At the beginning of the 2024-2025 academic year my colleagues at the Bodleian Libraries and I ran our usual induction programme in cooperation with the Oxford Law Faculty, including sessions for postgraduate students on plagiarism. Inevitably there were questions about the use of generative AI, and we had checked the University guidance with regard to assessments. It was (and remains) as follows: “Please note that artificial intelligence (AI) can only be used within assessments where specific prior authorisation has been given, or when technology that uses AI has been agreed as reasonable adjustment for a student’s disability (such as voice recognition software for transcriptions, or spelling and grammar checkers).”Footnote ¹

This pins down the use of AI in assessed work at Oxford to specifically authorised situations, or as an adjustment for students with special needs. Nevertheless, it was clear during our induction sessions that students were using generative AI – if not in their academic studies, then in their daily lives. There seemed to be a gap between what was happening in the commercial world and academia, and a further space between the social and academic realities of student life.

This is hardly surprising. In 2024, a report commissioned by the EU Commission noted that while until 2014 academic institutions were ahead of the private sector in areas of technological innovation, since then, the vast computational power required by advanced machine learning (ML) systems has resulted in industry and commerce taking the lead. Further, in the United Kingdom, although at that time we were ahead of our European counterparts, we were far behind the USA and China. According to the report, if we measure research activity by the number of citations “[t]he top-cited AI papers are from companies like Google, Meta, and Microsoft. The growth rate of AI-related publications is much faster than overall scientific production Globally.”Footnote ²

In short, AI is dominated by the tech industry, and so the apparent divide in take-up between academia and the legal profession is not surprising: extensive use and development of advanced generative AI tools demands resources that many universities have not got. This commercial domination has practical consequences and ethical implications, as the drivers for academia are not the same as those for organisations that are profit orientated.

I had four questions:

1. 1. Is it desirable to integrate the use of generative AI into academic law programmes?

2. 2. Does this risk undermining the quality and originality of students’ work as they seek to discover an authentic academic voice?

3. 3. Is it tenable to educate students, who will seek placements and ultimately jobs in law firms, without offering them the chance as part of their academic experience to develop a hands-on understanding of the strengths, weaknesses and practical applications of generative AI?

4. 4. Is it possible to resolve this perceived tension between academia and professional practice, and if so, how?

Since then, the picture at Oxford along with other universities has moved fast.Footnote ³ In fact in October 2024 there were a great many people in Oxford already thinking about generative AI; it takes time to evolve and agree on policies in a university, but Oxford has moved at such breakneck speed that the title of this article was out-of-date long before the 2025 BIALL Conference.

In Oxford, there is now guidance for students on the use of AI, and experiments are encouraged. To give just a few examples, the University now has an AI and ML Competency Centre,Footnote ⁴ and researchers at Oxford use AI to investigate fields that range from medical research to space travel, and from environmental protection to musical composition.Footnote ⁵ On 4 March 2025, the University announced a five-year collaboration with OpenAI aimed at enriching research, expanding the University’s AI capabilities and developing staff skills,Footnote ⁶ and the Oxford Law Faculty is currently offering all academic staff licences for ChatGPT Edu in a trial to support research.

Discussions about generative AI are everywhere, and I am very grateful to a number of colleagues and former colleagues who have been kind enough to talk to me and share their experiences.Footnote ⁷

My initial questions were essentially practical ones, but it quickly became apparent that answering them required consideration of a range of ethical and even philosophical problems. Several common themes emerged, in particular concerns around questions of attribution; privacy and data security; transparency and bias; equity; and questions about the impact of generative AI on jobs and the environment. None of these are trivial and they all bear upon our professional conduct and the ethics of using generative AI.

The academics and the librarians that I spoke to were for the most part already trying out or using AI for a range of tasks with mixed results. The tasks they found that generative AI can already perform effectively, if by no means perfectly, included:

For all these functions the generative AI tool needs to be sufficiently good to carry them out reliably, but the technology is not, and possibly does not have to be, perfect. The people I interviewed mentioned drawbacks that included hallucination, inaccuracy and cliché-ridden text. On the other hand, when it worked well it saved time, so long as the work involved in checking the outputs was not too onerous.

