Jamie Smith KC and Tim Chelmick consider yesterday’s Warning Notice published by the SRA and its timely reminder to the solicitors’ profession of two particular risks arising from the use of AI when providing legal services.
Hallucinations
Generative AI (or GenAI) tools of the type being used by legal practitioners rely at heart on a process of prediction: identifying a sequence of word(s) that probably is most apt to respond to the user’s input. As such, they are quite capable of making things up – so-called ‘hallucinations’.
Self-evidently, hallucinations are to be avoided. From the client’s perspective, they may receive bad advice. When it comes to court documents, there is the risk of the courts’ being misled and justice undermined.
In bringing the risk of hallucinations overtly to the attention of regulated persons, the SRA (it is to be hoped) is doing no more than telling solicitors what they already know – courtesy of widely publicised cases such as Ayinde and Cork.
The valuable aspect of the SRA Warning Notice is rather in identifying the shortcomings in policies and practices operated by legal firms that result in a failure to identify hallucinations; and, thereby, what good practice looks like.
In particular, the SRA places a focus on:
- The need for individual practitioners to take personal responsibility for the quality of their work product, i.e., carefully and conscientiously to check AI-generated work product for its accuracy.
- The need for firms to ensure that suitably qualified and experienced individuals are undertaking work.
- And/or, the need for firms to put in place an effective system of supervision to ensure that, before the final work product is released, it is subject to proper scrutiny by a suitably qualified and experienced person.
By way of a gloss, we would make two points. First, it is arguably not apt for the SRA to refer to the existence of hallucinations as evidence of “the misuse of AI”. As we have explained above, hallucinations are (at the moment, at least) an inevitable risk of any use of AI. Rather, if hallucinations survive into the final work product, it is evidence that the individuals in the law firms are not using AI responsibly and/or that the law firm has weak systems that fail to identify and manage shortcomings of the current generation of AI tools.
Our second point leads on from the first. We would suggest that firms should focus on adopting a culture that recognises the risks inherent in GenAI, views AI-generated work product with due scepticism and demands authentic and diligent review of such work product. ‘Tone from the top’ is very important here. It is the responsibility of senior practitioners, who have spent much/most of their working lives without AI, to hand down good habits to junior colleagues, who may otherwise rely on the accuracy of AI to an unhealthy degree.
Client confidentiality
The second particular risk identified by the SRA Warning Notice concerns client confidentiality, namely that allowing GenAI tools to access client data runs the risk of impermissible use of the data and loss of confidentiality (and, thereby, potentially loss of client privilege).
Standing back from the detail, one striking feature of the Warning Notice is that the SRA do not prohibit the use of GenAI to analyse client data. To the contrary, the SRA refers to the need appropriately to mitigate the risk of breach of client confidentiality. This will be most welcome to the profession (and to members of the Bar, who are presently grappling with materially similar issues).
To the extent that the SRA warns practitioners about allowing open source GenAI tools to access client data, we agree with the analysis – see our article on the distinction between ‘open’ and ‘closed’ systems here[1]. In short, don’t do it!
More nuanced are three interlinked aspects of the Warning Notice.
First, practitioners must not unthinkingly assume that ‘paid for’ GenAI legal tools provide sufficient guarantees to preserve client confidentiality. The devil will be in the detail of the contractual terms and conditions, e.g., does the foundational model (LLM) provider reserve the right to use the data to train the system? This needs specific due diligence and practitioners need to be satisfied that the contractual arrangements with the GenAI provider mean that the system can be treated as ‘closed’.
Second, a difficult issue arises for those legal firms that are building in-house GenAI systems. On the one hand, such systems may be viewed as more robust when it comes to the data journey, e.g., because the firm partners with a foundational model provider (and thus has close control over the contractual arrangements) and/or houses the LLM within a ringfenced cloud. However, what if the aim is to use client data to improve the tool? Questions then arise as to whether Client A has given the requisite consent to allow the firm to use the AI tool when providing services to Clients B, C and D etc. So, too, unexpected problems may arise when it comes to client conflicts and information barriers.
Third, it is necessary to consider the situation of the in-house lawyer operating within a non-legal entity. The entity may have GenAI tools that are not designed for legal work. Also, a conflict may arise between the commercial interests of the entity (in allowing the tool to access certain data) and the professional and ethical responsibilities of the in-house lawyer. An example might be where the entity is engaged in litigation with a competitor. The in-house lawyer would need to consider carefully whether it was appropriate to allow the entity’s AI tools to have access to materials provided by the competitor during the litigation.
Summary
As we see it, there is no stopping the momentum towards ever-great use of GenAI[2] by members of the solicitors’ profession (and the Bar). The SRA Warning Notice identifies two major risks, albeit there are others. If we were to identify one takeaway, it is that the safe and effective use of AI is a firm-wide issue: it is a question of good practices, systems, culture and ‘tone from the top’.
[1] Legal Professional Privilege in the Age of AI: Open Systems, Closed Systems and the Data Journey
[2] The SRA does not address Agentic AI (and nor have we), but the risks are no less and likely greater where AI tools/systems are being designed to achieve outcomes rather than merely respond to prompts.
© Jamie Smith KC and Tim Chelmick, 4 New Square Chambers, August 2026
This article is not intended as a substitute for legal advice. Advice about a given set of facts should always be taken.
