The Making of OLH's AI Policy

In January 2026, the Open Library of Humanities (OLH) published the first iteration of our Artificial Intelligence (AI) Policy in response to the growing need for guidance on the use of generative AI in scholarly publishing. The journey to producing a workable policy for all our journals has been more than a year-long undertaking. The OLH's Managing Editor, Dr Simon Everett, has overseen an extensive editorial consultation and policy drafting process to arrive at a flexible and appropriate response to the rise of generative AI use in scholarly humanities publishing. In this blog post, he explains the context around the necessity for an AI Policy, and how the OLH's AI Policy was created and implemented.

The Current AI Landscape.

Generative AI has surged in popularity and use cases in the last few years to become seemingly ubiquitous. Large language models (LLMs) are used frequently to produce outputs for all kinds of purposes. You can ask generative AI a question and it will produce an answer. You can tell an LLM to write a story and it will try to, with varying degrees of coherence. You are given an AI generated summary drawn from many websites whenever performing a simple web search, and you're expected to make use of it. Such AI output loudly precedes actual search results at the top of the page.

Not providing any guidance on the use of generative AI, for an open access publisher, is no longer an option. At the OLH, we have been monitoring the progress of adequate guidance of the use of generative AI for more than two years—but anything beyond very basic measures, particularly those outlined by the Committee on Publishing Ethics (COPE), has not at the time of writing been formalised into a set of universal standards which publishers can rely upon.

This is perhaps the most difficult aspect of forming a policy on AI for a publisher: not knowing exactly what will come next. We can work with what has already been established, but beyond the basic concept that generative AI is not a person and is therefore not a legal entity, there is almost nothing: a lack of concrete guidance and general awareness of the problems raised by AI technologies, whilst they are being rolled out at lightning speed across online platforms and software applications.

Here is a taste of the potential issues we have had to take into account as a publisher, to which there are not always satisfactory answers or solutions:

• Authors using generative AI knowingly, with purpose for their research, and unknowingly, perhaps as a tool integrated into software;
• Authors not knowing that generative AI has no legal claim to authorship and therefore cannot be listed as a co-author, contributor, etc.;
• Authors having no reasonable control over what happens to their published work in terms of LLM model training;
• Editors inappropriately using generative AI to make an editorial decision that they pass off as their own;
• Reviewers using generative AI to fabricate peer review reports for an author's work;
• Editors and reviewers using generative AI to help produce reports or to help make editorial decisions, which may lead to feeding an author's work into an LLM without their consent.

The OLH can take a stand on requiring editors, reviewers and authors to be appropriate actors in the publishing process, informed about why they should not undertake questionable and unethical activities that would affect research integrity and editorial best practice. However, some of the above issues, in many ways frustratingly, lie beyond our control. How do we, for example, prevent an author's research from being fed into an LLM, which could be used for model training, when generative AI is integrated into applications and its use now encouraged so widely? The risk of one actor in the publishing workflow, perhaps without thinking, asking an LLM to paraphrase a section of an author's work is not something anyone can reasonably police or prevent from happening.

The UK Government's proposals for ‘AI, the creative industries and copyright reform’ (6th March 2026) has only recently been scrutinised by the House of Lords, who have robustly advocated for the need for ‘a clear public statement setting out an expectation that commercial AI developers operating in the UK should obtain appropriate licences when using copyrighted works to train generative AI models’, which would mean ‘a licensing-first approach as the baseline for the UK’ (Summary of Conclusions and Recommendations, Chapter 2, 12, paragraph 71). There are signs that the ethical haziness surrounding generative AI use may have some much-needed clarity approaching, at least for AI developers operating in the UK, and particularly with regard to an author's consent of the use of their work for AI model training. This is a positive development, especially for open access publishing where there are currently insufficient protections in place to guarantee proper use of copyrighted work published under an open access license (e.g. Creative Commons licenses).

At the time of writing, we inhabit a continually shifting, lawless terrain, where ethics and integrity are at constant risk of being sacrificed for the benefit of productivity and commodity. The onus falls to us—the publisher—to put in place guidance on the acceptable and responsible use of generative AI for our journals and authors when there is currently not much actionable guidance from governments, courts and other organisations across the globe.

Finding a Way Through.

Although the response from many is that generative AI should be banned or controlled somehow to counter the threat it has on upholding the best standards of research integrity in publishing, there is no sufficient means of being able to practically enforce anything that is binding.

Detecting the use of generative AI LLMs that have written whole articles by using generative AI detection tools can lead to false positive results (University of San Diego Legal Research Centre, ‘Generative AI Detection Tools’). Many of these detection tools are also beyond the affordability of smaller not-for-profit publishers, who are constantly managing resources as efficiently as possible.

Trying to spot AI written work with a human editorial eye also raises some difficult questions: what if the author has required the use of a generative AI tool or LLM to assist with writing their work due to a disability? What if English is not an author's primary spoken language and they have required some generative AI-assisted help with translation into English? This is plausibly a matter of discrimination if a manuscript is rejected on that basis, when the research itself is otherwise rigorous, original and productive. Trying to assess whether an article has been output by generative AI by the way it has been written is no longer a viable approach as LLMs are trained on more material and become increasingly sophisticated.