In this paper, I largely leave aside questions about how good or reliable generative AI is: it is going to get better than it is now, and as a recent paper from the UCL Law Faculty, ‘Artificial Intelligence, Education and Assessment at UCL Laws’, notes, it would be a mistake to hinge analysis on current capabilities.Footnote ⁸ That is not to say that we should ignore present shortcomings. To take just one issue, that of privacy: the administrative or semi-administrative tasks listed above may, on the face of it, be unlikely to pose ethical problems, but that is not necessarily the case. A summary of a lengthy email chain about when to hold a meeting might not be sensitive (we usually just need the information about the final decision to be clear). However, if the string includes an explanation from someone about why they can’t make a particular date because they have a hospital appointment, the data immediately becomes sensitive and should not be fed into a language model. Even banal and ordinary tasks often involve filtering out private information that not everyone should access.

It is, however, easy to see how some administrative work can lend itself to AI applications; routine tasks that demand minimal human creativity and judgment can be well suited to automation. Looking at more routine academic work, generative AI has the capacity to take the labour out of document screening and summarisation. On its web page, the AI research tool Elicit claims that its AI-enabled systematic reviews are up to 80% faster than manual reviews.Footnote ⁹ Generative AI’s ability to identify relevant documents and analyse, extract and synthesise their data not only saves time, but more importantly enables researchers to focus on innovation and criticism. Medical research, with its extensive use of systematic reviews, has been an early adopter, and in the field of legal research some law firms are using AI tools to review large numbers of documents; for example, Australian firm Minter Ellison uses an in-house generative AI tool called Lantern, which can process 3500 documents an hour, as opposed to a human reviewer’s 20 to 25.Footnote ¹⁰ This is phenomenal, and immediately raises questions about which parts of our legal research ‘work’ can, or should, AI tools do for us?

It is not unethical, per se, to use technology to make legal research more efficient: we have been doing that for years. However, a common theme that emerged from my conversations with colleagues was the need, firstly, to fully understand the nature of our work, and what we are trying to do when we carry out research, so that we can make wise judgments about the situations in which generative AI can help us.

The example above, from Minter Ellison, illustrates the potential of generative AI to bring us to new starting points. In some situations, beyond the legal sector, this is already facilitating significant advances in human knowledge, accelerating research simply by advancing the stage at which the research can begin.

For example, in the sciences it took over 20 years to make the first three-dimensional model of a protein molecule using X-ray crystallography; now generative AI can model proteins in seconds.Footnote ¹¹ This means that, thanks to AI, researchers can start from a completely different point, with clear benefits for medical and pharmaceutical innovation. To take just one instance, researchers in Cambridge have used ML to screen millions of entries in a chemical library to identify compounds with potential to treat Parkinson’s Disease.Footnote ¹²

This example from the sciences seems a very clear case of a situation in which the technology is benefitting us without apparently diminishing or devaluing human intellectual input. However, acceleration does not devalue the achievements or methodologies of earlier researchers either.

I draw a comparison from a different field: 19th Century classical philology, specifically Wortphilologie, which involved finding every instance of a word in a classical text and then using comparison and analogy of all the instances to understand every nuance of meaning.Footnote ¹³ Today we can use the function key ‘Ctrl f’ to find a word every time it appears in a text in a matter of seconds, whereas classical scholars such as Gottfried Hermann (1772-1848) must have spent hours doing that preparatory work.Footnote ¹⁴ Does the fact that we can complete that aspect of the task in a few seconds invalidate their research or negate the importance of their work? I argue that it does not, because a significant part of the ‘work’ was having the idea in the first place, and the actual research was interpreting what was discovered through a process of close attention to the text. The fact that we can now do things more quickly than was possible in the past does not detract from, or render unnecessary, the core tasks of contextual human thinking and critical analysis.

Turning to philosophy, could AI have resolved Mr Ramsay’s predicament in Virginia Woolf’s book To the Lighthouse? Mr Ramsay is a philosopher, who sees thought in terms of the letters of the alphabet to be worked through firmly and accurately, like the keys on a piano, and although he can reach the letter Q, he struggles to get beyond it to R.Footnote ¹⁵ Generative AI can sometimes take us straight to Z and make that our starting point. It seems to work for protein folding, and might have worked for Wortphilologie but is that enough? Is Mr Ramsay’s goal as a philosopher simply to arrive at the letter Z, or is the process of arriving there also important? He needs to get to the logical conclusion of his thoughts knowing that he has not omitted any necessary stages; if he is to justify his solutions, he must understand how he gets there. This is part of what it is to be a philosopher; but more than that, one becomes a philosopher by doing philosophy.

