Verified Authority: artificial intelligence in law firms and the duties it already tests

Artificial intelligence in law firms has already been tested in court, and the published record now shows what follows when nobody checks the output.

Nothing in this article constitutes legal advice. It is offered to support discussion and further review with qualified advisers.

What does artificial intelligence in law firms actually change?

It changes the order of the work, leaving every duty in place. The solicitor duty of competence, along with confidentiality, supervision and candour to the court, survives intact. What has gone is the habit by which reading the authorities verified them, so verification now has to be designed in as a recorded step.

• The Solicitors Regulation Authority issued a warning notice on 17 August 2026 stating that regulated individuals remain accountable for their work and outputs however that work was prepared.
• A Divisional Court considered two matters together in June 2025, one involving five fabricated housing authorities and one involving 45 problematic citations of which 18 were entirely fictitious.
• An Upper Tribunal decision promulgated in November 2025 held that uploading confidential documents into an open public tool places the information in the public domain and waives privilege.
• A study of 202 legal queries reported hallucination rates of 17 per cent and 33 per cent for two purpose-built research tools, against 58 to 82 per cent reported for general models.

What have the regulator and the courts said about legal AI governance?

Four authorities sit behind almost every argument in the paper, and they say compatible things. The warning notice of 17 August 2026 records that these systems generate fictitious cases and seemingly factual assertions with no basis in fact, requires that named case law authorities be genuine, relevant and supported by a verifiable citation, and warns that information entered into them may be retained. It offers no safe harbour to a firm following a particular process. My own reading, on which a firm should take its own advice, is that a firm’s documented standard becomes the thing an investigator examines, so a firm with nothing written down would probably be measured against what a competent firm should have done.

Why does AI hallucination case law matter to a careful firm?

Sitting in June 2025, a Divisional Court considered two matters together under its jurisdiction over those who appear before it. Counsel in the first had cited five fabricated housing authorities, while a solicitor in the second put forward 45 problematic citations, 18 of them entirely fictitious, having failed to verify research supplied by the client. The outcomes were regulatory. Adopting the words of Dias J, the court recorded that the administration of justice depends upon being able to rely without question on the integrity of those who appear before it, and on their professionalism in only making submissions which can properly be supported. The operative half is the second, since a submission that can properly be supported is one somebody has checked.

The wider scale is measurable in outline. A public tracker held 2,036 such decisions worldwide as at 10 September 2026, of which 69 were in the United Kingdom, and rather more involved self-represented litigants than lawyers. Its compiler calls the collection a work in progress, so the count partly measures detection.

How reliable are the leading legal research tools?

Measured evidence exists here, which is unusual in this field. A study of 202 hand-constructed legal queries reported that one leading tool produced accurate grounded answers in 65 per cent of cases with a hallucination rate of 17 per cent, and that a second answered accurately in 41 per cent of cases with a rate of 33 per cent. Earlier work put hallucination by general-purpose models on legal questions at between 58 and 82 per cent. Both vendors disputed the publication, and the objection about product selection was answered when the study was extended before peer review in 2025. Purpose-built tools therefore perform substantially better than general models, which is an argument about procurement, but none of those measured performed well enough to remove the verification step, which is an argument about workflow.

How should a firm handle client confidentiality and AI?

Three sources point the same way. A tribunal has held that uploading confidential documents into an open tool such as a public chatbot places that information in the public domain, breaching confidentiality and waiving privilege, while distinguishing open models from closed or enterprise systems. The regulator warns that inputs may be retained, and the representative body warned in September 2025 that data often leaves the United Kingdom. One qualification travels with that holding, because the Upper Tribunal binds the First-tier Tribunal, and its authority stops below the High Court. On my reading, a firm should be able to demonstrate technically that matter material reaches only services it has contracted with, since a written policy is unlikely to be treated as adequate once a breach has occurred.

What does law firm AI supervision look like in practice?

