‘Don't look at the Sun!’ instructed the radio show host.
‘Don't look at the Sun!’ echoed his co-host.
Yesterday's solar eclipse, visible from much of Western Europe, was the first in 27 years (and a day). The last time there was a total solar eclipse in the British Isles, I was a fresh faced 29-year old. My partner and I had driven to Portland Bill on the Dorset coast, as close to 100% totality as was possible without getting wet. The smash hit of that summer was Rendez Vu by Basement Jaxx and every time I hear that track, it takes me back to August 11th 1999.
We are right now living through another sort of eclipse, a technological one, in which old practices are rapidly being erased by artificial intelligence (AI) with particular impacts on medical communications and scientific publishing.
Just a day before the solar eclipse, the anniversary of the 1999 one, the Journal of the American Medical Association (JAMA) released an update to its guidance for authors on the use of AI in preparation of manuscripts [1]. It’s well worth reviewing the updated guidance and editorial [2] as the requirements laid out by top journals like JAMA are a good temperature gauge for the broader medical publishing environment.
Recent studies suggest that most leading medical journals provide guidance on authors’ use of generative AI. A 2025 study found that 95% of the 200 highest-ranked medical journals and 86.7% of a random sample of non-top-ranked medical journals provided guidance for authors on generative AI [3]. Led by the 2026 update to the International Committee of Medical Journal Editors recommendations on AI use [4], medical journals are increasingly converging on more standardised AI disclosure requirements, including the tool, purpose and extent of use. Looking at academic journals in general, rather than medical journals specifically, approximately 70% of journals have adopted AI policies, predominantly policies requiring disclosure of AI use [5]. Based on this evidence medical journals (particularly high-ranking ones) appear to have been among the more proactive sectors of academic publishing in developing generative-AI guidance and requiring disclosure from authors. This might be a reflection of early concerns about perceived threats of AI use to research integrity, particularly in a sphere where a mature framework of ethical and reporting standards has been well established to safe-guard accuracy, patient safety and research integrity. Despite this, the rate of disclosure of AI use by manuscript authors remains very low [6,7]. As Tim Hardman, our Managing Director, says: Disclosure only works if people disclose!
The JAMA guidance update starts out with an affirmative statement of AI’s potential, stating that they “…embrace AI as a tool that can be helpful to researchers and authors by reducing the burdens of some tasks and improving efficiency in some contexts.” While the journal recognizes the ‘rapidly evolving and extraordinary influence of AI on research and scientific publishing’ this comes with a variety of ‘threats and concerns’.
As expected, authors remain responsible for content accuracy, however it was generated. To this end, authors must review and confirm the accuracy of all AI-generated content. Critically any use of AI must be disclosed in the Acknowledgements section. Areas where AI use is permitted include study research, preparation of the manuscript, conducting literature searches, language translation, and formatting data or graphic visualisation of data. Each instance of AI use must be fully described in the Methods or Acknowledgements sections, as appropriate. Use of AI to check grammar and spelling does not require disclosure. In contrast use of AI for writing opinion pieces, letters, comments, conducting peer review and creating clinical images is prohibited.
And then there's the references! JAMA discourages use of AI tools to build bibliographic references, stating that ‘AI tools may cause problems with the accuracy and integrity of cited references’ and instead encourages use of standard reference managers. Talk about an understatement! As I have written before (8, 9, 10) the propensity of AI platforms to create fictitious references has achieved considerable notoriety. AI platforms invent citations that look completely real with authentic author names, real-sounding journal titles, real-looking misappropriated DOI numbers and they place papers that were never written in reference lists next to a dozen papers that were formatted identically, styled identically and impossible to tell apart by the casual gaze. Time-pressured reviewers are unlikely to fact-check every claim in a paper, every table, every method and may wave through a citation that simply doesn't exist, because it has the appearance of authenticity. Get this wrong and your reputation may be toast!
Taken in hand with more conventional approaches, I often use AI to help me identify references, but only in cases that require a deeper level of information than is achievable through regular literature search engines. It is rather like walking through a minefield. I am not exaggerating when I say that in almost every case the references I am provided with are incorrect, or worse are hybrids of authentic references and hallucinated AI slop. The first time I experienced this, I was given a reference which had authentic author names from one paper, the authentic title of another related paper and journal issue number, date and page numbers from yet another authentic reference. Only by going to each journal’s online archives, was I able to resolve the mystery. Suffice to say, I’ve learnt that double checking absolutely every AI-generated reference is critical and when errors have occurred, fixing them can take a lot longer than one may have bargained for.
I often find that interrogating the AI platform about a specific reference can result in it retracting it’s previous recommendations and coming up with ‘better’ options, providing explanations like ‘I have checked the original papers and PubMed records, and I would suggest a slightly more careful approach than the one I gave you previously.’ I have to confess that sometimes I feel a strange sense of pity for the voice behind the platform, not least when it replies with a cringy confession about how it got this wrong.
Another area to watch carefully is reference identifiers like PMID or DOI numbers. Again, in my experience these are almost routinely incorrect. Finding correct numbers using the information provided by your AI platform can be frustrating and very time-consuming. Sometimes the reference just has to be removed completely.
In a recent confessional from the AI platform I was using, it told me:
‘When generating exact, long-string identifiers—such as DOIs or PMIDs—AI models predict plausible-sounding alphanumeric sequences rather than pulling live, verified keys from an active database query. Even when referencing real, high-impact clinical trials, the AI-generated numeric strings frequently contain transposed digits or fabricated suffix patterns that fail to resolve on PubMed or DOI.org.’
Straight from the horse’s mouth!
It’s just as JAMA says, ‘…we advise authors to not use AI to generate, format, or otherwise manage references.’
Simple really, just like that radio DJ telling us all, ‘Don’t look at the Sun!’
References.
Get our latest news and publications
Sign up to our news letterResources
Social
Contact us
Address
Niche Science & Technology
Unit 26 Falstaff House
Bardolph Road
Richmond TW9 2LH
United Kingdom
