AI promised to democratise academic publishing – the evidence says otherwise
Generative AI has been widely embraced as a tool that could level the playing field in academic publishing, giving non-native English speakers and researchers in under-resourced settings a fairer chance. Brendal Aformeziem draws on a growing body of evidence to show that the reality is more troubling. AI is being adopted fastest by the researchers who need it most, yet is failing to remove the structural barriers they face, and may be adding new ones.
For decades, the dominance of English in scientific publishing has imposed a measurable and well-documented burden on researchers who did not grow up speaking it. A landmark study published in Written Communication found that writing research articles in English, compared to Spanish, is perceived as 24% more difficult, 21% more anxiety-inducing, and 11% more dissatisfying. With English accounting for up to 90% of scientific publications globally, this is not a minor inconvenience. It is a structural disadvantage that shapes whose research gets written, submitted, reviewed, and read.
When generative AI arrived, it carried a compelling promise. Language models capable of drafting, restructuring, and polishing academic text seemed, at last, to offer a way around this disadvantage. Researchers from the Global South, early-career scientists at under-resourced institutions, and multilingual scholars could, in theory, produce prose that met the linguistic expectations of English-dominant journals. The playing field, long tilted, might finally level.
The evidence is beginning to tell a different story.
The researchers turning to AI are those who can least afford to be penalised for it
A large-scale analysis of over two million biomedical publications found that AI-assisted writing grew by approximately 400% in non-English-speaking countries between 2022 and 2024, compared with 183% in English-speaking ones. Adoption was highest among early-career researchers, those with fewer publications and citations, and scholars at lower-ranked institutions. A separate analysis of 5.2 million papers published in the Proceedings of the National Academy of Sciences confirmed the same pattern that authors from non-English-speaking countries are more likely to rely on AI writing tools, and this reliance is growing faster than anywhere else.
This is not surprising. These researchers are turning to AI for the same reason anyone turns to a tool, namely because they need it. The linguistic burden of writing research in a second language as documented by Hanauer and Englander did not disappear with the arrival of ChatGPT, but it created a demand that AI tools are now filling. The question is whether filling the demand is really helping?
AI polishes the prose, it does not remove the bias
A study examining 76,453 peer reviews across 20,827 papers over seven years, drawing on interviews with researchers across five continents, found that reviewers continue to penalise authors from countries where English is less widely spoken, even after ChatGPT became widely available. Bias decreased only slightly, and in most cases not significantly. More troubling still, reviewers have adapted, and ChatGPT-style phrasing has become a new cue. Words like “delve“, now strongly associated with AI-assisted writing, are being used by reviewers to infer author identity. Authors describe a “damned if you do, damned if you don’t” dynamic. Write naturally and risk being flagged as a non-native speaker; write with AI assistance and risk being flagged as an AI user.
Bias against non-native English speakers in peer review persists even when authors use large language models to improve their writing.
Research from the Stanford Graduate School of Education corroborates this finding. Bias against non-native English speakers in peer review persists even when authors use large language models to improve their writing. Reviewers use language as a proxy for identity, and when AI erases one set of linguistic cues, they shift to another. The discrimination does not disappear. It adapts.
This matters because it dismantles the core premise of the democratisation argument. AI can improve the surface of a manuscript. It cannot change who wrote it, where they are from, or how reviewers respond to those inferences. Therefore, the problem was never purely linguistic but structural, and AI tools are not designed to address structural bias.
The disclosure trap is falling unevenly
A further layer of inequity is emerging around how AI use in manuscript preparation is disclosed. The PNAS analysis of 5.2 million papers found that only 0.1% of more than 75,000 papers published since 2023 disclosed AI use, despite 70% of journals now requiring such disclosure. The ratio of detected AI use to disclosed AI use is approximately 40 to 1. The gap between what is happening and what is being declared is the norm, and the reasons are not hard to understand.
The researchers most likely to be penalised for non-disclosure are the same researchers who turned to AI out of structural necessity in the first place.
For instance, a peer-reviewed guide published in Medical Teacher found that a substantial proportion of authors who do disclose AI use originate from Global South countries, and warns that disclosure policies must be examined for their potential to disproportionately burden multilingual scholars, early-career researchers, and those from institutions without formal AI guidance. Fear of judgment, stigma, and what the authors describe as “AI shaming” are documented barriers to disclosure. Researchers at well-resourced institutions, with access to legal advice and clear institutional policies, can navigate these risks with relative confidence, but those without such support cannot.
The Committee on Publication Ethics (COPE) has warned that if enforcement mechanisms become punitive, they risk reinforcing existing academic inequities, favouring elite institutions that can afford editing services, AI subscriptions, and compliance infrastructure. The researchers most likely to be penalised for non-disclosure are the same researchers who turned to AI out of structural necessity in the first place.
What needs to change
To be clear, this article is not an argument against AI tools or against disclosure. It is an argument that the current framework is being built without adequate attention to who bears its costs.
Publishers need to move beyond one-size-fits-all disclosure requirements and develop tiered, context-sensitive guidance that accounts for the different circumstances of their authors. A researcher at a well-funded institution in a native English-speaking country and a researcher at an under-resourced institution writing in their third language are not in the same position, and disclosure policy should not treat them as if they are.
Funders have a role too. Equitable access to AI tools, and to the training needed to use them effectively and responsibly, should be part of how research capacity is supported in lower-income research environments. The Frontiers whitepaper on AI in peer review, drawing on a survey of 1,645 active researchers, identifies equitable access and structured training as among the most pressing policy priorities for the research community.
Democratisation requires more than access to a tool. It requires dismantling the structures that make the tool necessary in the first place.
And the research community itself needs to reckon with the persistence of language-based bias in peer review. AI has not solved this problem. In some respects, as the evidence on ChatGPT-style phrasing as a new bias cue suggests, it has added a new dimension to it. Journals investing in reviewer training on bias, and in double-blind review processes, will do more for equitable publishing than any AI writing tool.
A level playing field requires more than a better tool
The promise of AI in academic publishing was not unreasonable. The technology does reduce some barriers, and for some researchers in some contexts, it is genuinely useful. But the evidence emerging from large-scale analyses of millions of papers, from peer review data spanning seven years and five continents, and from the researchers navigating these systems in practice, points consistently in the same direction. AI is being adopted fastest by those who face the greatest structural disadvantage. It is not removing the bias those researchers encounter. And the policies being built around it risk adding new burdens to those who are already carrying the most.
Democratisation requires more than access to a tool. It requires dismantling the structures that make the tool necessary in the first place. On that, the evidence is clear, and the research community has not yet caught up.