Open Access and Artificial Intelligence: a practical agenda for publishers – part 1

Please note: This is an AI-generated summary of a session organised by OIPA on AI in publishing, held on 5th June 2025, at which five speakers presented their views. These speakers were George Walkley, Andrew Gray, Rupert Gatti, Jane Secker and Tanya Kant. It has been edited by staff to ensure it’s an accurate reflection of the session, and the speakers have all given their permission to be quoted. The blog reflects AI-related challenges, discussions and practices in circulation at the time of the event. However the fast pace of change in  technical, legislative and workplace contexts mean that some described processes and practices may have significantly evolved.

Artificial intelligence is already changing scholarly publishing. There is no doubt that AI is affecting the way books are marketed, discovered and read. For open access publishers, the implications can be particularly complex.

Openly licensed scholarship can make research available to anyone, anywhere. The same accessibility also makes content easier for artificial intelligence companies to collect and use, often without meaningful attribution or an obvious return to the organisations that have funded, edited and published it.

These questions were explored in an OIPA webinar on AI and publishing, held on 5th June 2025. Speakers George Walkley, Andrew Gray, Rupert Gatti, Jane Secker and Tanya Kant considered the practical opportunities offered by AI alongside questions of research integrity, copyright, transparency, sustainability and control.

Their contributions suggest that open access publishers should neither adopt AI uncritically nor attempt to avoid it altogether. Instead, we need a pragmatic and values-led approach based on experimentation, disclosure and collective action.

AI is already part of publishing

AI is no longer a specialist technology operating at the edges of publishing. Generative AI is being incorporated into familiar platforms, including office software, creative tools, search engines, web browsers and document readers.

Many people working in publishing are already using it. Yet, as George Walkley observed, a substantial proportion of professional users do not disclose their use to colleagues or managers. This presents a governance problem. If organisations do not understand how employees are using AI, they cannot identify successful practices, manage risk or ensure that confidential and copyrighted material is being handled responsibly.

The discussion also highlighted an important distinction between using AI to support individual tasks and using it to replace professional roles. The most credible current applications tend to accelerate specific, bounded activities. They do not remove the need for editorial judgement, subject expertise or human accountability.

For smaller institutional and scholar-led publishers, this distinction is important. AI may help publishers expand what they can deliver without pretending that technology can reproduce the full range of expertise involved in scholarly publishing.

Where AI can offer practical value

Rupert Gatti described how Open Book Publishers has been experimenting carefully with AI across its workflows. Its work offers several practical examples for OIPA members.

Potential applications include:

• producing an initial draft of a book description from a proposal, introduction and contents page;
• creating draft alt text for authors and editors to review;
• identifying names and possible topics for an index;
• supporting marketing and social media planning;
• identifying journals and other potential review outlets;
• extracting and checking citations, URLs and DOIs;
• producing code documentation or tests;
• creating additional formats, including audiobooks;
• supporting translation and language checking.

These uses are not fully automated. The output is treated as a starting point and reviewed by staff, authors or subject specialists before publication.

Accessibility is one of the most promising areas for smaller open access publishers. Producing alt text, audio editions, translations and other accessible formats has traditionally required resources that may be beyond the reach of a small press. Used carefully, AI could reduce some of these barriers.

There are nevertheless good reasons to set boundaries. Open Book Publishers, for example, was not using large language models for research quality assessment or substantive copy-editing at the time of the webinar. These are areas in which context, disciplinary understanding and editorial judgement remain essential.

A useful principle is therefore: AI can assist with a task, but a named person must remain responsible for its outcome.

Research integrity depends on disclosure

The presence of AI-generated or AI-assisted language within scholarly publications is difficult to measure. Andrew Gray’s research has examined changes in the frequency of words and phrases appearing in academic literature since generative AI became widely available.

Some indicative terms have increased sharply, with Andrew’s research suggesting the prevalence of generative AI usage in scholarly publishing is far greater than is currently being declared. While these corpus-wide approaches cannot not show that a particular text was produced or edited using AI, they do indicate that the undisclosed use of AI is widespread throughout academic writing, and there are signs that authors and AI tools are actively adapting to avoid using well-known terms.

There are a continuum of possible uses for AI, from widely accepted approaches such as simple copyediting at one end, through to supervised (or unsupervised) translation, up to things like the generation of entire papers. But the distinctive style and vocabulary of AI-generated text is consistent across all these uses, meaning that without clear and explicit disclosure, a reader or a publisher has no easy way to identify how deeply involved the AI tool was.

Research has also found LLM-marked papers to be more problematic in a variety of ways. For example, reviews with distinctive LLM styles were more likely to be short, inconclusive, last minute and were less likely to engage with feedback.

The challenge for publishers will be the need to distinguish between low-level language support and more substantial intellectual contribution when writing their policies, to encourage authors to disclose their use of AI appropriately, and to convey that information to readers in an informative way.

Does free to read mean free to train?

For open access publishers, copyright and licensing are central to the debate. Jane Secker challenged the assumption that making a work freely available necessarily means accepting every possible reuse.

Creative Commons licences specify the conditions under which a work may be reused. They do not remove copyright altogether, except where a work is placed in the public domain using CC0. Most Creative Commons licences require attribution, but current AI systems frequently provide little or no reliable information about the sources on which an answer is based.

This creates a tension between the legal and ethical expectations of open licensing and the practices of AI companies.

Open access publishing depends on a knowledge commons sustained by authors, editors, reviewers, librarians, institutions and funders. Large-scale automated collection can impose direct infrastructure costs while transferring much of the resulting value to commercial technology companies. If these companies weaken the organisations and communities that produce reliable open knowledge, they risk damaging the resources on which their own systems depend.

The response should not be to retreat automatically from open access. Restricting access to scholarship would also affect researchers, students and members of the public, while offering no guarantee that commercial systems would respect those restrictions. Instead, publishers need a clearer understanding of what their licences permit, what their communities expect and what additional measures may be required.

The legal position also continues to develop. Questions remain about text and data mining, fair dealing and fair use, the copyright status of AI-generated material, liability for infringing outputs and the extent to which an AI system creates or retains copies of protected works. Publishers should avoid making confident legal assumptions where the position remains unsettled and seek specialist advice when necessary.

Come back for part 2!