Can AI Improve Legal Research Quality?
Legal research tools now answer questions in prose and cite the cases they rely on. Grounding answers in a real legal database cuts invented authority compared with a general chatbot, but the best independent test still found the leading products wrong or misleading on one query in six or more. Australian courts have answered by making the lawyer verify every citation.
On this page
Better: Note and confidence
Moderate confidenceGrounding reduces invented authority; it does not remove itIn a preregistered study (figures from the 2024 preprint), legal research tools hallucinated on 17% to 33% of queries against 43% for GPT-4. An industry benchmark in which vendors opted in scored AI tools above a lawyer baseline. We found no study that measures missed authority directly.Faster: Note and confidence
Low confidenceFaster drafts in trials; checking time not countedA randomised trial with law students reports productivity gains of 50% to 130% in five of six legal tasks (abstract only). The abstract reports no measure of the time practising lawyers spend verifying citations.Cheaper: Note and confidence
Low confidenceNo independent cost evidenceWe found no independent study of what AI legal research saves once verification is included. The costs of getting it wrong are visible in court decisions and regulator action.Safer: Note and confidence
Governance-gatedSafe only with human citation checksWhere Gen AI helped prepare submissions, the NSW Supreme Court requires authors to verify that every citation exists, is accurate and is relevant, and says that check must not be done solely with AI. The Federal Court's practice note expects the responsible person to have confirmed that cited authorities exist and support the proposition stated.The problem, in numbers
Legal research has one job: find the authority that governs the question, and make sure it says what you claim. Generative AI has made the second half of that job visible. A database maintained by researcher Damien Charlotin had identified 2,241 court and tribunal decisions worldwide, as at 29 October 2026, in which a party was found or implied to have relied on AI-hallucinated material, usually fake citations. 112 of them were Australian.Source 1 It is a floor, not a census: it counts only decisions that address the issue.Source 1
The consequences are now regulatory. In Victoria, a solicitor who tendered an AI-generated list of authorities containing inaccurate citations and summaries lost the right to practise as a principal or handle trust money, and must work under supervision for two years.Source 2
This review asks the quality question: does AI find the right law more reliably, and what does checking it take? Our review of discovery costs covers document review, and client data covers confidentiality and privilege.
What improves, and why
A general chatbot answers from patterns in its training data, and on law it does badly. In a peer-reviewed study of models available in late 2023, chatbots hallucinated on between 58% (GPT-4) and 88% (Llama 2) of direct, verifiable questions about US federal court cases.Source 3 They did worst on district court cases and often accepted false premises in the question.Source 3
Legal research products add retrieval. The tool first searches a legal database, then writes an answer from what it found and cites it. In our reading, that is the mechanism behind the "better" claim: an answer tied to a real case can be checked, and a tool drawing on the whole body of case law has a better chance of finding the controlling authority than one drawing on a narrow set of documents.
The evidence supports the direction. In a preregistered Stanford-led study of 202 US legal queries, run in March to May 2024 (figures from the preprint, later published in a peer-reviewed journal), Lexis+ AI gave accurate, properly cited answers to 65% of queries against 49% for GPT-4.Source 4 Ask Practical Law AI, which draws only on the publisher's own practice articles, gave incomplete answers, either refusals or answers without supporting sources, on 62% of queries.Source 4 In our reading, coverage matters as much as the model.
Grounding turns an invented case into a real one. It does not guarantee the real case says what the answer claims.
What the evidence actually says
Grounding reduced hallucination but did not remove it. In the same study, Lexis+ AI and Ask Practical Law AI each hallucinated on 17% of queries, Westlaw AI-Assisted Research on 33%, and GPT-4 on 43%.Source 4 Hallucination there includes misgrounding: citing a real source that does not support the claim.Source 4
202 preregistered US legal queries, tested March to May 2024. Hallucination includes citing a real source that does not support the claim.
The causes are instructive. Among Lexis+ AI's hallucinated answers, 47% involved naive retrieval, failing to find the most relevant sources, and 38% inapplicable authority, such as the wrong jurisdiction, statute or court. For Westlaw, reasoning errors dominated, at 61%, often over sources that seemed relevant.Source 4 The authors conclude that users must keep verifying that key propositions are supported, and that the tools work best as the first step of research, not the last word.Source 4
Newer evidence is more encouraging, with caveats. A peer-reviewed randomised trial assigned upper-level law students to a retrieval-based legal tool, a reasoning model or no AI. The published abstract reports that both tools significantly improved work quality, that productivity rose by 50% to 130% in five of six tasks, and the retrieval tool did not introduce more hallucinations.Source 5 We read only the abstract, and the participants were students, not practising lawyers. An October 2025 industry benchmark scored four AI products between 74% and 78% on US research questions against 69% for a lawyer baseline, with legal products ahead of ChatGPT on authoritativeness.Source 6 Vendors opted in and approved their evaluation, and LexisNexis and Thomson Reuters products did not take part.Source 6
In our reading, the evidence shows grounded tools beat general chatbots and may match unaided lawyers on accuracy. We found no study that measures missed authority directly, and no Australian evaluation.
