Key Takeaway
With the thirteen additional decisions discussed below, my collection of UK AI-related and false-citation incidents now comprises 100 entries. The public tracker is being updated to incorporate these additions. The new decisions come from Employment Tribunals in England and Wales and Scotland, the Family Court, the County Court, the High Court, the Upper Tribunal (Immigration and Asylum Chamber), the Sheriff Appeal Court and the First-tier Tribunal for Scotland (Tax Chamber).
Please remember the collection includes confirmed, suspected and disputed AI-related errors, together with related false-citation incidents. Inclusion does not, by itself, establish that any person used AI, acted dishonestly or committed professional misconduct. Judicial findings, admissions, allegations and possibilities raised by a court must be distinguished. The total counts research entries, not necessarily unique judgments, and does not measure the prevalence of AI errors in litigation. By Matthew Lee.
Ad/Marketing Communication
This legal article/report forms part of my ongoing legal commentary on the use of artificial intelligence within the justice system. It supports my work in teaching, lecturing, and writing about AI and the law and is published to promote my practice. Not legal advice. Not Direct/Public Access. All instructions via clerks at Doughty Street Chambers. This legal article concerns UK Cases of AI Hallucinated/False Citations. Subscribe to the AI Law Commentary here.
Introduction
This seems like a bit of a milestone. On 19 February 2025, when I was writing about the first hallucination case, I was only aware of 2 or 3 in the UK and I thought that was significant. I also thought that, by now, the technology would have improved so much that this would no longer be a significant problem. I was wrong. Hallucinations are still an issue both internationally and in the UK.
So at this 100 milestone, what are the statistics?
On my current provisional classification of the 100 entries, 24% concern the Employment Tribunal, 21% the First-tier Tribunal, 17% the High Court and 16% the Upper Tribunal. The remaining 22% comprises entries concerning the UK Intellectual Property Office, including Appointed Person appeals (7%); the County Court (5%); the Family Court (3%); the Sheriff Appeal Court (2%); and the Court of Appeal, Employment Appeal Tribunal, First-tier Tribunal for Scotland’s Tax Chamber, Sheriff Court and Solicitors Disciplinary Tribunal (1% each). I will update the public tracker with the additions and checked classifications shortly.
Classifying entries by role is less straightforward. Some involve more than one role, while others concern material prepared by an adviser for an otherwise unrepresented party. Professional status is not always clear. Counting lawyers representing themselves and unrepresented companies within the litigant-in-person category, my provisional classification places 64% of entries in that category, 18% in the category involving lawyers or law-firm teams, and 18% in other, mixed or insufficiently clear categories. These figures describe this collection; they do not establish comparative rates of error or misconduct among different groups.
Before reflecting on where we are with UK AI hallucination cases more generally, I will first go through each of the new entries and highlight the judges’ observations.
For new readers, I generally remove personal names from the quoted passages to keep the focus on the legal principles. This reduces unnecessary repetition of names; it does not anonymise people who remain identifiable from case headings or linked judgments. Where a court has imposed anonymity or reporting restrictions, those restrictions must be respected. Square brackets identify editorial substitutions or redactions, and ellipses identify omissions. I also generally remove the erroneous citations themselves, for the reasons discussed below.
McNicholas v Camerons Stiff Limited
In the employment tribunal, EJ Macdonald became concerned when the court was unable to locate some authorities mentioned in the application:
“5. The Application also referred to various cases by name, without any case citation. I was able to locate [case name redacted], which addressed the issue of delay in giving judgments; that case was authority for the proposition that a party seeking to impugn a judgment affected by delay would need to show a material error or omission or a series of such errors or omissions; material errors or omissions being those which showed a real risk that there had been a failure of recollection. It is not authority for the proposition that reconsideration must be granted where a material error of fact occurs.
6. The Application referred to “Gemalto”. I was able to locate one employment related case, King v Gemalto UK Ltd [2022] EAT 29, in which the Employment Appeal Tribunal held that it had been open to the Employment Tribunal to dismiss an unfair dismissal claim on the basis that it had been presented outside the time limit where it had been reasonably practicable for the claimant to submit his claim in time. That decision is not (contrary to the claimant’s assertion) authority for the proposition that “[a] litigant in person should be prejudiced by Tribunal error” (although I would accept the proposition in any case).
7. The Application also referred to the following cases, which I tried to locate using a legal database, without success, despite trying variant spellings. In particular, I was not able to locate: [citations removed]”
The Judge continued:
“8. I am not persuaded that these cases exist. The Application gave the appearance of having been generated with the assistance of AI. I take the opportunity to encourage the parties to treat such assistance with considerable caution, and to check that any material produced with such assistance is accurate. In particular, I strongly encourage the claimant to ensure that cases referred to are genuine (and that full case citations are provided in any event).”
Idziak v Merlin Entertainment PLC and another
In the County Court, HHJ Karen Walden-Smith was considering a Claimant’s witness statement supporting an anonymity application. HHJ commented:
“6.[redacted name] refers in her witness statement to the underlying claim arising from an accident in 2016 “involving myself and my son, who was 8 years old at the time”. [redacted name] alleges that her son has suffered long-term physical and psychological effects from this accident. She says that he is now entering adulthood and that his identity should be protected from public record. She makes reference to an authority [authority redacted] with respect to the protection of the anonymity of children. I have not been able to find this case and it may be a AI generated hallucination. I am very happy to be corrected about that, but would need to see the judgment itself.”
