journals show more visible uptake. Editorial and reviewer implications (Priority: 4/5): AI may reduce desk rejections based on prose quality and increase reviewer efficiency, but it raises concerns about creativity, policy ambiguity, and the quality of AI-assisted peer review. Measuring AI use and study limitations (Priority: 4/5): The authors used keyword-based lexical markers, Scopus data, and embedding methods to infer AI-assisted writing, but acknowledge limitations from using only titles/abstracts and not full text. Broader innovation and U.S. competitiveness (Priority: 5/5): The conversation expands to how AI may boost global innovation while potentially eroding U.S. relative advantage, especially if basic research funding weakens. Basic research, disruption, and policy (Priority: 4/5): Bacher argues strong basic science ecosystems support long-term innovation, and policymakers should be careful about picking winners while preserving university-industry links.

Key Arguments: Generative AI is spreading quickly in scientific publishing, with uptake increasing across countries and fields after ChatGPT's release. The strongest AI adoption occurs in linguistically distant regions such as the Middle East and Asia, implying AI helps overcome English-language barriers. Non-U.S. papers are converging linguistically toward U.S. papers, especially when all authors are domestic non-native English speakers and in lower-impact journals. High-impact journals see less of the effect because authors there often already have strong English editing and more stringent submission standards. AI appears to make desk rejection harder because better-written papers are less likely to be rejected on presentation alone, shifting more screening work to reviewers. The study’s AI-detection approach is conservative, using pre/post-ChatGPT lexical markers and threshold tests; the authors view it as a lower bound. A more advanced approach using modern tools and full-text analysis would likely improve detection, but current computational limits constrained the study. AI may broaden participation and potentially increase scientific productivity, but it could also crowd out some previously publishable work and intensify competition. More participation and better writing do not necessarily mean more creativity; Bacher argues current LLMs remain weak at genuine novel idea generation. Strong basic research ecosystems historically support innovation and commercialization, and weakening them may reduce U.S. competitiveness over time. AI could counteract the slowdown in disruptive innovation, but if it is widely democratized, the benefits may accrue globally rather than disproportionately to the U.S.

Data Points: Paper corpus size: 5.6 million papers - Scale of the dataset used to analyze AI uptake in non-English-speaking authors' publications. Keyword list: 65 common GenAI keywords - Lexical markers used to detect possible AI-assisted writing in titles and abstracts. Pre/post window: 2 years before and 2 years after ChatGPT - Time window used to compare publication language and AI-marker changes. Keyword threshold: More than 300% increase - Base criterion for classifying a paper as GenAI-assisted based on marker growth from 2021 to 2024. Alternative threshold: Up to a five-fold increase - Robustness check showing results were similar under stricter definitions. Text representation: 768-dimensional text vector - Computational embedding dimension used in the similarity analysis. Author productivity example: Reviews take about a day; some people reduce them to two hours with ChatGPT - Used to illustrate reviewer incentives to use AI assistance. Scientific field buckets: 4 buckets - Physical science, life science, engineering and technology, and social science. Scope of analysis: Across many Scopus categories - The paper examines broad field-level patterns rather than one discipline only.

Pivotal Quotes: "Generative AI as a Linguistic Equalizer in Global Science" — Scott Walston: Title of the paper being discussed and the framing concept of the episode. "This is kind of the democratization effect of AI." — Jeff Bacher: Explaining the main positive interpretation of broader AI adoption in scientific writing. "What I see ... is more stuff is going out to reviewers than is being desk rejected simply because it's written better." — Jeff Bacher: On how AI-assisted writing may change editorial screening and review burden.

Implications: AI is likely to broaden global participation in science and change publishing norms, but it may also intensify competition, shift editorial standards, and raise new questions about innovation policy, peer review, and U.S. research competitiveness.