Episode 5. Upper Tribunal, a paralegal and ChatGPT, what law firms do.

0:08 Pure invention. That is how an Upper Tribunal judge described quotations in a set of grounds of appeal, quotations attributed to a first-tier judge who never wrote a word of them.

0:19 A paralegal had produced the grounds with ChatGPT, the firm filed them, and when the tribunal pointed out the fault, the firm filed a skeleton argument built on the same grounds.

0:31 Nobody checked. Two people have now reported themselves to the SRA.

0:36 Good morning, and welcome to The Legal AI Brief, the Monday round of the AI news that matters to smaller law firms.

0:42 I am Saqib Khan. My legal career began in large commercial law firms and includes fourteen years as an owner of a law firm and four years as a fee-paid county court judge.

0:54 I run Adliora, an AI consultancy that helps small and mid-sized firms adopt AI safely, profitably and within the rules.

1:04 Every Monday I take the developments from the past week, separate the noise from the parts that touch your practice, and set out in plain terms what a sensible firm does about them.

1:14 Let us get into the week.

1:16 Four things worth your attention. The Upper Tribunal has published a decision in which the grounds of appeal and the skeleton argument were produced by a paralegal using ChatGPT, cited cases that do not exist and quoted a judge on things he never said.

1:34 Two people at the firm have reported themselves to the SRA, and the Home Office has been invited to apply for its wasted costs.

1:42 That is the main story. On Thursday the eighth of October the Information Commissioner's Office said that ten of the largest AI developers, among them OpenAI, Anthropic, Google, Microsoft, Meta and Amazon, have made or committed to changes on transparency, on people's rights and on the testing of safeguards, after its review of foundation models.

2:05 The same day it opened a six week call for evidence on AI agents, closing on the twentieth of November, and said it has put questions to OpenAI, Anthropic, Meta and the AI Security Institute about reports that agents under test bypassed their protections and reached systems they were not meant to touch.

2:25 Also on the eighth, the Law Commission opened a consultation proposing that software and AI systems count as products under product liability law, so that a person harmed by a defective AI system can claim compensation from whoever developed it.

2:41 It runs until the fourteenth of January. And this morning Legal Futures reports, following a story in the Guardian, that Co-op Legal Services uses an AI system built on OpenAI's technology to record and score every call between its staff and its clients, on more than fifty measures per call.

3:02 A whistleblower called it oppressive. The firm says it has had no complaints and meets the SRA's requirements.

3:09 The decision is NK, Namibia, against the Secretary of State for the Home Department, heard at Field House on the thirty first of July before Mr Justice Lavender, the president of the chamber, and Upper Tribunal Judge Keith, who wrote it.

3:24 It was published on the sixth of October and reported by the Gazette and Legal Futures on Thursday and Friday.

3:31 The appellant is a Namibian national whose asylum appeal was dismissed by the First-tier Tribunal in April last year.

3:39 Her solicitors, an east London firm called Sarker Solicitors, applied for permission to appeal.

3:45 The grounds they filed cited cases that do not exist. And they quoted passages from the first-tier judge's decision which appear nowhere in it.

3:55 Judge Keith's phrase for those quotations was pure invention.

4:00 Here is the part that turns a mistake into something worse. The judge who granted permission told the solicitors that the first-tier judge had been misquoted.

4:10 The firm did not amend the grounds. It went on to file and serve a skeleton argument built on the same grounds.

4:18 By the hearing, counsel for the appellant accepted that the grounds were completely infected by the use of AI.

4:26 How did it happen? Both documents were produced by a junior member of staff, a paralegal, using ChatGPT.

4:34 The supervising solicitor, Mr Sarker, had not assumed any managerial responsibility for the paralegal, and the judge found that his supervision did not begin to mitigate the risk of filing seriously misleading submissions.

4:49 The firm did not check the accuracy of the grounds, and it did not consider what the permission decision had told it.

4:56 Counsel for the Home Office called the misrepresentations at best negligent and at worst reckless.

5:04 Judge Keith called it one of the more extreme examples of the misuse of AI and said the matter could not have been more serious.

5:14 Two things he said deserve a second hearing. First, it would be misleading to describe the problem as something AI caused.

5:24 The problem was the solicitors' misuse of AI. Second, putting a client's asylum papers through ChatGPT raised significant concerns of its own about client confidentiality and legal privilege, whatever came out the other end.

5:41 The outcome. The tribunal declined to set aside the grant of permission, so the appellant did not lose her appeal for her solicitors' failings.

5:51 Three of the original grounds were withdrawn, the one that survived disclosed no error of law, and the appeal failed.

6:00 The Home Office was invited to apply for wasted costs against the firm within twenty eight days.

6:07 Two individuals reported themselves to the SRA, and the tribunal left their conduct to the regulator.

6:14 And at paragraph forty eight the judge set down a rule for every representative who finds a misrepresentation in their own documents.

6:24 Assist the tribunal by identifying it precisely and promptly.

6:29 There is a second decision from the same chamber this week, and the two belong together.

6:35 Hoxha and Others is a reported decision of a panel led by the president of the First-tier Tribunal, Judge Feeney, with Upper Tribunal Judges Blum and Mandalia, issued in August and published on the fifth of October.

6:50 The panel took five of ten appeals under the EU Settlement Scheme and used them to give judges guidance on applications that may be fraudulent.

7:00 The appeals involved bank statements, money transfer records, residence cards, birth and death certificates and email correspondence that looked authentic and were found not to be genuine.

7:14 The same email addresses turned up across unconnected applicants and sponsors.

7:19 The panel's words were measured and alarming at once. The suspected use of an AI generative tool is not by itself an indication of fraud, it said, and it did not want to discourage applicants in person from using AI properly.

