Ghost in the Journal

Academia’s Strange Dance with AI Writing

Even St. Jerome might be tempted by a laptop with ChatGPT (created by ChatGPT 4o - one shot)

In August 2023, editors at the journal Physica Scripta faced an unusual problem. A submitted math paper looked ordinary enough – until a sharp-eyed reader spotted the telltale phrase "Regenerate response" buried in the text. For anyone who has toyed with ChatGPT, that phrase was a smoking gun: the authors had apparently copied-and-pasted text directly from the AI, accidentally including ChatGPT's interface prompt. Caught red-handed, the authors owned up to using the chatbot to draft their manuscript. The result? The paper was swiftly retracted for violating the journal's ethics policy by failing to disclose AI assistance.

Depending on whom you ask, this episode was either a minor comedy of errors – a "storm in a teacup" – or an ominous sign of how AI is infiltrating academia.

Since OpenAI's ChatGPT burst onto the scene in late 2022, academics across disciplines have been grappling with a burning question: When, if ever, is it okay to use AI in scholarly writing? What started as scattered experiments (and a few sneaky attempts like the one above) has morphed into an urgent debate. Journal editors, faced with a flood of AI-assisted writing, are hastily updating author guidelines. Researchers are unsure whether using AI makes them efficient innovators or integrity violators. And everyone is trying to figure out the new rules of a game that didn't even exist a couple of years ago.

Grab your quill – or your laptop – and let's explore this strange new world where a "co-author" might be a clever algorithm.

Journals vs. ChatGPT

Not long after ChatGPT went public, journal editors realised they needed to say something about it. By early 2023, a few high-profile incidents – like an Oncoscience paper listing "ChatGPT" as a co-author, or a handful of submitted manuscripts obviously written by AI – set off alarm bells. Journals and publishers scrambled to issue policies. The result was a flurry of editorial notes, guidelines and updates to "instructions for authors." By late 2023, virtually every major academic publisher had weighed in – but not all in the same way.

Let's start with the basics. Almost everyone agrees on one thing: AI can't be an author. The idea of crediting ChatGPT as a co-author on a scholarly paper provoked near-universal rejection. The Committee on Publication Ethics (COPE), an influential publishing watchdog, declared that "AI tools cannot be listed as an author of a paper" because they can't take responsibility for the work. This became a new first principle in publication ethics.

But beyond the "no AI authors" rule, policies began to diverge. Journals had to wrestle with a harder question: Can human authors use AI as a tool in writing, and if so, under what conditions? Here, the spectrum ranged from outright bans to cautious acceptance with strings attached.

At the strict end were journals like Science and its sister publications. In a Feb 2023 editorial, Science announced a bold policy: no text generated from AI is permitted in their papers, period. The editors wrote, "Text generated from AI, machine learning, or similar algorithmic tools cannot be used in papers published in Science journals... In addition, an AI program cannot be an author. A violation of this policy constitutes scientific misconduct." In other words, if you try to slip some ChatGPT-written paragraphs into a Science article, you're potentially committing the same sin as faking data.

Nature, while not going quite as far as banning all AI-generated text, likewise prohibited AI co-authors and urged caution. Nature's editors set "ground rules" in early 2023: authors could use tools like ChatGPT to improve readability or sift information, but they had to disclose it and AI would not be accepted as an author.

Many other journals and publishers took a more moderate stance: AI tools are okay as long as you're transparent about using them. This quickly became the emerging norm. For instance, the JAMA (Journal of the American Medical Association) network of journals doesn't forbid AI-assisted text, but it requires authors to clearly describe any AI-generated content and to cite the tool (mentioning the model and version) in the methods or acknowledgments.

The International Committee of Medical Journal Editors (ICMJE), whose guidelines influence thousands of journals, updated their recommendations in mid-2023 along similar lines. ICMJE now insists that authors disclose AI use in either the cover letter, the acknowledgments, or methods – with specifics on where to put it depending on what was done. If AI helped write or edit the text, say so in the acknowledgments; if it helped analyse data or generate figures, mention it in methods.

