Journalology #152: AI disclosure
Hello fellow journalologists,
In this edition of Journalology I’ve weaved together threads related to the use of AI in research and publishing. So the focus of today’s newsletter is firmly on AI; the usual eclectic mix of publishing stories will return soon.
Actual intelligence sought
We’re not prescriptive, but for certain article types we’ve decided to take a harder line. For rapid responses, letters, and opinion articles, permitted AI use is now limited to improving spelling and grammar and assisting with translation. No AI use is allowed among our columnists, who are paid for providing their expert clinical opinion and insights. For other articles, particularly analysis and editorials, we continue to prioritise the deep analysis of ideas, nuance, complexity, and argumentation that AI alone is unlikely to provide.
In the past week two of the leading general medical journals (BMJ and JAMA) have released updated policies on how authors should use AI.
The BMJ editorial quoted above spells out how AI should (not) be used in the creation of review and opinion content. The policy seems eminently sensible to me. There are far too many bland and vapid AI-generated opinion pieces being published (and not just in journals).
The Economist recently explained How to spot AI writing and conducted its own experiment by comparing content written by Economist journalists with articles generated by large-language models. One of the conclusions was:
Bots’ sentences tend to be long; paragraphs are rarely interrupted with short, punchy statements. How dull.
On Tuesday Anthropic published How Claude marks AI-generated content, in response to the recent EU AI Act (this is a useful primer on the Act).
Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch. Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported.
In other words, every time someone uses Claude from the EU their content will include a digital watermark detectable by (for example) publishers.
This game of cat and mouse will not end well. Costs will spiral as editors and publishers spend more time trying to sort the wheat from the chaff. No one wins in this scenario, but it’s the one we have to live with.
The JAMA Network Updated Guidance for Author Use of AI in Medical Publication contains a helpful table that outlines how AI tools can be used in different scenarios. JAMA Network editors are following a similar principle to the BMJ: LLMs should not be used to draft comment and opinion articles. I wish more outlets would adopt a similar policy.
Using AI for peer review, to create images, or to generate / edit reference lists is a strict ‘no’, too:
Even frontier models can introduce realistic-looking references, reflecting real authors and titles that support a particular point but that do not exist. As a result, we have updated our guidance to advise authors to not use AI, including large language models (LLMs), to generate, format, or otherwise manage references.
The journal ecosystem is highly fragmented; AI policies vary widely between journals and subject areas. It’s no wonder that authors are confused.
In September last year the STM Association published Recommendations for a Classification of AI Use in Academic Manuscript Preparation. It announced a follow up project in January 2026, which involved three other organisations, and invited feedback from the community:
To support a shared understanding of how AI should be disclosed in research, STM is part of a joint harmonisation initiative to work towards a Global Reporting Standard for AI Disclosure in Research, together with the Committee on Publication Ethics (COPE), the International Science Council (ISC), and the Global Young Academy (GYA) as key partners.
On July 28 STM provided a further update:
We are proud to announce that on July 9, the Second Consultation Round of this focus track towards a global reporting standard was launched. To ensure the standard reflects diverse, real-world perspectives, we warmly invite all STM Member Organisations to contribute to this consultation. It’s about five questions:
1. Which AI use should be disclosed? (Thresholds)
2. Where in the article should it be disclosed? (Placement)
3. How should it be structured? (Taxonomy)
4. Should non-empty AI disclosure be mandatory?
5. How to signal responsibility and accountability?By the end of the year, STM will invite its members again for a third and final consultation, sourcing feedback on a draft reporting standard on AI disclosure for the scientific publishing industry.
The deadline to provide feedback is October 16 (you can find out more here).
A common view of AI best practice will be helpful, especially if it’s widely adopted. However, a clear disclosure policy is unlikely to be sufficient on its own.
Below the fold you’ll learn about:
A review article on AI use in publishing
Eric Schmidt’s views on agentic AI and science
A survey of 10,000 researchers on AI for publishing tasks
A discussion of the future of the research paper