Close
Authorship in the Era of AI – Panel Discussion
By Naomi, on 9 July 2025
Guest post by Andrew Gray, Bibliometrics Support Officer
This panel discussion at the 2025 Open Science and Scholarship Festival was made up of three professionals with expertise in different aspects of publishing and scholarly writing, across different sectors – Ayanna Prevatt-Goldstein, from the UCL Academic Communication Centre focusing on student writing; Rachel Safer, the executive publisher for ethics and integrity at Oxford University Press, and also an officer of the Committee on Publication Ethics, with a background in journal publishing; and Dhara Snowden, from UCL Press, with a background in monograph and textbook publishing.
We are very grateful to everyone who attended and brought questions or comments to the session.
This is a summary of the discussion from all three panel members, and use of any content from this summary should be attributed to the panel members. If you wish to cite this, please do so as A. Prevatt-Goldstein, R. Safer & D. Snowden (2025). Authorship in the Era of AI. [https://blogs.ucl.ac.uk/open-access/2025/07/09/authorship-in-the-era-of-ai/]
Where audience members contributed, this has been indicated. We have reorganised some sections of the discussion for better flow.
The term ‘artificial intelligence’ can mean many things, and often a wide range of different tools are grouped under the same general heading. This discussion focused on ‘generative AI’ (large language models), and on their role in publishing and authorship rather than their potential uses elsewhere in the academic process.
Due to the length of this write-up, you can directly access each question using the following links:
1. There is a growing awareness of the level of use of generative AI in producing scholarly writing – in your experience, how are people currently using these tools, and how widespread do you think that is? Is it different in different fields? And if so, why?
1. There is a growing awareness of the level of use of generative AI in producing scholarly writing – in your experience, how are people currently using these tools, and how widespread do you think that is? Is it different in different fields? And if so, why?
Among researchers, a number of surveys by publishers have suggested that 70-80% of researchers are using some form of AI, broadly defined, and a recent Nature survey suggested this is fairly consistent across different locations and fields. However, there was a difference by career stage, with younger researchers feeling it was more acceptable to use it to edit papers, and by first language, where non-English speakers were more likely to use it for this as well.
There is a sense that publishers in STEM fields are more likely to have guidance and policy for the use of AI tools; in the humanities and social sciences, this is less well developed, and publishers are still in the process of fact-finding and gathering community responses. There may still be more of a stigma around the use of AI in the humanities.
In student writing, a recent survey from HEPI found that from 2024 to 2025, the share of UK undergraduates who used generative AI for generating text had gone from a third of students to two thirds, and only around 8% said they did not use generative AI at all. Heavier users included men, students from more advantaged backgrounds, and students with English as a second or additional language.
There are some signs of variation by discipline in other research. Students in fields where writing is seen as an integral part are more concerned with developing their voice and a sense of authorship, and are less likely to use it for generating text – or at least are less likely to acknowledge it – and where they do, they are more likely to personalise the output. By comparison, students in STEM subjects are more likely to feel that they were being assessed on the content – the language they use to communicate it might be seen as less important.
[For more on this, see A. Prevatt-Goldstein & J. Chandler (forthcoming). In my own words? Rethinking academic integrity in the context of linguistic diversity and generative AI. In D. Angelov and C.E. Déri (Eds.), Academic Writing and Integrity in the Age of Diversity: Perspectives from European and North American Higher Education. Palgrave.)]
2. Why do you think people are choosing to use these tools? Do you think that some researchers – or publishers – are feeling that they now have to use them to keep pace with others?
Students in particular may be more willing to use it as they often prioritise the ideas being expressed over the mode of expressing them, and the idea of authorship can be less prominent in this context. But at a higher level, for example among doctoral students, we find that students are concerned about their contribution and whether perceptions of their authorship may be lessened by using these tools.
A study among publishers found that the main way AI tools were being used was not to replace people at specific tasks, but to make small efficiency savings in the way people were doing them. This ties into the long-standing use of software to assist copyediting and typesetting.
Students and academics are also likely to see it from an efficiency perspective, especially among those who are becoming used to working with generative AI tools in their daily lives, and so are more likely to feel comfortable using it in academic and professional contexts. Academics may feel pressure to use tools like this to keep up a high rate of publication. But the less involvement of time in a particular piece of work might be a trade-off of time spent against quality; we might also see trade-offs in terms of the individuality and nuance of the language, of fewer novel and outlier ideas being developed, as generative AI involvement becomes more common.
Ultimately, though, publishers struggle to monitor researchers’ use of generative AI in their original research – they are dependent on institutions training students and researchers, and on the research community developing clearer norms, and perhaps there is also a role for funders to support educating authors about best practices.