Six technical controls carry most of the assurance burden, aligning with security guidelines co-sealed by agencies from 18 countries. Matter-level entitlements resolve before retrieval and before any tool call, contracts define training use, retention, hosting region, deletion and audit rights, and every production system records the model, prompt, corpus and permissions behind an output. The same tribunal decision made supervision concrete, observing that a supervisor who fails to prevent a junior fee earner’s false citations is likely to be more culpable than one who fails to catch their own errors. Conflicts and ethical walls stay deterministic, so retrieve-then-filter architecture is a disqualifying finding. Generated language should be marked as generated until a lawyer adopts it, because a reviewer reading polished text applies less scrutiny than one reading a colleague’s rough draft.

Who pays for the verification?

A business case that counts hours saved in production counts one side of the ledger. Compare the cost of producing the work by hand with the cost of generating it and then checking it to the standard the work requires. Generation becomes nearly free, and the binding constraint moves to the senior lawyers whose time is the most expensive in the firm, so a resolving citation is what converts verification from re-performance into spot-checking.

The fee question is live elsewhere and open here. An American ethics opinion of July 2024 states that a lawyer may not bill for time saved while charging as though it had been spent. I could find no equivalent statement from the regulator or the representative body in England and Wales, which is itself a planning point for any firm on hourly rates.

What the paper argues is narrow, and it holds. Value sits wherever an authoritative source exists to check the output against, and exposure sits almost entirely on the other side of that line. Research against an approved primary corpus qualifies, as does clause extraction from a permissioned document set, while a confident answer about the law from a general model falls on the far side and the measured reliability figures explain why.

Underneath that principle sits a records discipline, and this is where legal AI governance either becomes real or stays on a slide. Verification should be a step in the workflow with a name attached, because an instruction to be careful has never survived deadline pressure. A matter file should show which authorities were checked, against what, and by whom. That record costs minutes, it answers a regulator or a court, and it is what makes law firm AI supervision demonstrable. The companion decision is to settle in writing which services may receive matter material, since client confidentiality and AI now meet at the point where architecture is decided.

Two questions remain genuinely open, and both are testable inside a single firm within two intakes. Whether these tools erode the reasoning capability of junior lawyers is unresolved, and I found no longitudinal evidence either way. Whether broader coverage in document review produces better outcomes or simply more findings to discount is equally unresolved, and a firm keeping its own record of what it decided and why will answer both long before the profession does.

Frequently asked questions

Does using artificial intelligence remove a solicitor’s responsibility for the work?

No. The Solicitors Regulation Authority stated in its warning notice of 17 August 2026 that artificial intelligence has no separate legal personality and that regulated individuals remain accountable for their work and outputs regardless of how that work has been prepared. Competence, confidentiality, supervision and candour to the court stay with the solicitor.

Is it safe to put client documents into a public chatbot?

A tribunal decision promulgated in November 2025 held that uploading confidential documents into an open tool such as a public chatbot places that information in the public domain, breaching client confidentiality and waiving legal privilege. The tribunal distinguished open public models from closed or enterprise systems held under contract, which makes the choice of service a technical control.

How often do purpose-built legal research tools invent authorities?

A study of 202 hand-constructed legal queries reported that one leading purpose-built tool answered accurately in 65 per cent of cases with a hallucination rate of 17 per cent, while a second answered accurately in 41 per cent of cases with a rate of 33 per cent. Both vendors disputed the findings.

What did the Divisional Court decide about fabricated citations?

Sitting in June 2025, a Divisional Court considered two matters together, ordered wasted costs of £2,000 in the first, and held that the threshold for contempt was arguably met while declining to begin proceedings as a matter of discretion. Five referrals followed, covering counsel, the instructing solicitor, a law centre and the firms involved.

What should a firm do first?

Publishing interim guidance within a fortnight usually matters more than choosing a vendor, because a shadow estate grows while a programme designs itself. That guidance should say which services may receive matter material. Naming the lawyer accountable for the verification gate is the second decision worth taking early, since that gate is what leaves a record on the matter file.