Two cases, same tool, different outcomes. Artificial intelligence (“AI”) generative tools are the talk of the town, forcing courts and legal professionals to grapple with how to regulate and responsibly use a technology that, only a few years ago, was not part of everyday legal practice.
Within the span of a few weeks, the courts have seen both the promising potential and the serious risks of AI. On one side, Garfield Law Ltd celebrated what is claimed to be the world’s first victory for an AI-powered law firm, successfully recovering £7,000 for a freelance consultant at Wandsworth County Court. On the other, Pinsent Mason LLP had to refer themselves to the Solicitors Regulation Authority after its AI-generated research relied on a statutory power that did not exist, resulting in an investigation by the High Court (Cork v Smith).
The timing of these two cases creates a striking juxtaposition. It captures, almost perfectly, the current uncertainty surrounding AI in legal practice. AI is often presented as a shiny new tool that everyone is eager to try, but these cases remind us that it remains just that: a tool.
AI is a form of Legal GPS
A useful way to think about AI is a form of legal “GPS”. It may suggest a route, organise information and help users reach an answer more quickly. However, it should not be mistaken for the destination itself. The user still needs to check the route, read the signs and decide whether the direction being suggested is legally sound.
The risk is that users may treat the prompt box as though it is a search engine, assuming that it is retrieving factual information from reliable databases. In reality, free generative AI does not work in that way. It predicts language based on patterns in its training data. It may know what a legal statute sounds like and generate something that appears highly convincing, even where the authority itself does not exist. Meaning, without proper supervision, the legal “GPS” may identify a road that is not really there. If the user follows it without checking, they may find themselves driving straight into the lake without realising. Once the error has reached the court, the opportunity to apply the brakes may already have passed.
The positives:
The reported Garfield case is a useful example of how AI may be implemented in legal practice in a constructive way. Garfield Law Ltd was authorised by the SRA in May 2025 and has been described as the first law firm authorised to provide legal services through AI. Its model focuses on lower-value debt recovery claims up to £10,000, where the cost of traditional legal support may sometimes exceed, or come close to exceeding, the amount being pursued, Garfield’s use of AI helps lower costs significantly.
It is reported that in just over a year they have been able to take on 600 claims and resolved more than £500,000 for their users. That kind of scale is only possible because the AI is doing all the heavy lifting and drastically reducing the human friction required to process a claim.
Yet, this is victory where we can see AI have a meaningful role. For individuals and small businesses, pursuing a modest unpaid invoice can feel commercially unrealistic if legal costs quickly outweigh the value of the claim. The result is that some valid debts are never pursued at all, not because they lack merit, but because enforcing them is too expensive. This creates a real access to justice gap, where legal rights exist in theory but are difficult to enforce in practice.
The facts of the Tamires Camal Taquidir case also show where the limits of AI still lie. Ms Taquidir, a freelance HR consultant, used Garfield’s AI platform to pursue a £7,000 claim for unpaid fees. The platform reportedly assisted with the claim from the pre-action stage through to trial preparation, including generating the letter before claim, preparing and issuing the claim, responding to the defence and counterclaim, assisting with the direction’s questionnaire, drafting witness statements and preparing the trial bundle.
However, despite that level of AI assistance, the case was not conducted by AI alone. In readiness for the trial, Garfield instructed a junior barrister at One Essex Court, Dominic Li to conduct the advocacy. That is an important detail. If AI could simply replace legal professionals, then in theory the platform could have produced a script, and the claimant could have presented the case herself.
The dispute reportedly turned on the existence and terms of an oral agreement, which is exactly the type of issue that requires judgment, responsiveness and advocacy. A prepared script may help organise what needs to be said, but it cannot assess a witness in real time, respond to a judge’s concerns, adapt to an opponent’s submissions or make tactical decisions as the hearing unfolds. The Garfield case therefore demonstrates both the benefit and the boundary of AI: it may assist with preparation and efficiency, but legal judgment, strategy and advocacy remain human responsibilities.
The negatives:
The cautionary counter-narrative to Garfield’s success is found in the High Court judgment of Malcolm Cork & Anor v Smith [2026]. The case involved what should have been a routine insolvency block transfer application, but it became a warning about the risks of relying on AI-generated legal research without proper verification.
In this instance, a junior lawyer used an AI pilot to research a court’s power to grant a liquidator’s “release” during a block transfer application. The AI provided a seemingly perfect answer, citing Insolvency Rule 12.37(5) and even providing a detailed, italicised “quote” of the rule. The problem was that the text did not exist. It was a pure “hallucination”, a fabricated legal provision generated by the AI’s pattern-matching algorithms.
This was not merely an incorrect citation or a wrong turn in the research process. The AI had mapped out a road that did not exist. The danger was not that the answer looked obviously wrong. The danger was that it looked right, sounded and was formatted as if it were real, and thus, appeared confident enough to be relied upon. That is precisely what makes AI hallucinations so risky in legal practice.
What makes the case even more significant is that the warning signs were there. The AI tool reportedly indicated that its answer should be checked against primary sources before being relied upon. In other words, the GPS had effectively warned the driver to check the map. The failure was not only that AI produced inaccurate information, but that the human verification stage did not prevent that information from reaching the court.
This is where we see automation bias arising. Information can feel more reliable simply because it is well-presented, confident and immediate. However, legal professionals cannot allow presentation to replace verification. AI-generated material must be checked carefully against primary sources, and professional judgment must remain at the centre of legal work. It can be argued that removing the friction, removes the brakes of the legal practice which would have helped avoid this issue from ever occurring.
The High Court’s response makes clear that AI-generated errors will not be excused simply because they come from a new or developing technology. A lawyer remains responsible for the material submitted to the court. If a submission relies on a case, statute or rule, it must be checked. The duty to the court cannot be delegated to a machine. This was seen earlier in the year where the Upper Tribunal in Munir also addressed the risks of AI-generated legal material after incorrect citations were placed before the Tribunal in two cases. The Tribunal reiterated that lawyers must carry out rigorous checks on the accuracy of AI-produced work and warned that failures to properly supervise the use of AI are likely to result in referrals to the SRA.
This leaves an important question. If AI can help bring thousands of small claims before the courts, while simultaneously generating false legal material that requires costly regulatory and disciplinary proceedings to correct, will it truly democratise the justice system, or risk overwhelming the very courts it is trying to assist?
That tension may become one of the defining legal challenges of the coming decade. Until the dust settles on this technological Wild West, the safest approach remains the simplest one: use AI as a tool but always check the source.