Among all users, a significant – and potentially less controversial – role for generative AI is to help non-native English speakers with language and grammar, and to a more limited degree translation – though quality here varies and publishers would generally recommend that any AI translation should be checked by a human specialist. However, this has its own costs.
With English as a de facto academic lingua franca, students (and academics) who did not have it as a first language were inevitably always at a disadvantage. Support for this could be found – perhaps paying for help, perhaps friends or family or colleagues who could support language learning – but this was very much support that was available more to some students than others, due to costs or connections, and generative AI tools have the potential to democratise this support to some degree. However, this causes a corresponding worry among many students that the bar has been raised – they feel they are now expected to use these tools or else they are disadvantaged compared to their peers.
3. On one end of the spectrum, some people are producing entire papers or literature reviews with generative AI. Others are using it for translation, or to generate abstracts. At the other end, some might use it for copyediting or for tweaking the style. Where do you think we should draw the line as to what constitutes ‘authorship’?
In some ways, this is not a new debate. As we develop new technologies which change the way we write – the printing press, the word processor, the spell checker, the automatic translator – people have discussed how it changes ‘authorship’. But all these tools have been ways to change or develop the words that someone has already written; generative AI can go far beyond that, producing vastly more material without direct involvement beyond a short prompt.
A lot of people might treat a dialogue with generative AI, and the way they work with those outputs, in the same way as a discussion with a colleague, as a way to thrash out ideas and pull them together. We have found that students are seeing themselves shifting from ‘author’ to ‘editor’, claiming ownership of their work through developing prompts and personalising the output, rather than through having written the text themselves. There is still a concept of ownership, a way of taking responsibility for the outcome, and for the ideas being expressed, but that concept is changing, and it might not be what we currently think of as ‘authorship’.
Sarah Eaton’s work has discussed the concept of ‘Post-plagiarism’ as a way to think about writing in a generative AI world, identifying six tenets of post-plagiarism. One of those is that humans can concede control, but not responsibility; another is that attribution will remain important. This may give us a useful way to consider authorship.
In publishing, ‘authorship’ can be quite firmly defined by the criteria set by a specific journal or publisher. There are different standards in different fields, but one of the most common is the ICMJE definition which sets out four requirements to be considered an author – substantial contribution to the research; drafting or editing the text; having final approval; and agreeing to be accountable for it. In the early discussions around generative AI tools in 2022, there was a general agreement that these could never meet the fourth criteria, and so could never become ‘authors’; they could be used, and their use could be declared, but it did not conceptually rise to the level of authorship as it could not take ownership of the work.
The policy that UCL Press adopted, drawing on those from other institutions, looked at ways to identify potential responsible uses, rather than a blanket ban – which it was felt would lead to people simply not being transparent when they had used it. It prohibited ‘authorship’ by generative AI tools, as is now generally agreed; it required that authors be accountable, and take responsibility for the integrity and validity of their work; and it asked for disclosure of generative AI.
Monitoring and enforcing that is hard – there are a lot of systems claiming to test for generative AI use, but they may not work for all disciplines, or all kinds of content – so it does rely heavily on authors being transparent about how they have used these tools. They are also reliant on peer reviewers flagging things that might indicate a problem. (This also raises the potential of peer reviewers using generative AI to support their assessments – which in turn indicates the need for guidance about how they could use it responsibly, and clear indications on where it is or is not felt to be appropriate.)
Generative AI potentially has an interesting role to play in publishing textbooks, which tend to be more of a survey of a field than original thinking, but do still involve a dialogue with different kinds of resources and different aspects of scholarship. A lot of the major textbook platforms are now considering ways in which they can use generative AI to create additional resources on top of existing textbooks – test quizzes or flash-cards or self-study resources.
4. Do you think readers of scholarly writing would draw the line on ‘authorship’ differently to authors and publishers? Should authors be expected to disclose the use of these tools to their readers? And if we did – is that something that can be enforced?
There is a general consensus emerging among publishers that authors should be disclosing use of AI tools at the point of submission, or revisions, though where the line is drawn there varies. For example, Sage requires authors to disclose the use of generative AI, but not ‘assistive’ AI such as spell-checkers or grammar checkers. The STM Association recently published a draft set of recommendations for using AI, with nine classifications of use. (A commenter in the discussion also noted a recent proposed AI Disclosure Framework, identifying fourteen classes.)
