AI has already entered the law school library, the seminar room and the late-night essay crisis. The question is not whether law students will use it. They already are. The real question is whether it helps them develop legal judgement – or quietly replaces it.
According to the most recent HEPI/Kortext Student Generative AI Survey (February 2025), 92% of UK undergraduates now use AI in some form, up from 66% just a year earlier. 88% have used generative AI for assessments – up from 53% in 2024. In a single academic year, generative AI has gone from a curiosity to a default study habit. Pretending otherwise is no longer a serious position.
But the more revealing figure in the same survey is this: only 36% of students have received any training from their institution on how to use AI well, while 67% believe AI skills are now “essential”. A generation of law students is teaching itself how to use the most consequential study tool of its lifetime — without much guidance, and without much consensus on what “using it well” even means.
That gap is the real story. And for law students specifically, it matters more than for almost anyone else.
The AI reality in UK higher education
The Quality Assurance Agency (QAA), which oversees academic standards across UK higher education, frames generative AI as carrying both “far reaching implications” and “definite opportunities” for learning, teaching and the value of academic awards. The Russell Group’s Principles on the use of generative AI tools in education – adopted by all 24 of its member universities – commits institutions to making students “AI-literate,” to upholding “academic rigour and integrity,” and to adapting assessment rather than banning the technology outright.
In other words, the official sector position is neither prohibition nor laissez-faire. It is critical, supervised, educated use. The problem is that the official position and the lived reality on the ground are not the same thing. Only 29% of HEPI’s respondents felt their institution actively encouraged AI use; 40% disagreed. Students described policy as “vague,” “mixed messages,” “discouraged but not banned.” Into that vacuum, generic chatbots step in and offer a confident answer to almost any question a stressed law student can type.
That, not AI itself, is the real concern. A confident answer is not the same as a well-reasoned one. And a generic chatbot is not the same as a worked example produced by a qualified lawyer.
Why law is different from other subjects
Most subjects have to think carefully about AI. Law has to think harder.
In Artificial Intelligence, Education and Assessment at UCL Laws (UCL Legal Studies Research Paper 04/2025), a faculty group led by Michael Veale and including Isra Black, Orla Lynskey, Colm O’Cinneide and Eloise Scotford set out, in unusually frank terms, why legal education sits in a particularly exposed position:
“Foundational topics are, in some ways, more vulnerable to the use of AI to achieve passing grades. Law degrees, by intent and by regulation, impart foundational legal knowledge. There are few topics taught at universities around the world with more structured text existing online about them than common law systems. Language models are soaked in core legal knowledge because law is an open, text-driven field.”
Statutes, judgments, textbooks, commentaries – the entire raw material of legal study – has been digitised, structured and indexed for decades. That makes it ideal training data for a large language model, and it means a generic AI system can produce competent-looking text on a tort problem question or a contract law essay in seconds. Competent-looking is the operative phrase. As the UCL authors observe, generic AI “generates statistically median content, median structure, median style and median substance.” For a student aiming at a 2:1 or a First, median is precisely the wrong target.
Law also has a regulatory dimension that students often underestimate. Qualifying law degrees function as a “trusted and authenticated marker of achievement” for legal services regulators. The SRA and BSB require disclosure of plagiarism or academic misconduct findings as part of their character and suitability assessments. A two-minute shortcut on a coursework deadline can become a question on a regulatory form years later.
The danger of “cognitive offloading”
The UCL paper’s central distinction is the one we want to dwell on, because it reframes the entire debate:
“There is a critical balance between using AI as a true study aide, facilitating meaningful learning, and using it to cognitively offload tasks in a way that hinders learning.”
“Cognitive offloading” is the key phrase. It does not mean cheating, exactly. It means handing over to the machine the very thinking that the degree exists to develop – and it can happen invisibly, even when a student types every word themselves. The UCL authors are particularly clear-eyed about structural offloading: letting the AI generate the outline, the points, the counterpoints, the order of argument. The prose may be the student’s own, but the legal thinking is not.
This is the single most important distinction in the whole AI debate. Not “did you use AI?” but “at which stage did you use AI, and was it the stage that was supposed to be teaching you something?”