Where it falls short
Real cases, wrong use. Misgrounding and inapplicable authority are harder to spot than an invented case, because the citation checks out.Source 4
A snapshot of moving products. The independent test dates from 2024 and US law. Products have changed since, and no comparable Australian test exists.Source 4
Narrow sources, narrow answers. A tool limited to its publisher's own content gave incomplete answers to 62% of queries.Source 4
Local law is not guaranteed. The NSW Supreme Court warns that the data behind these tools may be out of date or not relevant in New South Wales or Australia.Source 7 Every study we found tested US law.
Where governance decides the outcome
Australian courts have made verification the lawyer's job, in writing. The NSW Supreme Court requires that, where Gen AI helped prepare submissions, the author verify in the document that every citation and authority exists, is accurate and is relevant, and that check must not be done solely with an AI tool.Source 7 The note names lawyer-directed products such as Lexis Advance AI and CoCounsel as Gen AI, while excluding dedicated research software that only searches legal materials.Source 7 The Federal Court's practice note expects the person responsible for a document prepared with Gen AI to have confirmed that cited authorities exist and support the proposition stated, and warns of adverse costs orders; according to law firm commentary, the Federal Court, Queensland and Victoria also flag possible referral to regulators.Source 8 Victoria's regulator has already restricted one solicitor's practice over AI-generated citations.Source 2
How to think about adoption
Choose tools by what they search: coverage of Australian legislation and case law matters more than the fluency of the answer. Use them to find a starting set of authorities, then read every case you rely on, checking jurisdiction, currency and whether it supports the proposition. Record that check, because NSW requires it in the document where Gen AI was used, and do not leave it to an AI tool alone, which the same practice note rules out. Before rolling a tool out, test it on questions your own lawyers have already researched, including questions with a false premise, and compare its authorities with theirs; the published studies show results vary sharply between products. Track how often your team finds a misgrounded citation, and treat any vendor claim of hallucination-free research as untested until an independent evaluation supports it.
BFCS.ai does not fabricate figures. The strongest independent test of legal research tools is US, preregistered and dates from 2024, and the products have changed since; we read it in preprint because the published version was not accessible. The benchmark that scores AI above lawyers is an industry report in which vendors opted in, and the trial showing productivity gains used law students. We found no study that measures missed authority directly, and no Australian evaluation.
Sources
- 1
Charlotin, D., "AI Hallucination Cases Database," last updated 29 October 2026 (researcher-maintained database, CC BY 4.0, not peer-reviewed). damiencharlotin.com
Back to textWhat we took from it
2,241 cases identified so far; the site's State filter lists Australia (112) and USA (1514). Tracks court and tribunal decisions in which AI use, established or alleged, is addressed in more than a passing reference, where the decision finds or implies that a party relied on hallucinated content; it does not track the wider universe of all fake citations or AI use in filings, and describes itself as a work in progress.
- 2
Victorian Legal Services Board and Commissioner, "Statement on the Mr Dayal matter," 2 September 2025. lsbc.vic.gov.au
Back to textWhat we took from it
The practitioner tendered to the Federal Circuit and Family Court a list and summary of authorities generated using AI that contained inaccurate citations and summaries (Dayal [2024] FedCFamC2F 1166). On 19 August 2025 the Board varied his practising certificate: he may no longer practise as a principal, handle trust money or operate his own practice, may practise only as an employee solicitor, and must complete two years of supervised legal practice, with he and his supervisor reporting to the Board quarterly.
- 3
Dahl, M., Magesh, V., Suzgun, M. & Ho, D. E., "Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models," Journal of Legal Analysis 16(1), 2024, pp. 64-93, published 26 June 2024, doi:10.1093/jla/laae003 (peer-reviewed journal; open access). academic.oup.com
Back to textWhat we took from it
Tested GPT-4 (gpt-4-1106-preview), GPT-3.5, PaLM 2 and Llama 2 on direct, verifiable questions about randomly selected US federal court cases, over 800,000 queries in total: LLMs hallucinate between 58% (ChatGPT 4) and 88% (Llama 2) of the time; hallucinations were lowest for Supreme Court cases and highest for district court cases; models often accepted false premises in questions, and struggled to predict their own hallucinations.
- 4
Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D. & Ho, D. E., "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools," Journal of Empirical Legal Studies 22(2), June 2025 (online 23 April 2025, per Crossref), pp. 216-242, doi:10.1111/jels.12413 (peer-reviewed). arxiv.org
and law.stanford.edu
Back to textWhat we took from it
The published version could not be opened (publisher returned access denied); all figures in this article are from the full-text preprint, arXiv:2405.20362v1 (submitted 30 May 2024; the PDF is dated 3 June 2024), which includes the Westlaw evaluation; the published abstract also reports hallucination of more than 17% for the LexisNexis and Thomson Reuters tools (Stanford Law School publication page). Preregistered dataset of 202 US legal queries, run on Lexis+ AI, Ask Practical Law AI and GPT-4 between 22 March and 22 April 2024 and on Westlaw AI-Assisted Research between 23 and 27 May 2024; responses hand-scored. Hallucinated means incorrect or misgrounded (a citation that does not support the claim). Accurate responses: Lexis+ AI 65%, Westlaw 42% (41% in one passage of the text), Ask Practical Law AI 20%, GPT-4 49%. Hallucinated: 17%, 33%, 17% and 43%. Incomplete: Lexis+ AI 18%, Westlaw 25%, Ask Practical Law AI 62%. Ask Practical Law AI draws only on Practical Law articles, which the authors link to its low responsiveness; Lexis+ AI connects to a wider body of case law and primary sources. Contributing causes among hallucinated responses (not mutually exclusive): naive retrieval 47%, 20% and 34%, inapplicable authority 38%, 23% and 34%, reasoning error 28%, 61% and 49% (Lexis, Westlaw, Practical Law). The authors conclude that users must continue to verify that key propositions are supported by citations, and that the tools can offer value as the first step of research rather than the last word. Bar widths in the figure are scaled to the highest value (our calculation).