However, we know that some people, especially undergraduates, spend a lot of time interacting with generative AI tools in a whole range of capacities, around different aspects of the study and writing process, which can be very difficult to define and describe – there may not be any lack of desire to be transparent, but it simply might not fit into the ways we ask them to disclose the use of generative AI.
There is an issue about how readers will interpret a disclosure. Some authors may worry that there is a stigma attached to using generative AI tools, and be reluctant to disclose if they worry their work will be penalised, or taken less seriously, as a result. This is particularly an issue in a student writing context, where it might not be clear what will be done with that disclosure – will the work be rejected? Will it be penalised, for example a student essay losing some marks for generative AI use? Will it be judged more sceptically than if there had been no disclosure? Will different markers, or editors, or peer-reviewers make different subjective judgements, or have different thresholds?
These concerns can cause people to hesitate before disclosing, or to avoid disclosing fully. But academics and publishers are dependent on honest disclosure to identify inappropriate use of generative AI, so may need to be careful in how they frame this need to avoid triggering these worries about more minor use of generative AI. Without honest disclosure, we also have no clear idea of what writers are using AI for – which makes it all the harder to develop clear and appropriate policies.
For student writing, the key ‘reader’ is the marker, who will also be the person to whom generative AI use is disclosed. But for published writing, once a publisher has a disclosure of AI use, they may need to decide what to pass along to the reader. Should readers be sent the full disclosure, or is that overkill? It may include things like idea generation, assistance with structure, or checking for more up-to-date references – these might be useful for the publisher to know, but might not need to be disclosed anywhere in the text itself. Conversely, something like images produced by generative AI might need to be explicitly and clearly disclosed in context.
The recent Nature survey mentioned earlier showed that there is no clear agreement among academics as to what is and isn’t acceptable use, and it would be difficult for publishers to draw a clear line in that situation. They need to be guided by the research community – or communities, as it will differ in different disciplines and contexts.
We can also go back to the pre-GenAI assumptions about what used to be expected in scholarly writing, and consider what has changed. In 2003, Diane Pecorari identified the three assumptions for transparency in authorship:
1. that language which is not signaled as quotation is original to the writer;
2. that if no citation is present, both the content and the form are original to the writer;
3. that the writer consulted the source which is cited.
There is a – perhaps implicit – assumption among readers that all three of these are true unless otherwise disclosed. But do those assumptions still hold among a community of people – current students – who are used to the ubiquitous use of generative AI? On the face of it, generative AI would clearly break all three.
If we are setting requirements for transparency, there should also be consequences for breach of transparency – from a publisher’s perspective, if an author has put out a generative AI produced paper with hallucinated details or references, the journal editor or publisher should be able to investigate and correct or retract it, exactly as would be the case with plagiarism or other significant issues.
But there is a murky grey area here – if a paper is otherwise acceptable and of sufficient quality, but does not have appropriate disclosure of generative AI use, would that in and of itself be a reason for retraction? At the moment, this is not on the COPE list of reasons for retraction – it might potentially justify a correction or an editorial note, but not outright retraction.
Conversely, in the student context, things are simpler – if it is determined that work does not belong to the student, whether that be through use of generative AI or straightforward plagiarism, then there are academic misconduct processes and potentially very clear consequences which follow from that. These do not necessarily reflect on the quality of the output – what is seen as critical is the authorship.
5. Do you think ethical use of AI will be integrated into university curriculums in the future? What happens when different institutions have different ideas of what is ‘ethical’ and ‘responsible’?
A working group at UCL put together a first set of guidance on using generative AI in early 2023, and focused on ethics in the context of learning outcomes – what is it that students are aiming to achieve in their degree, and will generative AI help or not in that process? But ethical questions also emerged in terms of whose labour had contributed to these tools, what the environmental impacts where, and importantly whether students were able to opt out of using generative AI. There are no easy answers to any of these, but they very much are ongoing questions.
Recent work from MLA looking at AI literacies for students is also informative here in terms of what it expects students using AI to be aware of.
6. Many students and researchers are concerned about the potential for being falsely accused of using AI tools in their writing – how can we help people deal with this situation? How can people assert their authorship in a world where there is a constant suspicion of AI use?
There was no easy answer here and a general agreement that this is challenging for everyone – it can be very difficult to prove a negative. Increasing the level of transparency around disclosing AI use – and how much AI has been used – will help overall, but maybe not in individual cases.
Style-based detection tools are unreliable and can be triggered by normal academic or second-language writing styles. A lot of individuals have their own assumptions as to what is a ‘clear marker’ of AI use, and these are often misleading, leading to false positives and potentially false accusations. Many of the plagiarism detection services have scaled back or turned off their AI checking tools.