The Great Policy Divergence

A comprehensive bibliometric study published in BMJ (January 2024) examined the top 100 academic publishers and top 100 high-impact journals to see who had AI-in-writing policies. The findings show a patchwork of rules:

• Many top journals have policies, but many publishers still don't. By late 2023, about 87% of the elite journals surveyed had some guidance on using generative AI (GAI). But only 24% of the largest publishers had issued any AI policy, meaning a lot of smaller journals were still flying blind.

• All those policies agree: no AI authors. 98% of journals with guidelines explicitly banned listing an AI tool as an author. That near-unanimity checks out – apparently only a fringe 2-4% hadn't explicitly said "no" yet, possibly because they felt it was obvious. (If any journal editors are reading this, and you haven't stated this clearly in your instructions to authors, maybe put it in bold now: "Don't list ChatGPT as an author. They didn't even go to school.")

• Almost nobody outright banned using AI to help write a paper. Only 1% of journals in that study forbade using GAI at all in manuscript preparation. In other words, the vast majority are not following the ultra-strict Science stance.

• Disclosure is usually required, but the where and how vary. Of the journals that allow AI assistance, only 43% gave specific instructions on how to disclose it, whereas 75% of publishers (the ones that had policies) did. Some journals said to put the info in the Methods or Acknowledgments; others wanted it in a separate "AI Contributions" section or even the cover letter to editors. A few created a new field in the submission forms. There was no single standard.

What kind of AI use is okay? Many policies were vague or differed in emphasis. Some journals simply state that AI tools may be used for improving writing (grammar, clarity) but must not replace "critical thought" or the author's own analysis. Others go further in listing forbidden uses. For example, the publisher Edward Elgar explicitly warns that using AI for "analysis of data or the development of written arguments" in a paper is "not permitted." That draws a bright line: AI can check your grammar, but it shouldn't be developing your argument or doing your data analysis for you.

Interestingly, even within the same publishing house, policies could conflict. The BMJ analysis found a dozen cases where a journal's posted policy didn't align with its parent publisher's rule. The result, circa late 2023, was a crazy-quilt of AI policies: nearly every top journal had something to say, but the specifics ranged from terse ("AI use must be disclosed") to thorough multi-point checklists.

What about fields outside the hard sciences and biomedicine? Top humanities and social science journals are often published by the same big houses (Oxford, Cambridge, Taylor & Francis, Sage, etc.) who have rolled out AI guidelines across their portfolios. For example, SAGE Publications (which covers social sciences and humanities) now tells authors: If you use an AI tool that produces content (text, images, references, etc.), you must disclose this use.

One might expect humanities scholars to be more relaxed about AI help with writing (since the "data" is often the prose itself), but actually the idea of undisclosed AI ghostwriting arguably cuts even closer to the bone in fields that value originality of expression. For instance, the Modern Language Association guidelines (for language and literature fields) emphasise that any AI-generated text should be "treated as a source" – i.e. cited or acknowledged appropriately – to avoid plagiarism.

Acceptable Help vs. "Outsourcing Thought"

It's one thing to say "disclose AI use," but another to agree on what kinds of AI assistance are acceptable. This is where we get into grey areas. Is using ChatGPT to fix your grammar like using Grammarly – a benign editorial aid? Is asking it to paraphrase a clunky sentence okay? What about generating a summary of related work? Or crafting an entire first draft based on your outline?

From surveying the new policies and editorial statements, we can generalise a bit: Clerical or mechanical uses of AI are broadly accepted; creative or analytical uses are viewed with suspicion. In practice, that often breaks down as follows:

Okay: spelling and grammar checks, smoothing out language, translating text, formatting references, checking for consistency. These are seen as extensions of existing tools or the kind of help a paid editor might provide. Many non-native English speaking academics, for instance, welcome AI tools to help polish writing – and journals are generally fine with this kind of help (as long as it's disclosed).