When tasked with trying to put together a workable generative AI policy for the OLH, given it is such a contentious and nascent area of policy, I decided that it was absolutely necessary to engage with our journal editorial teams across a wide range of fields within the humanities and social sciences. Across two full rounds of consultation, I was able to gather important feedback that was worked into subsequent drafts of the policy. A full record of this working can be read here, which details the kind of issues, responses and amendments that were made to the policy over the course of more than a year.

I quickly realised that the best that can be done, at least for now, is to lay out what we as a publisher clearly see as unacceptable uses of generative AI, alongside acceptable and responsible uses of generative AI. The problem with wanting to state that ‘an author should declare all uses of generative AI’, especially in all circumstances, is that this is no longer a realistic ask. As generative AI tools and LLM-driven applications have spread rapidly and sometimes seamlessly into our workflows, the requirement for an author to know and declare every instance of generative AI use becomes increasingly untenable. Knowing for certain when a prompt is input into a generative AI model or tool and then reproducing the output(s) in research is clearly a different case to being led by a (possibly unrequested) AI-driven suggestion on how to reframe presented content, restructure an article, or how to rewrite an introductory paragraph. Differentiating such cases is the core of what our AI policy seeks to achieve: filtering out concrete and known uses of generative AI for specific research purposes that should be cited and declared from light and minor editorial uses that have little to no bearing on the originality and methodology of the produced research.

We have been extremely lucky that our fantastic editorial teams across the OLH's roster of journals were thoughtful, receptive and constructive in their responses to our policy drafts. Integrating as much as possible into our AI policy from the legitimate concerns that they raised was not always easy, and there was a clear divide between those willing to accept that minor use of generative AI is inevitable, and those who are very cautious and sceptical about any AI use cases. The OLH's stance, far from being extremely welcoming of the use of generative AI in all cases, has tried to, where possible, urge caution against the overuse of and unethical use of generative AI, particularly where it presents clear and unacceptable malpractice (for example, AI-fabricated datasets passed off as an author's own work, or a reviewer fabricating a peer review report in its entirety using AI without reading an author's work in full).

It's in the spirit of one of the OLH's guiding principles—to empower rather than restrain journal editors—that we firmly believe it is for editors to make their own expert assessment of manuscripts in order to ascertain whether they constitute unacceptable academic quality for their journals. This at least has some bearing on eliminating wholly AI-generated manuscripts from being published, and focuses on human oversight to determine the accuracy and scholarly worth of submitted research.

Flexibility and Extensibility are Vital.

What we have sought to achieve at the OLH is a flexible, living AI policy that emphasises what can be done practically, right now, to ensure that the most important and impactful uses of generative AI on an author's research are transparently declared as a matter of record. However, our line that minor uses of generative AI need not be formally declared by authors is not a hard one. Asking authors to adequately cite and declare responsibility for known and substantial uses of generative AI is a start, but by no means the end of where we are headed in terms of future legal guidance and best practices adopted by other journals and publishers. Affording our journals some flexibility here means that journals can extend their own policies to work in addition to OLH's. We already have some journals seeking to add to our guidance at an individual level that is not needed across all our journals: for example, the disclosure of generative AI tool use in videographic scholarship for production and post-production.

Above all, we must trust our editors to lead on what they feel is needed in addition to our OLH AI policy for their journals. Ensuring that our AI policy is extensible means that, as a publisher, we are not in contradiction with the specific requirements of our journals' editorial teams. We are finding that getting the basic principles of acceptable, responsible, and unacceptable generative AI use right allows for robust iterations on our publisher AI policy, across the many diverse fields that constitute the humanities.

But what about the future? We can't know exactly what generative AI use will look like in a few years' time, or even a few months' time. Although we welcome the recent announcement by COPE that there will be a new ‘Global reporting standard for AI disclosure in research’ (25th March 2026) following an initiative at the World Conference on Research Integrity in May 2026, the outcome of this will not be available until at least late 2026. An area that also sorely needs to be addressed is author consent of the use of their research in widespread generative AI model training, particularly for a diamond open access publisher where the research we publish is available for all, for free. Ethically, this is of great importance for authors. But with this, nothing can be enforced until there is movement from governments and courts, particularly (as aforementioned) with adequate licensing. How much protection any legal outcomes will give to authors is not easy to see, and how this might be adequately policed is a further unknown.

We are still very much at a place where there are further problems to be solved, and impracticalities to be made practical. But we hope that the OLH's AI Policy is a much needed start, putting forward what we believe is robust and fair guidance on generative AI's use in scholarly publishing until we are required to adapt to a future we can't yet know.

 

Dr Simon Everett has been with the OLH since 2021, and in his role as Managing Editor he ensures that the publisher's policies and the policies across all the OLH's journals are in keeping with best practice standards.

 

Read the news item for the release of OLH's AI Policy here.

 

Like what we do? Recommend us!

If you like the work that the Open Library of Humanities is doing, please recommend us to your institution. We cannot operate without our library members. Your support helps keep us open and accessible for all.