By moving the cost of verification onto the person least equipped to carry it, and by removing the early-career work through which lawyers learned to carry it. Law runs on a currency that generative models counterfeit convincingly and cheaply: the citation. A fabricated authority has a name, a year, a court and a neutral citation number. The form is correct when the content does not exist. That single property explains most of what has happened to the profession since 2023.
This page uses the evidence base that law has and other professions do not: a public, dated, growing record of what goes wrong when the checking fails, maintained by an outsider and now cited by courts.
Nearly two thousand decisions, and the largest group is not lawyers#
Damien Charlotin's AI Hallucination Cases database tracks decisions in which a court or tribunal has explicitly found or implied that a party relied on hallucinated material. It excludes mere allegations. As at its update of 27 August 2026 it recorded 1,963 cases, the earliest from the second quarter of 2023.
By jurisdiction: United States 1,345, Canada 214, Australia 98, United Kingdom 62, Israel 57, with more than thirty other countries represented. By nature of the defect: fabricated material in 1,634 entries, misrepresented authority in 816, false quotations in 528.
Then the number that reframes the story. By the party responsible: self-represented litigants 1,127, lawyers 784, judges 29, expert witnesses 15.
The dominant coverage of this problem is professional embarrassment, and the majority of the cases are not lawyers at all. They are people without lawyers who found a tool that produced something that looked like a legal argument. That is an access-to-justice phenomenon wearing the costume of a professional scandal, and the two need different remedies. The database owner is careful that this counts only decisions where the court addressed the point, so the true universe is larger.
The 29 judges are the entry to sit with. Fabricated authority has reached the bench itself.
The tools sold as hallucination-free#
Magesh and colleagues at Stanford ran the first preregistered empirical evaluation of commercial AI legal research tools, published in the Journal of Empirical Legal Studies in 2025. Their stated reason for doing it was that providers had described retrieval-augmented generation as eliminating hallucinations, or had guaranteed hallucination-free legal citations.
They wrote over 200 legal queries across four categories, preregistered the dataset with the Open Science Foundation before running anything, and graded responses on whether they were both correct and grounded in the sources cited.
- Lexis+ AI, the strongest performer: accurate on 65 per cent of queries, incomplete on 18 per cent, and hallucinating on more than one in six.
- Westlaw AI-Assisted Research: accurate 42 per cent of the time, incomplete on 25 per cent, with a hallucination in one-third of responses, roughly twice the rate of the other legal tools tested.
- Ask Practical Law AI: incomplete on 62 per cent of queries, because it draws only on articles written by an in-house team rather than on primary law.
One mechanism in the paper matters more than the headline rates. Westlaw produced the longest answers, averaging 350 words against 219 for Lexis+ AI and 175 for Ask Practical Law AI. More words means more falsifiable propositions and more chances to be wrong, and it also means more to check. The paper puts it directly: lengthier answers require substantially more time to check, verify and validate, because every proposition and citation has to be independently evaluated.
That is the trap in one sentence. The tool that produces the most useful-looking output imposes the most verification, and the verification is invisible, unbilled and easy to skip. This research has called the general form of that problem the verifier's discount: the work of checking is systematically undervalued relative to the work of producing.
What the English courts settled in June 2025#
In Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), decided on 6 June 2025, the Divisional Court, Dame Victoria Sharp P and Johnson J, dealt with two referrals under the Hamid jurisdiction and used them to state the position for the profession.
On what the tools can do, at paragraph 6:
Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research. Such tools can produce apparently coherent and plausible responses to prompts, but those coherent and plausible responses may turn out to be entirely incorrect. The responses may make confident assertions that are simply untrue. They may cite sources that do not exist. They may purport to quote passages from a genuine source that do not appear in that source.
On the duty, at paragraph 7, lawyers using such tools must check the accuracy of the research by reference to authoritative sources before using it, and the court names them: the government legislation database, the National Archives judgments database, the official Law Reports, and the databases of reputable legal publishers. At paragraph 8 the duty extends to lawyers relying on the work of others who used AI, on the same footing as relying on a trainee or a pupil.
The facts underneath. In Ayinde, grounds of claim cited five authorities that do not exist; one of them carried a real neutral citation number belonging to an entirely different case. In Al-Haroun, a claim seeking damages of £89.4 million, a schedule of forty-five citations was put before the court of which eighteen did not exist, and of those that did, many contained neither the quotations nor the propositions attributed to them. One of the fabricated authorities was attributed to the very judge hearing the application.
The holding with the widest practical reach is at paragraph 81:
A lawyer is not entitled to rely on their lay client for the accuracy of citations of authority or quotations that are contained in documents put before the court by the lawyer. It is the lawyer's professional responsibility to ensure the accuracy of such material.
At paragraph 23 the court set out the full range of its powers: public admonition, a costs order, a wasted costs order, striking out, referral to a regulator, contempt proceedings, and referral to the police. At paragraph 31 it added that, save in exceptional circumstances, admonishment alone is unlikely to be a sufficient response. In Ayinde it found the threshold for contempt proceedings met and declined to initiate them, giving five reasons including that the barrister concerned was extremely junior and apparently operating outside her level of competence, and stating that the decision is not a precedent.
