Does Scientific Writing Converge to U.S. English? Evidence from Generative AI-Assisted Publications

Using a large-scale analysis of 5.65 million scientific articles, this study finds that generative AI tools are driving non-English-speaking authors to increasingly converge toward U.S. English stylistic norms, particularly in contexts where language barriers have historically been most significant, thereby reducing publication obstacles while raising questions about linguistic diversity.

The Gist

Imagine the world of scientific research as a massive, global library. For decades, to get your book (your research paper) on the shelves of the most prestigious sections, you had to write it in a very specific, polished dialect of English. If you were a brilliant scientist from a country where English isn't the first language, you faced a huge hurdle: you had to pay expensive editors or spend years studying in English-speaking countries just to make your writing sound "right." It was like trying to enter a VIP club where the bouncer only lets in people who speak with a specific accent.

Then, in late 2022, a new tool arrived: Generative AI (like ChatGPT). Think of this AI as a magical, instant translator and editor that anyone can use for free or cheap.

This paper asks a simple but profound question: Did this magic tool just help non-native speakers write better, or did it actually make everyone sound more like the "VIPs" (native U.S. English speakers)?

Here is the breakdown of what the researchers found, using some everyday analogies:

1. The Big Discovery: The "Accent" is Fading

The researchers looked at 5.65 million scientific papers published between 2021 and 2024. They used a special computer program (called SciBERT) that acts like a "linguistic fingerprint scanner." It doesn't just check for grammar; it measures how closely the style and vibe of a paper match the average style of papers written by scientists in the United States.

The Result: After ChatGPT came out, papers written by scientists from non-English-speaking countries started sounding much more like American scientific papers.

• Before 2022: Papers from different countries had their own unique "flavor" or "accent."
• After 2022: The papers that used AI tools started to lose that unique flavor and began to sound more like the standard "American" scientific voice.

2. Who Benefited the Most? (The "Underdogs")

The study found that this change wasn't equal for everyone. It was most dramatic for the people who needed help the most:

• Solo Teams: Scientists working alone in their home countries (without international partners) saw the biggest shift. It's like a solo musician suddenly getting a backing track that makes them sound like they are in a major orchestra.
• Linguistically Distant Countries: Scientists from countries where the language is very different from English (like China or Brazil) changed their writing style more than those from countries where English is already common (like the Netherlands).
• Smaller Journals: The change was strongest in lower-tier journals. In top-tier journals, the writing was already very polished, so there was less room to improve. But in smaller journals, the AI acted like a "leveling field," helping researchers there sound just as professional as the big shots.

3. The Double-Edged Sword

The authors compare this to global trade.

• The Good News: Just as lowering trade barriers allows more countries to sell their goods, lowering language barriers allows more brilliant ideas from around the world to enter the global conversation. It's fairer. It means a genius in a remote village can finally get their work heard without needing a fortune for an editor.
• The Bad News: If everyone starts speaking the exact same way, we might lose the unique "flavor" of different cultures. Imagine a world where every chef uses the exact same recipe book because it's the "standard." The food might be safe and consistent, but we might lose the unique, spicy, or weird dishes that made the world's cuisine interesting. The researchers worry that science might become too "homogenized" (all the same).

4. The "Magic Wand" vs. The "Real Thing"

The paper suggests that AI is acting as a partial substitute for human editors.

• In the past, if you wanted to sound like a U.S. scientist, you needed a human editor (or a human mentor).
• Now, the AI does that job. It helps non-native speakers bridge the gap.
• However, the researchers warn that while the words sound better, we don't know if the ideas are better. It's like using a filter on a photo: the picture looks sharper and more professional, but is the subject actually more beautiful, or just more polished?

The Bottom Line

Generative AI is breaking down the "language gate" in science. It is helping non-native speakers sound more like the dominant U.S. standard, which is great for getting their work published and heard.

But there is a catch: We are trading linguistic diversity for standardization. The big question for the future is: Do we want a world where everyone speaks the same "scientific language" to be heard, or do we want to find a way to keep the unique voices of different cultures alive while still being understood?

The researchers suggest that journals and policymakers need to be smart about this. They should encourage the use of AI to help people get in the door, but they shouldn't let AI detectors punish non-native writers, because those detectors often mistake "good AI help" for "fake writing."

In short: AI is the great equalizer, but it might also be the great homogenizer.

Technical Summary

Here is a detailed technical summary of the paper "Does Scientific Writing Converge to U.S. English? Evidence from Generative AI-Assisted Publications."

1. Problem Statement

The global scientific community is dominated by English, creating structural advantages for native speakers and persistent barriers for non-native speakers (e.g., higher writing burdens, peer-review bias, and career progression hurdles). While previous studies have documented that Generative AI (GenAI) improves readability and vocabulary, they largely focus on absolute changes in writing quality.

The critical unanswered question is whether GenAI use leads to systematic convergence toward the dominant linguistic standard (U.S. English) or merely improves writing in a vacuum. Understanding this distinction is vital: convergence suggests a reduction in language-based entry barriers but also raises concerns about the homogenization of scientific discourse and the loss of linguistic diversity.

2. Methodology

Data Source and Sample

• Dataset: 5.65 million English-language scientific articles published between 2021 and 2024, indexed in Scopus.
• Unit of Analysis: Publication $\times$ Country $\times$ Field cell (totaling ~14.3 million observations).
• Scope: Focuses on papers with at least one author affiliation in a non-English-speaking country.
• Fields: Aggregated into four broad categories: Life Sciences, Physical Sciences, Engineering & Technology, and Social Sciences.

