AI in UK Legal Research: Verification, Sanctions, and the SRA

AI-generated legal research now sits inside the daily workflow of a meaningful share of UK fee-earners in 2026, and the SRA has been unambiguous about where the responsibility lands when a hallucinated citation reaches the court. The fee-earner signs the work. The tool is not a party to the matter. This article sets out, in plain terms, what that means for a UK law firm: where AI research tools actually sit in the stack, what hallucination looks like in legal work, what UK courts and the SRA have said about it, how a defensible verification workflow is structured, and what a firm does next.

The framing is descriptive. Nothing in this article asserts SRA, Law Society, or Bar Standards Board alignment, endorsement, or accreditation on behalf of Learn AI. Any policy that would bind the firm should be reviewed by the firm's COLP, COFA, or external solicitor.

Where does AI-generated legal research sit in the UK stack in 2026?

AI-generated legal research sits in three places in a UK law firm in 2026: inside dedicated legal research products, inside general-purpose AI tools used by fee-earners off-platform, and inside drafting suites that surface case references while the fee-earner is writing. The three categories carry very different risk profiles, and a firm that governs one without governing the other two has an open flank.

Dedicated legal research products. Westlaw Edge UK, LexisNexis's generative-AI assistants, Thomson Reuters's CoCounsel, vLex's Vincent AI, and the research features inside domestic products such as Casetext (now within Thomson Reuters) all surface AI-generated summaries, suggested authorities, and drafted analysis grounded against curated UK and EU caselaw databases. The vendors have invested in retrieval-augmented generation against their editorially-reviewed content, and the output quality on mainstream UK authority is substantially higher than on a general-purpose model. The risk is not zero. The risk profile is that of a capable but occasionally-wrong research assistant who cannot be fired.

General-purpose AI tools used off-platform. A fee-earner who asks ChatGPT, Claude, or Gemini to summarise a point of UK law, find supporting authority, or draft a skeleton argument is using a tool that has no reliable grounding in UK caselaw and no incentive to decline. The tool will produce a plausible-sounding paragraph with what appears to be a citation. Whether the citation exists, whether the neutral citation is correctly formatted, whether the proposition cited is actually what the case held, and whether the case is still good law are four separate questions the tool has not answered. The fee-earner who pastes that paragraph into a pleading has not checked them either.

AI features inside drafting suites. Microsoft Copilot inside Word, the AI features inside iManage and NetDocuments, and the suggestion layer inside Harvey and Spellbook offer to draft, summarise, and surface authority inside the fee-earner's existing drafting environment. The provenance problem here is subtle: the fee-earner cannot always see whether a suggested citation came from a grounded legal database or from the underlying model's training data. A governance posture that depends on which tool produced the draft needs a way to tell.

The single most useful audit a Managing Partner or COLP can run in 2026 is a register: every AI tool currently producing research output inside the firm, the fee-earners using it, the matter types it is being used on, and the grounding model it runs against. The register is almost always longer than the firm expected, and the consumer-tool entries are almost always the highest-risk ones.

What does hallucination actually look like in UK legal research?

Hallucination in legal research is not a single failure mode. It is a family of overlapping failures that share one feature: the output reads as though it carries authority it does not have. Four patterns recur and are worth naming, because the verification response is different for each.

Fabricated authority. The model produces a citation that does not exist. The party names look right, the neutral citation format is correct, the court is plausible, and the year is consistent with the proposition being cited. Nothing in the output flags it as invented. This is the failure mode that has produced most of the public-facing sanctions in England and Wales and the United States. It is the most detectable failure mode because verification against any caselaw database will show the citation absent.

Real authority, misdescribed. The model produces a real case, correctly cited, but describes a holding the case does not actually support, or describes the ratio decidendi in terms drawn from a head-note rather than the judgment. Verification against the citation confirms the case exists; only reading the judgment reveals the proposition is wrong. This is the failure mode most likely to slip past a rushed verification pass.

Superseded authority. The model produces a case that was good law at the point the underlying training data was assembled but has since been overruled, distinguished, or materially qualified. The citation is real. The original holding is described accurately. The current status of the law is not. This is a harder verification problem because it requires a negative-authority check, not a positive-authority one.

Statutory drift. The model produces a quotation from a UK statute that is broadly correct but misstates the section, quotes a repealed provision, or renders the provision as it stood before a commencement order that has since altered it. The UK statute book's rolling amendment cycle makes this failure mode especially common on frequently-amended legislation (corporate insolvency, data protection, tax, immigration).

The common thread across the four patterns is that none of them is detectable by reading the AI output alone. All four are detectable, in minutes, against a curated UK legal database. The governance problem is not the availability of the check. It is the discipline and the time allocation inside the fee-earner's workflow.

What have UK courts and the SRA said about hallucinated citations?

UK courts and the SRA have said four things, consistently and across multiple 2024 and 2025 interventions. The positions have not softened in the months since.

Fee-earners are responsible for anything they submit. The SRA's published notice on the use of generative AI (2023, with follow-up statements through 2024 and 2025) makes explicit that fee-earners remain responsible for the accuracy of anything they file, regardless of whether the draft was generated by AI. The Code of Conduct for Solicitors requires competence and supervision; the Code of Conduct for Firms requires systems and controls. AI-assisted work does not relocate the competence obligation onto the tool vendor.

The courts have imposed costs sanctions for fabricated citations. Matter reporting in the legal press through 2024 and 2025 has recorded multiple UK tribunal and civil-court interventions where fabricated authority produced by generative AI was identified by the court. Outcomes have included wasted-costs orders, adverse references in judgments, and referrals to the SRA. The pattern mirrors earlier US decisions (Mata v. Avianca, 22-cv-1461, and subsequent matters) but is now documented inside the England and Wales system. The position that fee-earners cannot offload responsibility to the tool is not an SRA theoretical concern; it is a court finding.

