In the past few weeks, there has been a trio of important updates on the use of AI in legal work, from CILEX (1 October 2026), the Law Society (21 October 2026), and the Civil Justice Council (29 October 2026). Jamie Smith KC and Helen Evans KC have been through all the materials, and in this article they explain the core themes that emerge.
No “one risk fits all” approach
One thing that the Law Society, CILEX and the Civil Justice Council documents all have in common is a growing appreciation that AI poses different risks in different contexts.
As the Law Society’s Guidance puts it: “Risks posed by AI are not fixed. They depend on the type of system, how it is accessed, the task, the data involved, the level of autonomy and the organisation’s governance and controls”. CILEX makes similar points, identifying varying risk levels for what it calls:
- Lower risk uses, such as administrative support, document summary, first drafts of routine communications or organising information;
- Higher risk uses, such as generating substantive legal analysis, advice, or material intended for courts.
For its part, the Civil Justice Council is more doubtful about being able to sort lawyers’ AI use into neat categories. It suggests that the courts should take what it calls a “functional risk” approach, with the focus being on how AI is being used and the likely harm that might ensue. These insights are a helpful reminder that when deciding what level of oversight is required by lawyers, much depends on the task and the potential consequences.
Are the courts letting lawyers off lightly?
One of the core issues under consideration by the Civil Justice Council was whether court documents prepared by lawyers- such as statements of case and skeleton arguments- should contain a declaration explaining what (if any) AI had been used to prepare them.
By contrast to the stricter approach taken to witness statements and experts, the Civil Justice Council decided that they should not. The principal reasons behind their decision seem to be as follows:
- Documents of this nature already have to be signed by their author. There should be no difficulty in identifying who has professional responsibility for their contents;
- The starting point is that the legal regulators rather than the courts should have principal responsibility for ensuring that AI use by professionals is appropriate;
- Requiring declarations as to AI use without adequate need could serve to encourage satellite litigation.
One striking feature of the numerous cases involving “hallucinations” in lawyers’ work so far is that they have tended to culminate either in a referral to the regulator or in a wasted costs order. Unless being asked to determine a costs application, courts becoming involved in policing AI use by lawyers they risk inappropriately treading on the toes of the regulatory bodies. For its part, the SRA stated in July 2026 that it had received 42 reports arising from alleged AI misuse by solicitors in the previous year. [1] The fact that declarations are not going to be required for court-facing documents such as pleadings or skeletons does not mean that they are not at the higher end of the risk spectrum. It seems to us that anything being submitted to court automatically sits in that higher bracket.
How much oversight is enough oversight?
One issue that the Law Society and CILEX both have some difficulty pinning down is how much oversight a lawyer has to exercise over AI’s output. This conundrum is allied with the fact that risks levels can vary.
CILEX’s Guidance makes clear that work produced by AI has to be “reviewed and verified before it is relied on, shared with clients or third parties or submitted to a court or tribunal”.[2] However it appears clear that varying levels of scrutiny may apply to various tasks. CILEX only says that an “appropriate” level of review should take place and suggests that it should be “proportionate to the nature and purpose of the task, the reliability of the tool and the potential impact of an error”. It points out that “particular care is required where AI-generated content forms part of legal advice or is provided to clients, courts, tribunals, regulators or other third parties”.[3]
The Law Society highlights that reserved legal activities can only be performed by authorised humans and that they should “apply appropriate level of human involvement and safeguards to mitigate risk, for both business of law and practice of law use cases, based on the level of risk posed by the particular AI use”.[4] It singles out the need to check documents that will be provided to a court with particular care, referring to checking citations, ensuring that legal arguments are properly reasoned and supported and that AI generated information has been checked against primary sources. [5]
Are there some tasks that AI just cannot be allowed to do?
As widely expected, the Civil Justice Council has made clear that AI should not be used in a substantive way in the preparation of witness statements of fact.[6] In particular it is concerned about AI being used to “generate” or “influence” the substantive content of such statements.
To guard against this, it recommends that a witness of fact should be required to make a specific declaration that the statement has not been prepared through use of an AI tool that has generated or influenced its substantive content.