In May 2025, the choreographer William Forsythe gave a public lecture at the Blavatnik School of Government in Oxford in which he discussed the experiences and influences that made him a choreographer. Choreographers, he explained, learn through observation, participation in dance, the application of skills transferred from other disciplines in the arts and sciences, and through practice. The experience of ‘doing’ choreography and behaving like a choreographer gradually changes them, until they are transformed from being someone that can make up dances into someone that has ‘become’ a choreographer.

In the same way, legal research is not only about finding things out; by doing legal research, students advance on the road of understanding and development towards ‘becoming’ lawyers, whether in academia or the professions. Although starting legal research from a more advanced point, thanks to a generative AI tool, is likely to be time-saving and advantageous, a lawyer must still be able to deploy skills that range from the ability to interpret citations and track down cases, to familiarity with legal concepts and a thorough understanding of the application of the law. This legal practice, which begins the moment a student begins to study the law, is all part of the work involved in gradually becoming a lawyer.

Whether we should work with generative AI is not the question: legal researchers have always been open to new technologies, from the 19th Century when we began systematically to use indexes and citators, through to the 1970s when Lexis was founded and the European institutions were beginning to develop CELEX, and on to the 1990s, when we began to have a computer with Internet access on every desk.Footnote ¹⁶

The question is, how should we work with this new technology? What could possibly go wrong? There have already been problems. To take an Australian example (Valu v Minister for Immigration and Multicultural Affairs (No 2) [2025] FedCFamC2G 95), in January 2025 a court referred a lawyer to the Office of the New South Wales Legal Services Commissioner (OLSC) for including citations to non-existent cases in court filings.

This is what the judge reasoned:

In this paragraph, the judge notes firstly that the use of generative AI in legal proceedings is an issue of current concern, and secondly that not all courts have developed guidelines to manage it.

It is important that we work on the basis that legal researchers and practitioners will use technology for more and more tasks, but if we wait for the technology to be ready and more widely adopted before we start planning, we shall be ‘playing catch-up’, just as we are currently nationally and internationally in the face of the rapid and largely unregulated expansion of social media.

What happened in Valu v Minister for Immigration and Multicultural Affairs was clearly wrong, as it is wrong to mislead the Court, but it was also a failure of professional integrity. I believe that it is concerns about honesty and integrity that unite us as library and information professionals, whichever sector we work in. The BIALL Professional Skills Framework, which complements the CILIP Ethical Framework, includes a section on compliance and ethics.Footnote ¹⁸ CILIP’s seven ethical principles are:

• Human rights, equalities and diversity, and the equitable treatment of users and colleagues

• The public benefit and the advancement of the wider good of our profession to society

• Preservation and continuity of access to knowledge

• Intellectual freedom, including freedom from censorship

• Impartiality and the avoidance of inappropriate bias

• The confidentiality of information provided by clients or users and the right of all individuals to privacy

• The development of information skills and information literacyFootnote ¹⁹

There is no specific mention of AI, but all these principles have a bearing on how we should work with it. They mesh with the requirements of good academic practice, and for their part, the professional lawyers that we directly or indirectly serve, are subject to regulation and codes of conduct. Generative AI can undoubtedly save time, but that should never be to the detriment of our core values, and there are three key steps that we can and should take.Footnote ²⁰

The first requirement is transparency. Whether it is student coursework, an article for a learned journal, a presentation at a conference, or advice to a client, if we have used generative AI we must disclose it. This is part of good referencing and citation practice. In an academic setting, readers need to be able to evaluate what they read and possibly replicate findings, in order to follow the author’s path. The same applies in professional practice, with the additional proviso that clients in some areas, such as government departments, may be cautious about, or object to, the use of generative AI on their behalf.

The second requirement is that material generated by AI is always checked by a human being. This goes beyond checking grammar, hyperlinks and even facts. The person responsible for a piece of work, whatever it is, must examine the logic of the discussion or argument, refine language to make it as precise as possible, and make good any inaccuracies or omissions.