GenAI Identification Strategy

The authors do not rely on self-reported usage but instead detect GenAI adoption through lexical pattern shifts in titles and abstracts.

• Keyword List: A 65-term vocabulary (e.g., delve, groundbreak, intricate, pivotal, unlock) previously linked to GenAI content.
• Filtering: Terms are selected based on a field-specific frequency increase of at least 300% (quadrupling) between 2021 and 2024.
• Classification: A paper is flagged as "GenAI-assisted" if at least one filtered keyword appears in its title or abstract.

Linguistic Similarity Measurement

• Model: SciBERT (a BERT model pre-trained specifically on scientific corpora) is used to generate 768-dimensional text embeddings.
• Process: Titles and abstracts are concatenated, tokenized, and processed to create mean-pooled vectors.
• Metric: Cosine similarity is calculated between non-U.S. publications and a "pure" U.S. benchmark corpus (papers with only U.S. affiliations) within the same scientific field and publication year.
• Goal: To measure how closely non-U.S. writing aligns with the U.S. stylistic and semantic register.

Empirical Strategy

• Design: An Event-Study Difference-in-Differences (DiD) model with repeated cross-sections.
• Treatment: GenAI-assisted papers vs. non-GenAI-assisted papers.
• Shock: The late-2022 release of ChatGPT serves as the intervention point.
• Baseline: 2022 is the reference year.
• Controls: Fixed effects for Country, Field, Journal, Year, and Journal $\times$ Year interactions.
• Equation:
  $$y_{ifcjt} = \delta_0 \cdot \text{GenAI}_{ifcjt} + \sum_{\tau} \beta_\tau \cdot \mathbb{1}\{t=\tau\} \times \text{GenAI}_{ifcjt} + \text{Controls} + \text{Fixed Effects} + \epsilon$$
  Where $\beta_\tau$ captures the differential change in linguistic similarity for GenAI papers relative to the 2022 baseline.

3. Key Results

Primary Finding: Linguistic Convergence

• Post-2022 Shift: GenAI-assisted publications from non-English-speaking countries exhibit a statistically significant and increasing convergence toward U.S. scientific English compared to non-GenAI-assisted papers from the same countries.
• Magnitude: By 2023, the convergence effect is ~0.15%; by 2024, it rises to ~0.4% relative to the 2022 baseline.
• Timing: No pre-trend was observed in 2021–2022, confirming the shift coincides with the widespread adoption of GenAI tools.

Heterogeneity Analysis

The convergence effect is not uniform; it is strongest in contexts where language barriers are historically highest:

1. Author Teams: The effect is significantly larger for domestic author teams (all authors from the same non-English country) compared to international teams.
2. Linguistic Distance: Convergence is most pronounced for authors from countries linguistically distant from English (low Common Language Index scores).
3. Co-authorship: Papers with no English-speaking co-authors show stronger convergence than those with English-speaking collaborators.
4. Journal Impact: The effect is stronger in lower-impact journals, suggesting GenAI substitutes for professional editing services which are less accessible to authors in these venues.

Robustness Checks

The findings hold under various alternative specifications:

• Keyword Thresholds: Results persist with stricter (400% increase) and looser (200% increase) frequency thresholds.
• Identification Criteria: Results hold when requiring at least two GenAI-linked terms (reducing false positives).
• Benchmark Definitions: Convergence remains robust when the U.S. benchmark is restricted to:
  • Non-GenAI papers only.
  • Pre-2023 papers (pre-ChatGPT).
  • Top 10% of journals by impact factor (convergence is actually stronger here).

4. Key Contributions

1. Conceptual Shift: Moves beyond measuring "absolute improvement" (readability) to measuring convergence (alignment with a dominant standard), providing a new lens on how technology alters global scientific norms.
2. Scale and Scope: Provides the first large-scale, cross-field, and cross-country evidence (5.65M papers) linking GenAI adoption to linguistic shifts in the peer-reviewed record, surpassing previous studies limited to preprints or single disciplines.
3. Mechanism Identification: Identifies GenAI as a partial linguistic equalizer, disproportionately benefiting researchers with the highest baseline language frictions (domestic teams, distant languages, lower-tier journals).

5. Significance and Implications

• Reduced Barriers: GenAI appears to be lowering entry barriers for non-native English speakers, potentially diversifying the pool of ideas and contributors in international journals.
• Trade Dynamics Analogy: The authors draw parallels to international trade theory, suggesting that reducing "language tariffs" could increase the variety and quality of global research, similar to how trade liberalization increases product variety.
• Risk of Homogenization: While inclusion is a positive outcome, the study warns that convergence toward a single dominant register (U.S. English) may lead to stylistic homogenization, potentially eroding rhetorical diversity and unique cultural perspectives in science.
• Policy Recommendations:
  • Disclosure: Journals should adopt transparent, low-friction disclosure norms for AI language assistance rather than relying on unreliable AI detectors (which disproportionately flag non-native writing).
  • Equity: Funding bodies should support equitable access to GenAI tools for under-resourced institutions.
  • Evaluation: Reviewers and editors must be trained to distinguish between language clarity (aided by AI) and scientific merit.

Conclusion

The study concludes that GenAI is actively reshaping the linguistic landscape of science, driving a measurable convergence toward U.S. English norms. While this acts as a powerful equalizer for non-native speakers, particularly those facing the greatest linguistic and institutional constraints, it necessitates careful institutional management to ensure that the benefits of inclusion do not come at the cost of scientific diversity.