Beyond Ayinde: testing AI legal research

21 Aug 2026|Opinion|Add your comment

Law firms need defensible methods for evaluating research systems

Lawyers remain strictly responsible for verifying any authorities, legal propositions, and submissions produced with artificial intelligence, a position firmly underlined by R (Ayinde) v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin).

Yet while Ayinde exposed the dangers of unverified, fabricated citations reaching the court, phantom cases are merely the most visible failure mode of AI-assisted research.

A previous article in Solicitors Journal examined the lessons of Ayinde and explained the consequences of allowing inaccurate AI-generated material to reach the court without adequate checking. The judgment provides a serious warning about professional responsibility, supervision and the basic obligation to verify legal authorities.

Confirming that an authority exists tests citation authenticity, but it does not establish that the authority supports the proposition attributed to it, that the system has synthesised the case law correctly or that material authorities have not been omitted. Those less visible failures are harder to detect because the answer may remain fluent, plausible and supported by genuine sources.

The question for law firms is therefore no longer simply whether lawyers should verify AI output. They must. The more difficult question is how a firm should determine whether a particular system is suitable for a particular form of legal research before lawyers begin relying on it.

Consider a system that correctly identifies a leading judgment on anticompetitive exclusivity rebates and then presents the economic test examined in that judgment as the legal framework that must be applied in every subsequent case. The authority is genuine and the answer sounds convincing.

Yet the system has overlooked a wider line of case law establishing that the test is not legally required in every case. Checking that the judgment exists would not reveal the error. The failure lies in the system’s synthesis of the law: it has converted one potentially relevant analytical tool into a universal legal requirement.

This is more difficult to detect than a fabricated citation because the answer survives a superficial verification exercise. The authority exists, the legal vocabulary is correct and individual propositions may appear defensible. What is missing is the relationship between the authorities and the qualifications that determine the scope of the rule.

The emergence of roles in legal research, legal engineering, knowledge management and AI governance reflects this wider problem. Organisations do not merely need lawyers who know how to prompt an AI system. They need professionals who can evaluate legal outputs, translate legal and professional duties into operational requirements, design repeatable verification processes and determine when human review is sufficient.

The growth of these functions suggests that AI-assisted legal research is moving from individual experimentation towards organised adoption. The methods used to evaluate the systems have not necessarily developed at the same pace.

Productivity claims create a related difficulty. Generation speed is often compared with the time a lawyer would have spent researching from the beginning. That comparison excludes the time required to turn a fluent, unverified answer into work that a lawyer can responsibly put their name to.

If the lawyer must locate every authority, correct the analysis, search for omitted cases and reconstruct the research, the system may have reorganised the work rather than reduced it. A firm that records time apparently saved without measuring time spent verifying is counting only one side of the ledger.

Nor can the quality of the output answer every question about professional use. A legally accurate response says nothing about where the information entered by the lawyer has gone, whether it has passed through an intermediary, how long it is retained or whether the same underlying model creates different confidentiality risks when accessed through different services. These matters require examination of the particular product, access route and data-handling arrangements.

Existing guidance properly states the duties that remain with lawyers: verify legal material, protect confidentiality and maintain appropriate oversight. It does not, and does not purport to, provide a method for evaluating a particular system against a particular legal task. General policies restating professional duties cannot fill that gap by themselves. What is missing is a way of testing a given system against a given task before a law firm relies on it.

Further research must therefore move beyond demonstrations, isolated accuracy tests and general warnings about hallucinations. It must examine what evidence firms should require before approving a system, how legal quality and research completeness can be assessed, when expert judgment is unavoidable, whether productivity calculations include verification time and how confidentiality risk changes across different forms of access.

Ayinde supplied a necessary warning. The unresolved problem is how the profession should translate that warning into a defensible method for evaluating the systems it increasingly expects lawyers to use.

Legal News desk contact: editorial@solicitorsjournal.com|

Yet while Ayinde exposed the dangers of unverified, fabricated citations reaching the court, phantom cases are merely the most visible failure mode of AI-assisted research.

Preparing for the private rented sector database

The new database will impose further compliance requirements on landlords and expand the scope of rent repayment orders

SPP praises proposal to unlock DB surplus

The Society of Pension Professionals has expressed support for draft regulations aimed at unlocking DB scheme surplus with a focus on protecting members

Five new solicitor advocates appointed

The Law Society of Scotland welcomes five new solicitor advocates granted rights of audience in a ceremony

Strait of Hormuz disruption: contractual risk and dispute strategy

Strait of Hormuz disruption is exposing contractual fault lines, making early risk assessment and dispute planning essential

Class actions filings reach decade low

Recent findings indicate a significant reduction in class action filings while settlements remain notable and shareholder claims increase

Sentence length rises for drug smuggler

A man involved in a drone-assisted drug smuggling plot had his prison sentence extended by over a year

Support for sexual abuse survivors increased

Victims and survivors of rape and sexual abuse are set to receive specialist care around the clock

Preparing clients for effective mediation

Solicitors should raise mediation early, time it carefully and prepare clients for realistic, effective negotiation

Court finds Serisys directors stripped insolvent company of software IP in Garden House Software v Marsh

Fancourt J rules directors of failed fintech group assigned valuable Adypt IP away from creditors at undervalue.

West London NHS Trust v Palmer: retired Broadmoor nurse loses human rights challenge to eviction

High Court rejects discrimination and proportionality defences, orders possession of hospital-owned home.

Angela Ritchie v The Royal Wolverhampton NHS Trust: widow barred from fatal accident claim after husband's lifetime settlement

County Court rules earlier settlement with product manufacturer extinguished any later claim against NHS trust.