Thirdly, and related to this, every piece of work should have a responsible author (or authors), who should have made a substantial contribution to the end product. In some circumstances this might be met in part by designing prompts and curating data outputs, but the main point is that using generative AI should augment and improve the quality of the work. Passing off work generated by AI as one’s own is akin to copying and pasting from Wikipedia and claiming authorship, and in a university is likely to be a form of academic misconduct.

In summary, legal research requires acknowledgement of methods and sources, careful and critical assessment by a human being who takes responsibility, substantial input by each named author, and transparency to enable others to evaluate and replicate the work. Whether the setting is a university or a law firm, the use of a large language model (LLM) does not change the requirements of authorial responsibility.

Honesty and integrity are of crucial importance. No-one wants to be accused of plagiarism, but for far wider societal reasons it is important that this new and very exciting technology does not throw us off balance and make us forget the values that we hold dear. In the legal sector and beyond, library and information professionals are being encouraged to adopt generative AI, but the fact that we can do something does not necessarily mean that we should do it. It is for us to drive the technology, rather than for the technology to drive us.

The arrival of generative AI is a rupture on the scale of the French Revolution in the 18th Century or the Russian Revolution in the 20th. It has been likened to the discovery of fire, with the potential to benefit or destroy.Footnote ²¹ It will completely change the world as we know it, ‘know’ being a significant word, as there is so much still to do with large language models that we do not know. The accuracy of AI models depends on the quality of the underlying data, which is often not clear. In-house models, or those provided by legal aggregators, can mitigate this concern, but using them uncritically and assuming that their datasets are unbiased may still mislead us and reinforce particular ways of looking at the world, whether in terms of socio-economic norms or by shaping aesthetic taste.Footnote ²² It is as if generative AI is poised to take up the mantle of classicism as a lens through which to see, order, and judge the world.Footnote ²³

Further, if we don’t know the scope or origin of the underlying dataset, we cannot know whether or not the owners or subjects of the data gave informed consent to its use. These are copyright, security and privacy issues, of the highest concern, and beg the question: are LLMs based on unacknowledged theft? There is a case for arguing that generative AI that relies on LLMs is likely to be inherently unethical, and it has been proposed that data scraped from the web and online systems constitutes the raw material for a digital coup that is already bringing about the collapse of the international order.Footnote ²⁴

Although the outlook for wider society is not necessarily universally alarming, (according to an ILO report published in 2023 “most jobs and industries are only partially exposed to automation and are thus more likely to be complemented rather than substituted by AI”),Footnote ²⁵ the environmental implications are enormous; according to the Irish Central Statics Office, in 2023 21% of Ireland’s electricity went to data centres.Footnote ²⁶

These massive geo-political problems are essentially about equity and fairness. Fairness is not the same as justice, but it is a related concept, and as law librarians we all work, to a greater or lesser extent, with the profession that is responsible for the exercise and administration of justice in our various jurisdictions. I believe that our responsibility is to understand the possibilities of generative AI, and to be proactive, but not precipitate. It is better to move forward thoughtfully, preparing our guidelines and policies before we introduce it, rather than be forced to respond reactively afterwards, when it may be too late. Whatever our reservations about generative AI, even if we are not using it at work ourselves, others are. We cannot put the genie back in the bottle, so it is important that we start thinking now about what kind of work we are happy to leave to AI and how we shall manage it.

Returning to my questions, I think that it will be possible to integrate the use of generative AI into academic law programmes, if that is what we wish to do. I believe that there is a very real risk that without appropriate guidelines, students using generative AI could fail fully to develop the legal reasoning and quality of intellectual argument that is a necessary part of the process of becoming a lawyer, but that is not an inevitable outcome. For that reason, I think it is essential that we enable and support students to gain a practical understanding of the uses and missuses of generative AI, and that this experience and knowledge should be part and parcel of the critical thinking skills that they develop at university and continue to apply as academics or professional lawyers. I do not believe that there is a divide between academia and the profession, although there are significant superficial differences, because although we are subject to very different pressures, ultimately, we are all bound by ethical standards and a shared commitment to professional integrity.