Maybe okay (with disclosure): using AI to summarise background literature or suggest possible references, using it to generate ideas or questions as a form of brainstorming, or having it rephrase a paragraph for clarity. These venture beyond mere copyediting, but many policies implicitly allow this as long as the author remains in control.

Not okay (usually): letting AI drive the original analysis, or using AI-generated text or images as if they were your own creative work. For instance, feeding your data into an AI and having it write the Results section, or asking ChatGPT to "write the discussion section for me" – most journals would frown on this, even if you disclosed it. The concern here is that you're effectively outsourcing the intellectual work that peer reviewers expect you to have done.

Definitely not okay: hiding AI use, or attempting to pass off AI-generated content as if a human created it (without disclosure). This is viewed as a form of academic misconduct.

Are Authors Using AI?

Policies are one thing, but what's actually happening on the ground? It's like the speed limit – the sign might say 60, but are people going 75 when no one's looking? In the world of academic writing, are researchers quietly using ChatGPT and its kin to churn out text? And if so, are they disclosing this to journals?

We have some intriguing clues. By late 2023, one Nature survey of 1,600 researchers worldwide found that about 30% of scientists said they had already used generative AI to help write papers. Thirty percent! And that was still relatively early in adoption.

But how many are admitting it in their published work? There the number is much smaller. Throughout 2023, journals recorded only a trickle of disclosures in the acknowledgments. For example, JAMA in mid-2023 said that only a handful of manuscripts had come in with an AI usage statement, despite everyone talking about ChatGPT. It seems many authors are either not using it heavily enough to mention, or if they are, they're hesitant to declare it – perhaps fearing stigma or rejection.

Enter the detectives. A few computer-savvy researchers took it upon themselves to detect AI-written text in publications. One such effort by data scientist Dr. Andrew C. Gray made headlines in late 2023. Gray developed an approach to scan millions of papers for linguistic fingerprints of GPT-like text. Specifically, he looked for overuse of certain adjectives and adverbs that language models tend to favor more than humans do (words like "innovative", "meticulous", "commendable", "methodically", etc.).

His analysis of papers from 2019–2023 revealed a striking uptick in these AI-associated words in 2023. Gray estimated that perhaps on the order of 60,000 to 85,000 research papers published in 2023 showed signs of LLM (large language model) involvement. That's roughly 1–2% of papers in his sample.

Another study tried a more direct test. Desaire et al. (2023) took a sample of published scientific papers and used an AI detection algorithm on the introductions. They reported finding that perhaps 1–3% of those introductions showed evidence of ChatGPT involvement. They provocatively titled the paper "Almost Nobody Is Using ChatGPT to Write Academic Science Papers (Yet)", highlighting that (at least in mid-2023) the vast majority of authors were not handing their writing to ChatGPT – or if they were, they weren't leaving obvious traces.

Meanwhile, Retraction Watch (the watchdog blog) started a public list of papers and even peer reviews suspected to be AI-written. By late 2023 they had over a hundred entries. Some were caught because authors left in weird phrases like the infamous "As an AI language model, I…" at the start of a sentence – clearly a snippet of a ChatGPT reply that the author forgot to delete. Others were flagged by peer reviewers who noticed inconsistent writing style or unverifiable references (AI sometimes makes up fake citations).

So, the empirical picture is this: AI use among authors has grown very quickly, but formal disclosure lags behind actual use. In 2023, probably only a small fraction of the papers that had AI assistance explicitly said so, either out of forgetfulness, fear, or the authors not even considering an AI tool "important" enough to mention (which is risky). But the direction is toward greater transparency. As policies clarify and the stigma lessens, we may see authors being more upfront.

Ethics, Integrity and the Soul of Academic Writing

Let's step back from the stats and ask: Why is this such a big deal? What are the philosophical and ethical arguments swirling around AI in scholarly writing? After all, technology has always been part of research – from calculators to spellcheck to statistical software. But something about AI-generated text has touched a nerve in academia.

