Researchers Divided on AI's Role in Scientific Publishing, Survey Finds

A new global survey of more than 5,000 academics reveals sharp divisions over how generative AI should be used in writing and reviewing scientific papers. While researchers broadly accept AI tools for tasks such as editing and translating manuscripts, opinions diverge markedly when it comes to more substantive applications like generating text or contributing to peer reviews. The results highlight a complex landscape in which efficiency and accessibility clash with concerns about ethics, accuracy, and the integrity of the scholarly record.

Historical context: AI in academic writing has evolved rapidly over the past decade. Early adopters treated AI as a language aid and consistency booster, while skeptics warned of hidden biases, plagiarism risks, and the potential erosion of critical thinking. The latest poll reflects a maturation of this debate, moving beyond simple acceptance or rejection toward nuanced guidance about disclosure, provenance, and accountability. As universities and funding bodies increasingly emphasize reproducibility and transparency, researchers are weighing not only what AI can do, but how its involvement should be disclosed and reviewed.

Scope and methodology: Respondents hailed from a diverse mix of fields, career stages, and regions, with substantial representation from the United States, India, and Germany. The survey probed three core domains: the use of AI in manuscript editing and translation, the generation of original text for manuscripts, and the role of AI in the peer-review process. It also captured attitudes toward disclosure, potential benefits, and perceived risks. Across disciplines, responses show a broad willingness to adopt AI tools for routine editing tasks, coupled with caution about deeper textual authorship and evaluative functions.

Key findings: editing, translation, and abstract drafting

• Editing and translation: More than 90% of participants found AI acceptable for editing and translating research papers. This high level of comfort reflects the practical reality that AI can streamline language polishing, reduce turnaround times, and improve accessibility for non-native English speakers. The consensus here is anchored in perceived reliability, provided users review outputs and verify factual content.
• Text generation: Acceptance for generating text within a paper varied widely. About 65% deemed it acceptable to use AI for generating text in all or part of a manuscript, while roughly one-third opposed this practice outright. Opinions were more favorable for drafting abstracts than for methods, results, or discussion sections, where concerns about originality, accuracy, and critical interpretation were greatest.

Key findings: peer review and disclosure

• Peer-review assistance: A majority—over 60%—rejected the notion of using AI to generate an initial peer-review report, citing privacy concerns and the risk of biased or erroneous conclusions. In contrast, 57% supported AI assistance for specific tasks within peer review, such as answering factual questions about a manuscript or summarizing the content, indicating a preference for AI as a supplementary tool rather than a primary evaluator.
• Disclosure and provenance: Across respondents, there was broad agreement that significant AI involvement should be disclosed. However, details on how to disclose remain contested. Some favored simple acknowledgment in the methods or acknowledgments sections; others argued for more explicit statements about the extent and nature of AI contributions, akin to data provenance. The lack of a standardized disclosure framework mirrors broader discussions in science about reproducibility and transparency.

Actual usage patterns: real-world adoption

• Overall adoption: Despite favorable views of AI for editing and assistance, actual usage remains relatively modest. About 65% of respondents reported never employing AI in any publishing scenario. Among those who have used AI tools, 28% used them for editing, 8% for drafting a first version, and only 4% for initial peer reviews. Early-career researchers and respondents from non-English-speaking countries were more likely to have tried AI, suggesting a trend toward broader experimentation among newer cohorts and more linguistically diverse contexts.
• Disclosure practices: Among users, disclosure is common but inconsistent. Many respondents who used AI failed to disclose its use in their manuscripts, raising questions about accountability and the integrity of the scholarly record. The gap between favorable opinions and actual practice underscores a need for clearer guidelines and practical disclosure norms.

Regional and disciplinary patterns: comparisons and contrasts

• English-language publication hubs: In regions where English is predominant in academia, editors and authors reported higher comfort with AI-assisted editing, reflecting existing infrastructure, training, and access to language-support tools. Yet concerns persist about overreliance and the potential depletion of language proficiency and critical thinking skills.
• Non-English-speaking regions: Researchers from non-English-speaking backgrounds often viewed AI as a leveling tool that can mitigate language barriers. They were also more likely to experiment with AI in the drafting process, albeit with careful consideration of accuracy and originality. This dynamic emphasizes AI's potential to promote inclusivity in global science, while also highlighting the risk of uneven adoption across institutions with differing access to resources.
• Field variability: Acceptance of AI-generated text varies by discipline, with more uniform willingness in fields that emphasize rapid communication and broad dissemination of findings, and greater caution in areas where nuanced interpretation and methodological rigor are paramount. Across the board, however, there is a shared concern about maintaining scientific integrity and avoiding misattribution of authorship or ideas.