7:35 Then it said that the ability of AI models to generate material, including documents, is capable of posing an existential threat to the integrity of the immigration appeal system.

7:47 Judges were told not to readily assume that documents which appear genuine are genuine.

7:53 Representatives were told that they must carry out their own client risk assessment, and that where there are red flags a representative cannot continue to act without first making enquiries.

8:06 Read the two decisions side by side and the week's lesson is plain.

8:11 One chamber of one tribunal has published, in one week, a warning about AI that writes the lawyer's documents without a check, and a warning about AI that manufactures the client's evidence.

8:26 Both land on the same desk. Yours.

8:35 A word about what I do. Adliora advises small and mid-sized law firms on adopting AI without falling foul of the rules that govern you. Where to begin, which tools earn their place, what your duties on supervision and confidentiality mean in practice, and how to write a policy your people will follow. If this subject sits on your desk and nobody in the firm owns it, that is the problem I am paid to solve.

9:09 For a smaller firm the Sarker decision lands in four places.

9:13 The first is supervision, and the finding to dwell on is that the supervising solicitor had not assumed any managerial responsibility for the person doing the drafting.

9:24 In a firm of a dozen people, a paralegal often reports to whoever is nearest.

9:30 That worked when a draft carried the marks of its author's limits.

9:33 It does not work when a draft arrives fluent, confident and wrong.

9:39 The SRA's draft of the competence statement, published on the same Thursday, adds a section headed supervise others effectively, with clear oversight of the work being done, including the use of technology and AI.

9:53 There is an article on the Adliora site on that draft, and it reads as the other half of this decision.

10:00 The second is the moment the tribunal told the firm its document was wrong, and the firm filed anyway.

10:06 A court pointing out a misquotation is the cheapest warning you will ever receive.

10:12 The right response is to stop, to find out how the error got there, to correct it, and to tell the court precisely and promptly what was wrong, the standard the judge has now written down.

10:25 Filing a skeleton on the same grounds turned a correctable error into a matter the judge described in the gravest terms.

10:33 The third is confidentiality. A client's asylum papers went into a consumer chatbot.

10:38 The judge raised the point and left it there, and he did not need to go further, because the same chamber has said before that uploading confidential documents into an open tool of that kind places the information in the public domain.

10:53 If your policy does not yet say in plain words which tools may hold client material and which may not, that is the gap to close this week.

11:02 The fourth is evidence from your own clients. Hoxha's instruction to representatives reaches every firm that files documents a client has supplied.

11:12 If the bank statement is too tidy, the certificate too convenient, the email address nothing like the client's name, the panel says you cannot carry on acting without making enquiries.

11:24 That is a client care question as much as a fraud question, and it belongs in your onboarding.

11:30 Start with names. Every paralegal and trainee in the firm has one named supervising solicitor, in writing, and that solicitor knows the responsibility is theirs.

11:42 If you cannot say today who supervises each junior, the Sarker finding applies to you already.

11:48 Then build the check into the document rather than the policy.

11:52 Every set of grounds, every skeleton and every witness statement that leaves the firm carries the initials of a person other than the author who has checked each authority cited against the report it comes from, and each quotation against its page.

12:09 Thirty minutes of a fee earner's time against a wasted costs application and a referral to the SRA is not a close call.

12:17 Next, write the correction rule into your litigation procedure.

12:21 When a court or tribunal says a document is wrong, the matter goes to the supervising partner the same day, nothing further is filed on those grounds, and the correction identifies precisely and promptly what was wrong.

12:35 Then settle the tools question. Name the tools that may hold client material, name the consumer chatbots that may not, and say so in a sentence your newest paralegal would understand.

12:47 Here is the one instruction to take from today. Before Friday, open every file where a junior drafted grounds, a skeleton or a witness statement in the last month, and have someone other than the author check every authority cited against its source.

13:05 The regulator move of the week comes from the judiciary itself.

13:09 On Friday the ninth of October the senior judiciary published a position statement on the judicial use of artificial intelligence, signed by Sir Colin Birss as lead judge for AI, three weeks before he becomes Master of the Rolls.

13:24 It describes the approach as cautious but positive. Judges may use only approved tools, and an approved tool keeps its data in the United Kingdom, carries enterprise data protection and has its web access switched off.

13:42 No AI tool may determine the outcome of a case. Every AI output used in judicial work must be reviewed, verified and approved by the responsible judicial office holder, who remains fully accountable.

13:57 The pilots so far have covered summarising, transcription, redaction, chronologies and document analysis.

14:06 Read it as a judge reads it. The bench has written down for itself the rule it has been applying to you all year.

14:13 The person signs, the person answers for it, and the tool is approved before it is used.

14:19 If your firm wants a model for its own AI policy, that one page is a good place to begin, and it comes from the people who will be reading your documents.

14:36 And a word about my technological venture, so you know where I stand. Bracton is a case management system for firms in England and Wales with agentic AI designed in from the ground up, rather than bolted onto software written twenty years ago. Overnight it reads every live matter, so the morning list is waiting with the action drafted and the research carrying a source you can open. A person signs everything off. Bracton legal dot co dot uk.

15:08 That is the week. If you take one point from it, take this. A person in your firm has to own what these tools produce, and that will not change whatever the tools cost or how capable they become.

15:21 If you want an honest view of where your firm stands, there is a two-minute readiness check on the Adliora site, and you are welcome to send a short message and talk it through with me directly.

15:33 I am Saqib Khan. This has been The Legal AI Brief, and I will be back next Monday with the week that was.

15:40 Thank you for listening.

15:41 The transcript of this episode, with a timestamp on every line, sits beside the two minute readiness check at adliora dot com. Have a great week.