AI editing tools make more corrections but reduce writing quality

By Dr. Priyom Bose, Ph.D.Reviewed by Lauren HardakerFeb 10 2026

AI tools promise faster, cheaper manuscript editing for global researchers, but new evidence shows they may introduce hidden risks that could reshape equity in scientific publishing.

Study: Does ChatGPT enhance equity for global health publications? Copyediting by ChatGPT compared to Grammarly and a human editor. Image credit: Maxx-Studio/Shutterstock.com

A recent PLOS ONE study, based on a small, preliminary case comparison, evaluated the copyediting quality of U-M GPT, Grammarly, and a human editor on two draft research papers. ChatGPT made three times as many corrections as a human editor but was less precise, raising questions about whether artificial intelligence (AI) can truly enhance equity in academic publishing and under what conditions it may help or hinder researchers.

Language Barrier in Academic Publishing

English dominates academic publishing, yet most researchers worldwide speak it as a second language. Non-native English speakers spend up to 51% more time writing papers than native speakers, yet still face rejection due to grammar issues. Professional editing could help, but it is prohibitively expensive.

Often, professional editing services can cost nearly half an early-career researcher's annual salary in countries like Colombia. This could be why researchers from non-English-speaking countries remain severely underrepresented in the academic literature and why language has been described as a structural source of power imbalance in global science.

AI tools have been perceived to bridge this disparity. Although basic grammar checkers, including Microsoft Word, Google Docs, and Grammarly, have been available for years, large language models (LLMs), such as ChatGPT, offer new possibilities for more sophisticated editing support. However, it is imperative to assess the efficacy and accuracy of these models, as well as the new barriers they may introduce, including technical skill requirements and ethical concerns.

Early research yielded mixed results: one study found that ChatGPT provided helpful grammar corrections for non-native English writers, while another contradicted this, with some authors reporting that AI edits sometimes “over-polish” writing without improving clarity.

Copyediting Potential of LLMs Compared to Grammarly and a Human Editor

The current study defined copyediting as correcting grammar, spelling, syntax, and punctuation; ensuring proper terminology and conventions; checking structure, organization, and clarity; and improving readability, flow, and style. It compared corrections made by U-M GPT, a secure, University of Michigan-hosted generative AI tool, with those made by Grammarly and a human copyeditor on draft manuscripts written by Ugandan sexual and reproductive health researchers.

Two Ugandan researchers provided written approval for the use of their draft manuscripts on reproductive health. Both papers were later published in peer-reviewed journals. The key copyediting goal was to help Ugandan researchers get accepted for peer review.

The authors evaluated the introduction, methods, results, and discussion sections of both papers, as well as selected tables. In total, only eight paragraphs and two tables were analyzed, and a single comprehensive prompt was used to test real-world feasibility. A sensitivity analysis was conducted to identify potential limitations of this approach.

The free web-based version of Grammarly was used, which does not sell user data and restricts access to uploaded text to users only. Grammarly corrections were generated for an expert audience, a formal tone, and a general writing domain. Texts were submitted to U-M GPT one paragraph or table at a time, and the corrections generated were analyzed. A professional copyeditor from the CIRHT/PREPSS training program provided human corrections with access to the full manuscripts, unlike the AI tools that evaluated isolated excerpts.

The researchers classified corrections from the three editors, using fewer sentence-level categories than previous studies. They noted that U-M GPT sometimes removed key information, such as references, which they tracked separately because these deletions posed risks to meaning and accuracy. The authors also acknowledged that classification of edits was performed by an internal research team, which may have unintentionally introduced bias in the assessment of edit quality.

AI Tools Were Quick but Lacked Human Editing Efficiency

The human copyeditor required 3.75 and 4 hours to edit Papers 1 and 2, respectively. In contrast, U-M GPT generated corrections nearly instantaneously. However, creating a track changes document for each paper with U-M GPT required approximately 30 minutes. Grammarly also generated corrections within seconds, prompting users to accept or reject changes one at a time, a process that took approximately 5 minutes.

Only the human copyeditor could correct both text and tables. U-M GPT suggested pasting table content into the chat, but the researchers found this difficult and impractical. Grammarly does not allow table uploads.

It must be noted that U-M GPT prompted about three times as many corrections as the human editor and roughly ten times as many as Grammarly. The scope of corrections varied by editor. For instance, Grammarly corrected only spelling and grammar, while U-M GPT flagged errors in spelling, grammar, punctuation, spacing, and capitalization. In contrast, the human editor corrected grammar, punctuation, spacing, and capitalization, and flagged unclear text rather than attempting to rewrite passages whose meaning was ambiguous, allowing authors to clarify intent themselves.

For readability, U-M GPT yielded the most changes but showed the poorest judgment. Of its 83 revisions, only 61% actually improved the text, while 14% made it worse and 24% had no effect. Grammarly made just five readability corrections, two provided improvements, and three had no real impact. The human editor made 21 corrections: 90% improved the text, one was neutral, and one made it worse.

U-M GPT also deleted important content from one paper, including citations and a table reference, raising concerns that authors, particularly those less confident in English writing, might accept harmful edits uncritically.

Conclusions

The current study highlighted a fundamental difficulty in using AI for academic editing, particularly regarding quality. While U-M GPT generated three times more corrections than a human editor, only 61% actually improved the text. Even though this technology is fast and inexpensive, it introduces risks of deleting key information, making questionable revisions, and potentially blocking content.

The promise of AI-enhanced equity in academic publishing remains uncertain. In addition to quality concerns, the study emphasized issues of data privacy, the environmental costs of large language models, and the need for prompt-engineering skills, meaning the ability to carefully design AI instructions that may themselves create new inequities.

As these tools evolve, researchers need larger studies across diverse manuscript types to understand when AI editing helps and when it hinders. For now, authors using AI editing tools should proceed with caution, because the goal is not rapid editing but access to editing that genuinely improves research communication while preserving author voice, meaning, and scholarly independence.

Download your PDF copy now!