Authors, reviewers and editors should not be left to endure AI anxiety alone

The arrival of generative artificial intelligence in academic research has produced something more disorienting than a simple ethical dilemma. It has created a situation in which every participant in scholarly publishing is being asked to make judgements that the system itself has not yet learned how to make.

Authors are told to be transparent, reviewers to be vigilant, editors to protect integrity. But the result is not a new culture of clarity. It is a culture of suspicion.

I experience this problem first as an author. Like many researchers, I do not approach AI as either a miracle or a threat. I approach it as a tool, whose boundaries remain strangely unclear. It can help with phrasing, structure, summaries, coding, translation, visualisation, brainstorming and literature mapping. Yet each of these uses occupies a different ethical position. Asking AI to polish a paragraph is not the same as asking it to generate an argument. Using it to produce a chart is not the same as using it to interpret data. Asking it to suggest possible lines of enquiry is not the same as outsourcing the intellectual work of the article.

This is where the anxiety begins. At what point does assistance become authorship? If I use AI to test an interpretation, sharpen a research question or generate alternative ways of presenting data, is that part of my process or a dilution of my contribution? Academic writing has never been the pure expression of a solitary mind. We discuss ideas with colleagues, receive feedback from reviewers, work with copyeditors, attend workshops and use software for analysis. Yet AI unsettles this familiar ecology because it produces language and sometimes appears to produce thought. It can mimic the surface of scholarly reasoning while being indifferent to truth, context and accountability.

Appropriate acknowledgement, therefore, is not as simple as adding the standard sentence now required by many journals. “The author used AI for language editing” may give the appearance of transparency, but it rarely tells readers anything meaningful. Was it used to generate code for visualisation? Assist with translation? Summarise documents? Compare categories? Suggest alternative framings? The reader deserves more precision.

The difficulty is intensified when I sit on the other side of the process. Reviewers are increasingly expected to identify AI-generated work, but they have not been given reliable tools or fair procedures for doing so. AI-detection software is notoriously uncertain. It can produce false accusations, especially against non-native English writers, formulaic academic prose and highly polished writing.

At the same time, experienced reviewers do notice patterns: vacant generalisation, strangely smooth transitions, repetitive conceptual gestures, references that do not quite support the claims, citations that feel decorative rather than necessary, and paragraphs that sound plausible without doing real analytical work.

But noticing is not proving. A reviewer may feel that a paper has been generated, or heavily shaped, by AI, but scholarly judgement cannot rest on a hunch. We cannot reject a paper because it “sounds like ChatGPT”. The risk is very real of creating a new form of academic gatekeeping, whereby certain styles of writing are treated as suspicious because they are too polished, too generic or too distant from an imagined ideal of authentic scholarly voice. The irony is painful: for years, scholars working in English as an additional language were encouraged to improve fluency, clarity and idiomatic control. Now the same fluency may be read as evidence of machine involvement.

The problem becomes still more troubling when reviewers disagree. I have seen cases where a perfectly legitimate paper is suspected of being AI-generated by one reviewer while another sees no such issue. This is not a minor procedural disagreement. It can affect the fate of an author’s work, reputation and confidence.

As an editor, the situation is even more dire. Editors are expected to set AI policies, but it is not easy to come up with one that is sufficiently robust, fair and enforceable. A policy that simply says “AI must be acknowledged” is inadequate; one that bans AI outright is unrealistic. A policy that permits “language editing” but forbids “content generation” sounds clear until one asks where content begins. Is restructuring an argument content? Is suggesting categories content? Is generating a title content? Is translating interview material content? Is producing code content? Academic work is made of many small acts, and AI can now enter almost all of them. Torn between their duties to protect the journal, support the reviewers and treat authors fairly, editors are increasingly walking a tightrope.

The academic world, therefore, should move away from asking “Was AI used?” as if the answer alone determines integrity. The better question is: “What intellectual work does the author remain accountable for?” This shifts attention away from policing style and towards evaluating responsibility.

Authors should be asked to provide a short, specific AI-use statement when relevant: what tool was used, for which task, at what stage – and how was the output checked? Reviewers should not be asked to identify AI by instinct; they should be asked to evaluate argument, evidence, method, citation practice, originality and accountability. Editors should intervene when there is a mismatch between the declared process and the scholarly substance of the article. Disclosure, on all levels, should be normalised without treating every AI use as misconduct.

Most importantly, publishers must stop treating AI governance as an ethical burden to be outsourced to reviewers and editors who are already overburdened and under-rewarded. If AI has made writing and submitting papers easier and misconduct harder to detect, then, all other things being equal, the system will become even slower, more suspicious and less fair. Publishers should provide more than policy templates, reviewer training, appeal procedures and clear editorial guidance across disciplines. They should invest in stronger initial screening, better administrative support, transparent triage procedures and realistic workload models for editorial teams.

They should also confront a question that academia has avoided for too long: is it still defensible to rely so heavily on unpaid editorial and reviewing work? I think not, and so do many others. In the age of AI, academic publishing cannot keep asking scholars to defend the integrity of a system that continues to undervalue their work.

Mai Zaki is professor of linguistics at the American University of Sharjah, United Arab Emirates.