Economic impact: efficiency, costs, and equity

• Time savings and productivity: AI-assisted editing and translation can expedite manuscript preparation, enabling researchers to allocate more time to data collection, analysis, or collaboration. This efficiency gain is particularly meaningful for researchers in regions with limited language support services. The productivity boost may also accelerate the pace of scientific discovery and collaboration across borders.
• Cost considerations: Access to AI tools varies, with some institutions providing subscriptions or on-site capabilities, while others impose higher barriers through licensing fees. The economic dimension becomes a policy question for universities and research funders, who must balance investment in AI infrastructure with other critical needs, such as data stewardship, reproducibility initiatives, and researcher training.
• Equity implications: By lowering language barriers, AI can democratize participation in international science. Conversely, unequal access to AI tools could exacerbate existing disparities, privileging well-funded labs and institutions. A thoughtful approach to licensing, open-source alternatives, and targeted training can help ensure broader, fairer access.

Historical context and regional comparisons: lessons from the past

• Past tech adoptions in academia show a consistent pattern: tools that reduce friction tend to be adopted quickly in routine tasks, while deeper, integrity-sensitive uses are approached with caution. The current survey aligns with this trajectory, suggesting that AI will continue to permeate editorial workflows and support roles, with stricter governance around original content generation and evaluative tasks.
• Regional comparisons reveal how policy environments shape adoption. Countries with stronger research integrity frameworks and clearer disclosure norms tend to foster more disciplined use of AI in publishing. Where guidelines are evolving, researchers may experiment more freely yet face uncertainty about expectations, leading to inconsistent practices.

Public reaction and trust: the human element

• Perceptions of ethics and trust: Open responses from participants illustrate a spectrum of views. Some see AI as an everyday tool, akin to calculators or grammar-checkers, requiring minimal disclosure as its use becomes normalized. Others label AI-assisted practices as "pathetic cheating and fraud," warning of plagiarism risks, false citations, and a potential erosion of learning and critical thinking. The tension between convenience and trust is at the heart of ongoing debates about AI in science.
• Environmental and societal costs: Critics emphasize the energy consumption associated with large AI models and the broader environmental footprint of running powerful computing infrastructure. Proponents counter that responsible use and optimization can mitigate these impacts, while enabling researchers to focus more on novel ideas and collaborative work. The conversation increasingly includes sustainability considerations alongside scientific quality.

Policy implications: shaping guidelines for the future

• Disclosure standards: Most respondents agree on the need for disclosure when AI contributes significantly to a manuscript or its evaluation. The challenge lies in standardizing the language and location of disclosures to ensure consistency across journals, publishers, and disciplines. A practical approach may involve a dedicated section in manuscripts that details AI contributions and the specific tasks performed.
• Ethical frameworks and governance: The survey underscores demand for clearer ethical guidelines. Institutions, publishers, and funding agencies are likely to develop or refine policies that address authorship integrity, accountability, and the boundaries of AI-assisted work. Transparent governance structures can help maintain trust in the scientific record while preserving opportunities for innovation.
• Training and education: As AI tools become more prevalent, training programs for researchers at all career stages will be essential. This includes instruction on when AI is appropriate, how to verify outputs, and how to document AI involvement in research workflows. Emphasis on critical thinking and manual validation remains crucial.

Conclusion: a landscape of cautious optimism

The survey findings reveal a nuanced landscape in which AI is embraced for routine editorial tasks but approached with caution for substantive writing and evaluative functions. The widespread willingness to disclose significant AI involvement signals a collective commitment to transparency, even as specifics about disclosure mechanisms remain unsettled. Economic and regional dynamics further shape adoption, with accessibility and policy clarity likely to drive broader acceptance in the coming years.

As AI technologies continue to evolve, academia faces a critical question: how can tools that enhance efficiency coexist with the enduring standards of rigor, reproducibility, and ethical integrity? The path forward will require collaborative policymaking among researchers, institutions, publishers, and funders to craft practical guidelines that preserve trust in scientific publishing while unlocking the potential of AI to democratize access to knowledge, accelerate discovery, and strengthen the global research enterprise.