In publishing, a lot of processes have historically been run on a basis of trust – publishers, editors, and reviewers have not fact-checked every detail. If you are asked to disclose AI use and you do not, the system has to trust you did not use it, in the same way that it trusts you obtained the right ethical approvals or that you actually produced the results you claim. Many publishers are struggling with this, and feeling that they are still running to catch up with recent developments.
In academia, we can encourage and support students to develop their own voice in their writing. This is a hard skill to develop, and it takes time and effort, but it can be developed, and it is a valuable thing to have – it makes their writing more clearly their own. The growth of generative AI tools can be a very tempting shortcut for many people to try and get around this work, but there are really no shortcuts here to the investment of time that is needed.
There was a discussion of the possibility of authors being more transparent with their writing process to help demonstrate research integrity – for example, documenting how they select their references, in the way that systematic review does, or using open notebooks? This could potentially be declared in the manuscript, as a section alongside acknowledgements and funding. Students could be encouraged to keep logs of any generative AI prompts they have used and how they are handling them, to be able to disclose this in case of concerns.
7. Are there journals which have developed AI policies that are noticeably more stringent than the general publisher policies, particularly in the humanities? How do we handle it if these policies differ, or if publisher and institutional policies on acceptable AI use disagree?
There are definitely some journals that have adopted more restrictive policies than the general guidance from their publisher, mostly in the STEM fields. We know that many authors may not read the specific author guidelines for a journal before submitting. Potentially we could see journals highlighting these restrictions in the submission process, and requiring the authors to acknowledge they are aware of the specific policies for that journal.
8. The big AI companies often have a lack of respect for authorship, as seen in things like the mass theft of books. Are there ways that we can protect authorship and copyrights from AI tools?
A substantial issue for many publishers, particularly smaller non-commercial ones, is that so much scholarly material is now released under an open-access license that makes it easily available for training generative AI; even if the licenses forbid this, it can be difficult in practice to stop it, as seen in trade publishing. It is making authors very concerned, as they do not know how or where their material will be used, and feel powerless to prevent it.
One potential way forward is by reaching agreements between publishers and AI companies, making agreements on licensing material and ensuring that there is some kind of renumeration. This is more practical for larger commercial publishers with more resources. There is also the possibility of sector-wide collective bargaining agreements, as has been seen with the Writers Guild of America, where writers were able to implement broader guardrails on how their work would be used.
It is clear that the current system is not weighted in favour of the original creators, and some form of compensation would be ideal, but we also need to be careful that any new arrangement doesn’t continue to only benefit a small group.
The issue of Creative Commons licensing regulating the use of material for AI training purposes was discussed – Creative Commons take the position that this work may potentially be allowed under existing copyright law, but they are investigating the possibility of adding a way to signal the author’s position. AI training would be allowed by most of the Creative Commons licenses, but might require specific conditions on the final model (eg displaying attribution or non-commercial restrictions).
A commenter in the discussion also mentioned a more direct approach, where some sites are using tools to obfuscate artwork or building “tarpits” to combat scraping – but these can shade into being malware, so not a solution for many publishers!
9. We are now two and a half years into the ‘ChatGPT era’ of widespread AI text generation. Where do you see it going for scholarly publishing by 2030?
Generative AI use is going to become even more prevalent and ubiquitous, and will be very much more integrated into daily life for most people. As part of that integration, ideally we would see better awareness and understanding of what it can do, and better education on appropriate use in the way that we now teach about plagiarism and citation. That education will hopefully begin at an early stage, and develop alongside new uses of the technology.
Some of our ideas around what to be concerned about will change, as well. Wikipedia was suggested as an analogy – twenty years ago we collectively panicked about the use of it by students, feeling it might overthrow accepted forms of scholarship, but then – it didn’t. Some aspects of GenAI use may simply become a part of what we do, rather than an issue to be concerned with.
There will be positive aspects of this, but also negative ones; we will have to consider how we keep a space for people who want to minimise their use of these tools, and choose not to engage with them, for practical reasons or for ethical ones, particularly in educational contexts.
There are also discussions around the standardisation of language with generative AI – as we lose a diversity of language and of expression, will we also lose the corresponding diversity of thought? Standardised, averaged language can itself be a kind of loss.
The panel concluded by noting that this is very much an evolving space, and encouraged greater feedback and collaboration between publishers and the academic community, funders, and institutions, to try and navigate where to draw the line. The only way forward will be by having these discussions and trying to agree common ground – not just on questions of generative AI, but on all sorts of issues surrounding research integrity and publication ethics.