The paragraph the profession has under-read#
Paragraph 9 does something the coverage largely missed. It puts the obligation on people who did not touch the tool.
practical and effective measures must now be taken by those within the legal profession with individual leadership responsibilities (such as heads of chambers and managing partners) and by those with the responsibility for regulating the provision of legal services... For the future, in Hamid hearings such as these, the profession can expect the court to inquire whether those leadership responsibilities have been fulfilled.
Read alongside the reasons given for not pursuing contempt in Ayinde, where the court referred to the regulator the question of whether those supervising the barrister's pupillage had complied with requirements as to supervision, work allocation and competence, the direction is clear. The court treated a fabricated citation as a supervision failure with a junior at the end of it, and said future hearings will ask who was supervising. That is the same accountability question this research has put to leadership elsewhere, in who supervises work they cannot do.
The regulator's answer was to forbid the machine from proposing case law#
On 6 May 2025 the Solicitors Regulation Authority authorised Garfield.Law Ltd, described as the first purely AI-based firm permitted to provide regulated legal services in England and Wales. It handles small claims for unpaid debts up to £10,000.
The conditions are the interesting part. Among them: the user must approve each stage, named regulated solicitors remain accountable for all system outputs, and the AI is precluded from proposing case law.
A regulator willing in principle to authorise a firm with no human fee-earners drew its line at exactly the function the empirical evidence says is unreliable. That is a more sophisticated regulatory response than either the prohibitionist or the permissive reading suggests, and it points at where the profession is heading: not a ban on AI, but a boundary drawn task by task around the operations where fabrication is both likely and consequential. The general version of that boundary is set out in the Delegation Boundary Map.
The rung that is disappearing, and the number that does not exist#
The work most exposed in legal practice is document review, first drafts, research memoranda and citation checking. It is also, historically, how a junior lawyer learned what an authority is worth: by reading fifty of them badly, being corrected, and eventually developing the reflex that something is off.
Reading the Magesh findings next to the Charlotin database gives the uncomfortable version. The tools produce output whose defects are detectable only by someone with the judgement that used to be built by doing the work the tools now do. The verification requires the capability; the capability was built by the task; the task has been automated.
No published dataset tracks trainee solicitor or pupil barrister hours by task before and after 2023. This page does not have the number and will not manufacture one. The nearest available evidence is general rather than legal: Brynjolfsson, Chandar and Chen find employment among 22 to 25 year olds in highly AI-exposed occupations running about 19 per cent below where it would sit on the comparison trend, through reduced hiring rather than dismissal, and concentrated in occupations where AI substitutes rather than complements. Whether law sits in the substituting group is currently an argument rather than a measurement. The structural case is at missing rungs.
What remains unresolved#
- Nobody has measured whether AI use changes the quality of legal outcomes. Every number on this page concerns errors caught, tool accuracy or employment. None concerns whether clients win or lose more often.
- The Magesh evaluation is a snapshot of specific product versions. Providers update continuously; the finding that the claims outran the systems is durable, the specific percentages are not.
- The database counts decisions, not incidents. Cases where nobody noticed, or where the point was resolved without a written decision, are invisible by construction, and its author says so.
- No court has yet ruled on whether a lawyer who used AI competently and was still misled is negligent. Every reported case involves a failure to check at all.
- Whether disclosure of AI use should be mandatory in filings is unsettled, and jurisdictions are diverging on it.
What a firm or chambers can do this quarter#
- Check every citation against an authoritative database, and record that it was checked. The Divisional Court named the four acceptable classes of source. A verification step that leaves no trace is indistinguishable from no verification when it is examined afterwards.
- Never verify one model against another. Two systems trained on overlapping text agreeing with each other is not corroboration.
- Treat length as a cost signal. On the Magesh evidence, the longer the generated answer, the more verification it demands and the more likely it is to contain something false.
- Name the supervisor for AI-assisted work, in advance. Paragraph 9 says the court will ask.
- Protect the checking reps for juniors deliberately. If a trainee never reads the authority in full, the firm is buying speed with a capability it will need in five years.
- Ask what happens to the litigant in person. They are the majority of the cases in the database, they will not read a practice direction, and courts are absorbing the cost.
Key research and primary sources
- Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D. and Ho, D. E. (2025). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies, 22(2).
- Charlotin, D. (ongoing). AI Hallucination Cases database. Figures cited from the update of 27 August 2026.
- Divisional Court of England and Wales (2025). Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), 6 June 2025.
- Solicitors Regulation Authority (2025). SRA approves first AI-driven law firm, 6 May 2025.
- Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine?. Stanford Digital Economy Lab.
Judgment paragraphs are quoted from the published text on the judiciary website. Database figures were read from the source on the date stated and will drift, so the date is given.
Related SuperSkills research#
On the underlying failure mode, what an AI hallucination is and why AI sounds so confident. On who carries the checking, who owns verification and the verifier's discount. On the boundary, the Delegation Boundary Map and meaningful human oversight. On the profession-level method, deskilling risk by profession, and on the neighbouring cases, medicine and consulting. On the junior end, missing rungs.
About this research#
Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The Divisional Court judgment was read in full at the primary source and paragraph numbers are given for every proposition attributed to it. Database figures were taken directly from the source page on 28 August 2026 and reflect its update of 27 August 2026. No figure for trainee hours is given anywhere on this page because none could be found. This is commentary on the professional and evidential position, not legal advice.
Cite this
Hirji, R. (2026). How will AI change law? The SuperSkills Intelligence Company. Last reviewed 28 August 2026. thesuperskills.com/research/how-will-ai-change-law