- 5
Schwarcz, D., Manning, S., Barry, P., Cleveland, D. R., Prescott, J. J. & Rich, B., "AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice," Journal of Law and Empirical Analysis 3(1), pp. 220-250, published 9 April 2026 (per Crossref; CC BY-NC), doi:10.1177/2755323X261427048 (peer-reviewed; abstract only read, via the University of Michigan Law School repository; the publisher page refused automated access). repository.law.umich.edu
Back to textWhat we took from it
Randomised controlled trial assigning upper-level law students to complete legal tasks with a RAG-powered legal AI tool (Vincent AI), an AI reasoning model (OpenAI o1-preview) or no AI: both tools "significantly enhance legal work quality"; productivity rose significantly in five of six tasks, with gains of 50% to 130%; o1-preview improved analytical depth and "Vincent AI avoids introducing more hallucinations". Journal peer review confirmed from the Hebrew University of Jerusalem flyer, May 2024 ("a new peer-reviewed journal, published by Sage"). An earlier working-paper version reported different ranges (about 38% to 115% for Vincent AI and 34% to 140% for o1-preview); this article uses the published abstract. The working paper states that neither OpenAI nor vLex funded the study, but both gave participants free access.
- 6
Vals AI, "Vals Legal AI Report: Legal Research" (VLAIR), updated 14 October 2025. vals.ai
Back to textWhat we took from it
Industry benchmark, not peer-reviewed: vendors joined by open call or invitation and approved their evaluation, and Vals AI has a customer relationship with one or more participants. Participants Alexi, Counsel Stack, Midpage and ChatGPT (Team), so no LexisNexis or Thomson Reuters product was tested (the report does not name non-participants). 200 question-and-answer sets on US federal and state law (two later disregarded), responses collected in the first three weeks of July 2025; weighted score of accuracy 50%, authoritativeness 40% and appropriateness 10%. AI products scored 74% to 78% against 69% for a lawyer baseline from one unnamed US firm, whose lawyers could use non-generative research tools; legal AI products averaged 6 points above ChatGPT on authoritativeness; zero-shot prompts only, and some tasks, such as sourcing for specific arguments or formatting citations, were not covered.
- 7
Supreme Court of New South Wales, Practice Note SC Gen 23, "Use of Generative Artificial Intelligence (Gen AI)," issued 28 January 2025, commenced 3 February 2025. supremecourt.nsw.gov.au
Back to textWhat we took from it
Paragraph 3 names lawyer-directed programs, including Lexis Advance AI, Westlaw Precision and CoCounsel Core, as Gen AI; paragraph 6(b) states that the note does not apply to dedicated legal research software that uses AI or machine learning to conduct searches across legislation, judgments and legal texts; paragraph 7(a) warns of hallucinations, including false citations and fabricated legislative, case or other secondary references; paragraph 7(b) warns of dependence on the quality and reach of underlying data sets, which may include data that is not up to date or not relevant in New South Wales or Australia. Paragraph 16: where Gen AI has been used in preparing written submissions or summaries of argument, the author must verify in the body that all citations, legal and academic authority, case law and legislative references exist, are accurate and are relevant to the proceedings; paragraph 17: that verification must not be carried out solely using a Gen AI tool; paragraph 18: use of Gen AI does not absolve the author of professional or ethical obligations.
- 8
Federal Court of Australia, "Use of Generative Artificial Intelligence Practice Note (GPN-AI)," D S Mortimer, Chief Justice, 16 April 2026, paras 4.5(b) (the responsible person is expected to have confirmed that "legal authorities cited in submissions exist and support the proposition stated") and 5.1 (possible consequences include adverse costs orders and issues as to compliance with legal and professional obligations). fedcourt.gov.au
Other jurisdictions and referral to regulators, law firm commentary (secondary): claytonutz.com
Back to textWhat we took from it
Clayton Utz, "Learned Friend or just a Large Language Model? The Rise of Generative AI in Australian Litigation," 8 September 2026. According to the firm, Supreme Court of Victoria practice note SC Gen 25 requires court users to be prepared to identify the portions of documents produced using AI; Queensland, South Australia and Western Australia have their own directions or guidelines; and the Federal Court, Queensland and Victoria flag possible referral to legal services regulators.