The Case for AI as Just Another Tool (The Optimists)

Many authors and some technologist commentators argue that using AI in writing is an evolution of the same trajectory that gave us grammar checkers and citation managers. In this view, generative AI can handle drudge work (like rephrasing sentences or summarizing findings in plain language) which frees up human researchers for the creative and analytical heavy lifting.

Supporters often use analogies: Would you insist a mathematician do long division by hand instead of using a computer? If not, then why freak out if a historian uses ChatGPT to condense a verbose paragraph? As long as the scholar verifies and curates the output, they remain the intellectual author.

From this perspective, refusing to allow any AI assistance might even be seen as Luddite. If a non-native English speaker can make their paper clearer with AI, isn't that a win for knowledge dissemination? And if a busy clinician-researcher can save two hours of editing by having GPT suggest a structure for the introduction, couldn't that time be better spent designing the next experiment?

The Case for Caution and Integrity (The Sceptics)

On the other side, journal editors, ethicists, and many scholars emphasize the core values of academic writing that they fear AI use might erode. David Resnik, a bioethicist and journal editor, listed principles like honesty, transparency, accountability, rigor, and objectivity, and noted that AI challenges these because AI cannot be held accountable or be fully transparent about how it generates text.

The crux of this view is responsibility: when an academic puts forth a piece of writing, they are implicitly vouching for its accuracy and integrity. If an AI wrote a part of it, can the author truly vouch for it in the same way? Given that current AIs often produce plausible-sounding but incorrect or fabricated information, an author might inadvertently include errors or even nonsense if they trust the AI too much.

Moreover, even if the facts are correct, there's the value of original thought. Especially in fields like humanities or theoretical work, the writing isn't just packaging for the ideas – it is part of the intellectual contribution. Style, argumentation, and narrative flow are part of how scholars make their case. If those were machine-generated, is the work less original?

Academic integrity experts also invoke slippery slope arguments: Today it's a paragraph, tomorrow it's a full paper. Where do we draw the line between acceptable assistance and academic dishonesty?

Middle Ground Perspectives

Between the boosters and the alarmists, many voices call for a balanced approach. They acknowledge the benefits of AI tools – improving efficiency, aiding those less adept at academic English, even sparking creativity by offering suggestions – but insist on preserving human intellectual ownership and rigorous verification.

A telling quote from a journal (Patterns) author, James Zou Wang, encapsulates this balanced ethos: "When a powerful tool like ChatGPT is available, it is neither beneficial nor feasible to prevent its use. However, the scientific community must remain vigilant about the ethical implications and long-term impacts."

Tales from the Frontlines

We already recounted the Physica Scripta retraction, which is perhaps the most famous case so far of "AI misuse" in publishing. But it's not the only notable example. Let's look at a few case studies that have made editors sweat and put a spotlight on the need for clear policies:

The Accidental ChatGPT Manifesto: In mid-2023, a scholarly article (ironically about AI in education) included a bizarre sentence in the first person: "As an AI language model, I cannot access the internet, so I do not have real-time data or browsing capabilities." Imagine a research paper suddenly lapsing into a ChatGPT monologue about itself! Obviously, the human authors had at some point asked ChatGPT a question and got this boilerplate response, and somehow a piece of that made it into the paper's text.

Regenerate Respone (sic): Another case involved a paper on human-computer interaction that ended a paragraph with the phrase "...and these modalities regenerate response." It seems the authors might have invoked the "Regenerate response" button on ChatGPT, and part of that slipped into the text (though here it was missing a quote or context, making it a bit nonsensical).

The ChatGPT Urology Exam Paper: In early 2023, a team submitted a paper titled "ChatGPT Performs Poorly on XYZ Exam" (paraphrasing here) which analyzed how well ChatGPT answered questions from a standardised test in urology. This paper, published in a urology journal, was later retracted – not because using ChatGPT was itself wrong (the whole point was to test it), but because they failed to properly disclose ChatGPT's role and some content issues.