The explanation for taking such a strict approach to witness statement of facts is the need to secure the integrity of that evidence and ensure that it is in the witness’s own words. The restriction on AI use can therefore be seen as part of broader innovations in the court rules in recent years, including the introduction of CPR PD57AC and the increased emphasis on a witness’ evidence being their own recollections.
A more relaxed stance is taken to other types of witness evidence. For instance, procedural witness statements by solicitors are to be treated in the same way as skeleton arguments or pleadings.
Unpacking experts’ work
The Civil Justice Council report makes clear that expert evidence also falls to be treated more strictly than lawyers’ work products. This is because, unlike pleadings or statements of case, the opposing party needs to be able to understand and test the way that the expert has put their evidence together. AI poses particular dangers to the process of preparing expert evidence in terms of cutting corners, excluding evidence that undermines the instructing party’s case or putting a partisan spin on evidence.
Some of these dangers were very apparent in two American cases cited by the Civil Justice Council.
The first, Kohls v Elison (2025) WL 66514 concerned an expert witness whose report included two citations of non-existent articles. It was therefore similar to many of the lawyers’ “hallucination” cases. The expert admitted using GPT-4 to assist in drafting his report but sought to defend his work on the basis that the underlying propositions were correct even if the citations did not exist. Just as the King’s Bench Division dismissed that type of argument in R(Ayinde) v LB Haringey [2025] 1 WLR 5147, it also fell flat in Kohls. The court did not fault the expert for using AI in the first place, but for failing to check his work adequately. It decided to exclude the expert’s evidence as a consequence.
The second, Mathis, Kyle & Ors v 3M Company & Ors (11th District Court, Harris County, Texas), was a jury trial arising from an explosion where an expert gave prompts to AI to produce an unbalanced report. According to news reports[7] these prompts included: “I need to show that 3M met the standard of care and why” and “show how 3M is 0% at fault for the explosion”. Clearly, AI was not being asked to approach the task with the level of dispassion required of an expert.
Just as the USA led the way in terms of hallucination cases, we can expect similar problems with expert evidence to emerge in England & Wales, making the Civil Justice Council’s intervention a timely reminder that AI cannot detract from experts’ responsibilities.
Conclusions
If there is one clear message that these recent reports and guidance give to the legal profession it is that the answer to the question, “What amounts to safe and responsible use of AI?”, is “It all depends on what you are using AI to do and what type of document you are preparing”.
At one extreme, witness statements of non-lawyers prepared for court proceedings must not be the product of AI in any substantive way. At the other extreme, routine administrative tasks may be performed by AI with no or minimal personal supervision.
For all those situations in between, the better questions for the lawyer to ask are:
- To what extent it is necessary for me to take personal responsibility for the accuracy and propriety of the overall work product?”; and
- “In view of the extent to which and manner in which I have used AI to assist me in my task, what steps are needed for me to assume that degree of personal responsibility?”
Herein lies the paradox for litigation lawyers producing court documents. On the one hand, we have access to these remarkable ‘thinking machines’ that can absorb vast amounts of data (factual and legal) and produce seemingly polished pleadings and skeleton arguments. On the other hand, how do we take due personal responsibility for the work product without independently doing all the work ourselves?
© Jamie Smith KC and Helen Evans KC, 4 New Square Chambers, November 2026
This article is not intended as a substitute for legal advice. Advice about a given set of facts should always be taken.
[1] SRA | SRA cautions profession about safe and responsible use of AI in legal sector | Solicitors Regulation Authority
[2] Para. 9.
[3] Para. 13.
[4] Para. 3.4
[5] Para. 4.2.
[6] See also the excellent article by colleagues Anson Cheung and Ryan Fincham about the dangers of AI use for witness statements in “AI and Witness Statements: Risks, Retrials and Reforms”.
[7] ‘Show How 3M Is 0% at Fault:’ Expert Witness Used ChatGPT to Write Report Defending Company in Deadly Explosion Lawsuit
Expert Asked ChatGPT ‘Show How 3M Is 0% At Fault’ Making For Entertaining Deposition – Above the Law
