Generative AI use in academic publishing

Having realised the great value of the high-quality, peer-reviewed content they control for training AI language models, academic publishers are beginning to look closely at the opportunities provided by Generative Artificial Intelligence (Gen AI). A number of major academic publishers, including Taylor & Francis, Wiley and Oxford University Press have made multi-million pound licensing deals with AI companies like Microsoft to allow the use of their scholarly content as training data for large language models (LLMs). As these publishers own the copyright to these works they do not need the consent from authors to do so. Other publishers like Cambridge University Press have contacted their authors to request their permission before licensing their content for the training and development of LLMs.

Such approaches may cause uncertainty and disquiet for authors, who feel that Gen AI tools may use their work without proper attribution, may misinterpret their work, or use it in ways which are not foreseeable or plainly incorrect. The Library feels that CUP are to be applauded for making this clear and putting in place a contract that allows them to do this legally and to offer royalties. We would like to see more publishers following CUP’s example. 

Ithakar S+R have produced a Generative AI Licensing Agreement Tracker that is recording these licensing deals.

• Author Consent and Rights: There are concerns about authors retaining control over the use of their work, especially as some publishers may not provide clear opt-out choices or adequate notifications.
• Research Integrity: The accuracy and updates of scholarly content within AI models, including handling corrections or retractions, remain unclear.
• Citation and Attribution: It is yet to be determined how Gen AI outputs will handle the provenance and citation of the original works they utilize.
• Long-term Impact:  Gen AI is redefining how research is read, summarised and valued, with the long-term impact not clearly understood.

We advise researchers to:

1. Understand their rights by reviewing their publishing agreements.
2. Keep abreast of their publisher’s practices regarding AI licensing.
3. Consider if the venue where they publish is acting in their best interests.

We have received queries from authors who feel that the University’s requirement to deposit publications in an open access repository puts them at risk in the same way. We are increasingly seeing Gen AI’s consumption of research publications (and datasets) being presented as an objection to the principles of Open Access, primarily because authors feel they are losing control of their work and do not consent to their work generating profit for commercial companies. On the face of it this is a valid complaint, however we consider it a complete red-herring. Since most authors have willingly signed over their copyright to commercial publishers their works are already being exploiting to generate revenue - through selling journal subscriptions, charging open access or other publications fees, and now through licensing LLM training models.  Again we ask if academic publishers who are triple dipping are acting in authors best interests.

What is the Library doing?

Like many academic libraries, we have been experiencing slowness and outages on some of our open access platforms over the last few months. We have identified that this is the result of poorly behaved Gen AI crawling activity. As such we have had to try to disable the ability for AI -harvesters to access content from our repositories, via IP range blocking, geographical blacklisting, implementing limits to throttle bots and using robots.txt files to limit access. 

We acknowledge that this approach is not consistent with our open access policy which should require us to permit machine access to our content. Our limiting activity is not 100% fool-proof, and is a temporary measure to give services breathing space to perform while a longer termn solution is worked out by software providers and the community. This approach also has the added benefit of helping to assuage authors fear of losing control of their works.

We have discussed these issues within the Library and have consensus on the following points. 

• Gen AI has the potential to be a transformative and positive tool supporting high quality research. 
• Gen AI consumption of research publications is happening anyway, and there is no stopping it, regardless of whether licences are adhered to.
• We feel strongly that it is better for Gen AI models to consume content from high-quality sources. There will be far better outcomes for society if Gen AI models are trained on high quality datasets and peer-reviewed publications. It is better for this to happen in a controlled, regulated, legal environment.
• Restricted availability off high-quality content increases the risk of Gen AI models being trained on poor quality content. 
• Whilst some authors may not like that publishers are asking for permission to us their works to train AI models, we feel that CUP are to be applauded for making this clear and putting in place a contract that allows them to do this legally. We would like to see more publishers following CUP’s example. 
• The Library will listen to its academic community and act in their best interests.