The Bar Council's published position echoes the SRA. The Bar Council's guidance on the use of generative AI in legal practice, published in 2024 and updated through 2025, aligns with the SRA on the competence and accuracy points. Barristers operating under instructions from solicitors who are using AI-assisted research are not insulated from the verification responsibility when the authority is drawn into a pleading or a skeleton argument.

Judicial guidance has been published. The Judicial Office issued guidance on AI use by judges and judicial office-holders in December 2023, refreshed through 2024-2025, that sets expectations on verification, confidentiality, and disclosure when AI tools are used in judicial drafting. The guidance does not directly bind fee-earners, but it indicates the framework judges are applying when AI-assisted pleadings reach them. A pleading that arrives with a hallucinated citation inside that framework will not be treated leniently.

Read together, the four positions do not ban AI-assisted research. They require the firm to treat verification as part of the work, not an optional step after the fact.

What does a defensible verification workflow look like?

A defensible verification workflow is five steps. Every step is short. The value is in doing all five on every AI-assisted research output before the fee-earner treats the output as firm work product.

Step one: classify the output. Before anything is verified, the fee-earner records whether the output was produced by a grounded legal research tool (Westlaw Edge UK, LexisNexis, CoCounsel, Vincent AI) or by a general-purpose model (ChatGPT, Claude, Gemini, Copilot inside Word against a non-grounded back-end). The classification determines how heavy the verification needs to be. Grounded outputs still need checking; ungrounded outputs need every assertion to be independently verified.

Step two: verify every citation against a curated database. Each case citation, statutory reference, and regulatory reference in the output is looked up in a curated UK source: BAILII for primary judgments, Westlaw or Lexis for the editorially-reviewed version, and the official statute book on legislation.gov.uk for statutory text with the current commencement status. Fabricated authority is detected here. The check takes under a minute per citation when the database is already open.

Step three: read the holding. For every citation that survives step two, the fee-earner reads the judgment (or at minimum the headnote plus the relevant paragraphs of the ratio) to confirm the case actually supports the proposition the AI output cited it for. Misdescribed authority is detected here. This is the step that most often gets skipped under time pressure.

Step four: negative-authority check. For every surviving citation, the fee-earner runs the citator (Westlaw's KeyCite, Lexis's CaseSearch, vLex's Vincent citator) to confirm the authority has not been overruled, distinguished, or materially qualified. Superseded authority is detected here. The citator check is a parallel step to step three; on a well-configured workstation both can be in flight simultaneously.

Step five: record the verification. The fee-earner records, on the matter file, which AI tool produced the output, who verified it, the date of verification, and the citator position on each authority. The record is two or three lines. Its function is not to satisfy the fee-earner; it is to satisfy the Head of Legal Operations, the SRA, or the court if the matter is ever reviewed. The firm's systems-and-controls obligation under the Code of Conduct for Firms runs through this record.

The Secure AI for Fee-Earners workshop walks the firm through the five-step workflow against its actual tool stack and produces a one-page template the firm can adopt.

Where firms consistently get verification wrong

Three anti-patterns recur and each is worth naming.

Treating grounded tools as self-verifying. Even the best legal research products in the UK market occasionally misdescribe authority or surface a case with a stale citator status. The correct frame is that grounded tools reduce the rate of hallucination, not that they eliminate the need to verify. A firm that has told its fee-earners "you can rely on Westlaw AI" has misread the vendor's own terms.

Delegating verification downward. Partners who delegate AI-assisted research to a trainee and then verify the trainee's work are verifying the wrong artefact. The verification obligation attaches to the fee-earner who signs the output; it cannot be discharged by verifying that a trainee performed a verification step. A partner who signs an AI-assisted skeleton argument is responsible for every authority in it.

Relying on the model's own confidence signal. Current generative models do not reliably signal when they are uncertain. A hallucinated citation will be produced with the same apparent confidence as a verified one. Asking the model "are you sure about that citation?" produces a response that is uninformative about the ground truth. The verification has to be external.

What a UK law firm does next

The posture is available. It is a three-step sequence a Managing Partner can complete in the same quarter.

First, run the AI research tool inventory across the firm. Name every product, every consumer AI account in active use by fee-earners, and every AI feature inside a drafting suite. Circulate the register to every partner and require confirmation.

Second, adopt the five-step verification workflow as firm policy. Document it in a one-page note signed off by the COLP and circulated through the firm's learning-management platform. Incorporate a reference to it in the firm's engagement letter if the firm wishes to disclose AI use to clients, though the SRA does not currently require disclosure as a mandatory position.

Third, train the fee-earner base on the workflow against the firm's actual stack. A half-day session that walks partners and fee-earners through the five steps on real (redacted) matter examples produces an evidenced competence baseline the firm can point to if an SRA or court query ever lands. The AI Readiness Assessment is the fifteen-minute diagnostic that shows the firm where on that sequence its current posture sits.

AI-generated legal research is not going to retreat from UK practice. What is going to happen, over the next twenty-four months, is that the firms that built the verification workflow early will continue to use AI tools competitively, and the firms that did not will carry a cumulative exposure that lands in a single visible incident. The five-step workflow above is not novel. It is the research workflow a senior fee-earner has always done, applied to a new category of research output. Every firm currently using AI-assisted research already knows how to verify; the question is whether the workflow is documented, trained, and evidenced.