Peer Reviews by AI: A tangent but interesting – it's not just authorship where AI crept in; some reviewers have used ChatGPT to write peer review reports! A few journals discovered reviews that contained the same "AI tells on itself" phrases or a generic tone that made them suspect an AI wrote the referee's comments.

Each of these cases reinforces that transparency is key. The primary reason these became scandals is because the AI involvement was hidden (until a mistake exposed it). The actual content in some cases might not have been bad – e.g., maybe ChatGPT's summary was fine – but the trust was breached.

Emerging Best Practices

After a tumultuous year of experimentation, debate, and a few snafus, the dust is beginning to settle on how to handle AI in academic writing. While not everyone agrees on every point, there are signs of an emerging consensus:

1. Always Disclose AI Assistance: This is rule number one, enshrined in essentially all guidelines. If an author used AI in any part of the manuscript creation, they should state so in the paper. The disclosure should mention which tool, which version if relevant, and what was done.

2. Human Accountability is Non-Negotiable: No matter how much AI was used, the human authors must take responsibility for the content. Several policies ask authors to include a statement akin to: "After using the AI tool, we reviewed and edited the content as necessary and take full responsibility for it."

3. No AI in the Author List: It's become almost humorous how every policy reiterates this, but they want to be crystal clear: authors must be real humans. If a submission lists an AI as an author, it will be rejected or require revision.

4. Standardised Reporting: A push is underway to standardize how AI use is reported in papers. Some have proposed adding a simple check box in submission systems: "Did you use generative AI in preparing this paper? If yes, provide details."

5. Shifting Culture – from Stigma to Standard Procedure: Perhaps the most important "best practice" isn't a rule but a cultural shift. By treating AI use sort of like we treat statistical software use, we can integrate it without moral panic.

Rethinking Authorship in the Age of AI

If there's one thing the saga of AI in academic writing has taught us, it's that academia can adapt – albeit clumsily at first – to new technology without losing its core principles. We watched journal policies evolve at breakneck speed since January 2023: from initially no mention of AI at all, to knee-jerk bans in some cases, to now a more layered set of guidelines.

Throughout this journey, a central tension emerged: Does using AI diminish the intellectual rigor of research (an "outsourcing" of thinking), or is it just the next step in technological support (like using better software tools)? The answer isn't an either-or – it's both, depending on how AI is used. Use it as a crutch to avoid thinking, and yes, it's an intellectual shortcut that cheapens the work. Use it as a lever to amplify your thinking, and it could improve the clarity and reach of your scholarship.

Academia has long held certain rituals and values around writing: the lonely scientist pouring over drafts late at night, the scholar wrestling with words to articulate a new idea. There's almost a romantic notion of the academic author that AI seems to encroach upon. But consider this: the written form of academic output has changed many times. Now, perhaps, AI tools can be seen as the new lingua franca helpers – assisting those who struggle to express themselves to do so more effectively.

The key will be preserving trust. Science and scholarship run on trust: readers trust that authors are presenting work honestly, authors trust that peers reviewing are doing so conscientiously, and so forth. AI doesn't have to break that trust if everyone is open about its use.

To be sure, challenges remain. Plagiarism and fraud aren't vanishing just because we have policies – unscrupulous authors might try even harder to use AI covertly. Detection tools will improve, but it will become a cat-and-mouse game that will never (most likely) be 'won'. On the flip side, AI will keep improving too: maybe one day it will be capable of genuine insight and then the philosophical question of credit really explodes. That day may well be coming sooner than we think, but till then..

So, in concluding, let's return to the notion of the "ghost in the journal." It turned out the ghost – AI – doesn't have to haunt us if we shine a light on it. By acknowledging the ghost's presence, we take away its power to spook. We've begun to see AI not as a ghoul threatening academic integrity, but as a tool that, when used with eyes open, can sit quietly in the background of the writing process. The real "spirit" of the work remains the human scholar's curiosity, reasoning and insight. Journals, for all their stuffy guidelines and formats, are ultimately vessels for human knowledge. AI can help shape those vessels, but it's up to us to fill them with meaning.