But here is what the offloading argument is not saying
It is here that the conversation often goes wrong. A simplistic reading of the UCL paper would conclude that any AI-assisted writing support, any model essay, any worked example must be damaging. That is not what the evidence —-or the UCL authors – actually say. And it would be at odds with several decades of cognitive science on how complex skills are actually learned.
Learning by example is one of the most evidence-backed methods in education
Cognitive Load Theory, developed by John Sweller and refined over forty years of empirical research, is now mainstream in educational psychology. One of its central findings is the worked example effect: novices learn complex problem-solving skills more efficiently by studying fully worked examples than by attempting unguided practice. Sweller’s own summary is blunt – worked examples produce “superior performance compared to problem solving,” and this finding has been replicated across mathematics, science, programming and, increasingly, legal reasoning.
The reason is straightforward. A novice law student confronted with a problem question on remoteness in contract has to do several extremely demanding things at once: recall the relevant authorities, identify which facts trigger which rules, structure an argument, anticipate counter-arguments, write in legal register. Working memory simply cannot hold all of that simultaneously. A worked example reduces extraneous cognitive load and frees the student to focus on the underlying schema – the pattern of legal reasoning – which is the thing that actually transfers to new problems.
This is why textbook problem questions come with model answers. It is why moot judges give feedback on past performances. It is why every good legal writing handbook is full of annotated examples. It is why supervisors share First-class scripts (anonymised) with their students.
Scaffolding is how complex skills are taught – full stop
The other pillar of the relevant evidence base is scaffolding, derived from Vygotsky’s concept of the Zone of Proximal Development (ZPD): the gap between what a learner can do alone and what they can do with expert support. Wood, Bruner and Ross’s foundational 1976 paper formalised the idea: an expert tutor models, prompts, simplifies and gradually withdraws support as the learner takes over. The expert does not do the thinking for the learner. The expert does the thinking visibly, so the learner can see how it is done, and then practises with diminishing assistance until they can do it alone.
This is, almost exactly, how law has been taught for centuries. We read judgments not because we will be tested on the precise wording, but because we are watching expert legal minds reason in public. We read leading textbooks not to memorise them, but because Smith, Treitel, Hart and Honoré are modelling what good legal argument looks like. Tutorials are scaffolded discussions in which the supervisor guides the student through reasoning they cannot yet do alone. The First-class essay shown to the second-year is a worked example.
If you have never seen what a good legal argument looks like, the chance of you producing one from first principles is vanishingly small. That is not a controversial claim. It is decades of cognitive science applied to one of the most demanding analytical disciplines in the academy.
The real failure mode is unguided AI use, not modelled AI use
This is where the conversation about AI in law school needs to be more precise than it usually is. The UCL paper’s worry – and ours – is not that students should never see another worked example. It is that generic AI tools encourage students to skip the modelling stage and go straight to producing text they do not yet understand. The chatbot generates the structure; the student never internalises why that structure works. The chatbot produces the counter-argument; the student never learns to spot one unaided. There is no fading of support, no scaffold being withdrawn, no expert reasoning being made visible – just an output, untraceable, ungrounded, indifferent to whether the student learned anything at all.
That is the failure mode. Not “the student was helped.” But “the student was helped in the wrong way, at the wrong stage, by the wrong tool.”
Useful AI use versus risky AI use
Drawing the threads together, a clearer taxonomy emerges:
| Genuinely useful | Risky | Damaging |
|---|---|---|
| Studying a high-quality model answer to see how a strong argument is structured | Asking a generic chatbot to summarise a case you have not read | Submitting AI-generated text as your own |
| Annotated worked examples that make the expert’s reasoning visible | Trusting AI-generated case citations without verifying them | Asking AI to do the legal thinking the assessment exists to test |
| Using AI to test your own draft (“here is my argument – what is the strongest counter?”) | Using a generic chatbot to generate the structure of an essay before you have wrestled with the question | Cognitively offloading the moment of intellectual difficulty – the bit that is the learning |
| Specialist legal writing support that scaffolds your understanding before withdrawing | Mistaking median, plausible-sounding text for legally rigorous analysis | Pasting unverified AI output into a submission |
The pattern is consistent with both the UCL paper and the cognitive science literature. AI-assisted learning is most valuable when it operates the way a good supervisor operates: making expert reasoning visible, scaffolding the student’s own developing skill, and gradually fading away as the student becomes able to do the work alone. It is most dangerous when it is generic, opaque and substitutional – when it produces an output without producing any learning.
Why generic AI is not enough for legal study
There is a problem the major chatbots do not advertise. They are trained on vast corpora of internet text, which means they confidently produce material that looks legal but is not always legally reliable. Hallucinated case names, misstated rationes, jurisdictionally confused doctrine and subtly wrong statutory wording are all well-documented features of generic AI output. The UCL authors note that legal language model benchmarks show “even the most advanced models perform unsatisfactorily on basic legal tasks.”
Generic AI has a deeper, structural problem too. It produces median work – by design. It optimises for plausibility. Law school assesses for judgement. Those are not the same thing, and a student who learns from generic AI is learning to write the kind of essay that an examiner has already seen a thousand times.
This is precisely why specialist legal writing support – produced by qualified lawyers, grounded in real authority, explicitly designed as a teaching tool rather than a substitute – sits in a different category from a generic chatbot. The former is a worked example in the Sweller tradition. The latter is the antithesis of one.
What responsible legal writing support should look like
Pulling the threads together, responsible AI-era support for law students should:
- Make expert reasoning visible – show how the argument is built, not just present a finished output
- Scaffold rather than substitute – support the stages of thinking the student has not yet mastered, while requiring them to do the rest themselves
- Be jurisdictionally accurate and grounded in real, verifiable authority rather than hallucinated citations
- Help the student keep authorial control of their own voice, structure and argument
- Be honest about its limits – including the kinds of question no tool should be answering on a student’s behalf
- Respect academic integrity rules at the student’s institution
It should function, in short, like a good supervisor: an expert making expert reasoning legible, then standing back so the student can take over.
Used badly, AI can flatten legal thinking. Used properly, it can help students see how legal thinking is built.
How Essay Barrister fits into that future
Essay Barrister has been built around exactly this distinction. We are not a generic chatbot, and we are not a shortcut. We produce specialist legal writing support – model work, worked examples and structured guidance – designed to be learned from, in the long tradition of how law has always been taught: by reading the work of expert legal minds and absorbing how they reason.
We draw on:
- Barclay’s long history in academic support – years of working with law students at every stage from LLB to LLM to conversion
- Practising solicitor and former lecturer input into how guidance is written and what students actually need to develop
- Qualified law writers who understand the difference between an essay that sounds legal and an essay that is legal – and who structure work to make that difference visible
The UCL authors are right that generic AI, used unthinkingly, can hollow out legal education. But the answer to that is not to deprive students of expert modelling and scaffolded guidance. The answer is to make sure the support they receive is the right kind: guided by qualified lawyers, grounded in real authority, designed to teach how legal thinking is built – not to bypass the building of it.
That is what we exist to do.
References and further reading:
- Veale, M., Black, I., Dsouza, M., Fisher, M., Ghaus, M., Gibbs, T., Lynskey, O., O’Cinneide, C., Scotford, E., Thomas, O., Trapova, A., Trapp, K. (2025) Artificial Intelligence, Education and Assessment at UCL Laws, UCL Legal Studies Research Paper Series 04/2025
- Freeman, J. (2025) Student Generative AI Survey 2025, HEPI Policy Note 61 (HEPI/Kortext, February 2025)
- Russell Group (2023) Principles on the use of generative AI tools in education
- Quality Assurance Agency for Higher Education, Generative artificial intelligence (sector resources)
- Sweller, J. (2006) “The worked example effect and human cognition,” Learning and Instruction, 16(2)
- Wood, D., Bruner, J. and Ross, G. (1976) “The role of tutoring in problem solving,” Journal of Child Psychology and Psychiatry, 17(2)
- Vygotsky, L.S. (1978) Mind in Society: The Development of Higher Psychological Processes