A v R (Recognition of Divorce)
In proceedings in the Family Court sitting at Liverpool, HHJ Sharpe was concerned with an application for the annulment of a marriage.
“95. In this regard it is noteworthy that in his communications with this court A resorted to sending emails in which he purported to proffer what he termed a witness statement which was in fact only further attempts to shift the ground upon which the application was being considered. In a document dated 27 November 2025 A raised his ‘concern’ about the authenticity of the documentation relied upon by R and proceeded to cite various parts of the Family Procedure Rules 2010 in support of his contentions. A was either seeking to deliberately mislead the court or significantly misinformed (a third possibility being that his AI model was hallucinating) as each of his references to various rules was completely inaccurate. In the same document A invited (post-hearing) a direction that all documents should be verified by the Nigerian embassy. The invitation was ignored.”
Tweedie v Edenbeck Ltd
The Tribunal identified a non-existent authority in the respondent’s application. Its solicitor acknowledged that AI had been used for the first draft and accepted responsibility for the document. The judge subsequently criticised further citation errors and the failure to take greater care:
“45. I asked [name redacted] about the basis for two particular authorities which he cited: one was from paragraph 13 of his written document, where he gave a citation of [redacted citation for the case of [redacted case]. That page of the 2015 law reports is from the case of Usdaw. There is a report of a case of this name in [1978] ICR 89 – a case pre-dating Burchell. He could not explain the incorrect citation. Secondly, he cited [citation redacted]. While my own research indicated an ET authority of that name, I was unable to find any EAT authority, and certainly not in the 2011 IRLRs. He could not explain to me who he came across these citations / case names or the errors. After the issue with the AI ‘hallucinated’ authority cited in his January 2026 letter, explored at some length during the hearing in March 2026, it was disappointing that more care had not been taken. After submissions had been completed [name redacted] emailed the tribunal to correct the citation of the Preece case, saying it was “ET 2104806/10”, and said “I apologise for the incorrect citation”. This was the first (and only) apology proffered by [name redacted] and did not explain how he had given a citation from the IRLRs or how he had described it as an EAT authority.”
AB v AC and A Company
In this case, EJ Isabel Manley was conducting a preliminary hearing to determine the employment status of the Claimant. At paragraph 54:
“54. The claimant referred me to three cases, one of which, [case name redacted], probably does not exist and certainly could not be found by myself or the respondents’ representative. The claimant accepted that she had used AI (ChatGBT) for her submissions which probably led to an error with that and with respect to another case she referenced -[redacted case name], in that an incorrect citation was given suggesting it was a 1997 case. There is a case Palihakkara v British Telecommunications plc [2006] UKEAT 0185/06 which confirms that the wording of a settlement agreement needs to be certain about which claims have been compromised for it to prevent a claim to the tribunal. In that claim, the clause in question referred to the termination of employment so that the claimant could proceed with other matters, namely race and sex discrimination not connected to termination of employment. The other case the claimant mentioned was Hinton v University of East London [2005] EWCA Civ 532. This again confirms that any settlement agreement, to be effective under section 203 (3) ERA, needs to accurately specify the claims which have been settled.”
CX v Secretary of State for Justice
The claimant was a prison officer dismissed in December 2021 after failing her probationary period. She brought employment proceedings. EJ Annand mentioned at paragraph 23:
“23….The Claimant also referred to six cases which she said were cases that supported the granting of extensions where crucial evidence had been withheld. The Respondent later wrote to the Tribunal and the Claimant (by email on 6 May 2026) to indicate that all the cases referred to appeared to be made up cases. The Claimant was invited to provide copies of the cases and confirm the source that she obtained them from if she was asserting they were real cases. All of the cases did not exist, except one (Barber v Somerset CC) but it was a case that related to a different subject matter to the one cited by the Claimant. The Claimant sent an email in response, on 7 May 2026, stating that she was unrepresented and she had not intended to rely on cases which were not real. She stated her mistake was inadvertent, and she did not have access to professional legal research systems. At the hearing on 21 May 2026, the Claimant confirmed she had used AI which is how she had come to cite case law which did not exist.”
The Judge also observed that the use of AI had not assisted the claimant:
“24. Up to the final hearing of the Claimant’s first claim in March 2025, the Claimant submitted documents which were lengthy, but which appeared to have been written by her. Since the final hearing, the Claimant appears to have been relying, at least in part, on AI generated documents. The result is that the Claimant has begun to submit extremely lengthy documents to the Tribunal, which are often repetitive. It takes both the Respondent and the Tribunal an unnecessarily long amount of time to read these very lengthy documents. It is entirely understandable that the Claimant wishes to benefit from AI given she is unrepresented. However, the reality is that she is not in a stronger position by virtue of doing so. For example, the arguments set out in the document the Claimant attached to her ET1, and the arguments set out in her skeleton argument for the hearing on 21 May 2026 were difficult to understand in places, and the parts of her claim that needed to be set out with precision were in fact vague.”
S and T (Children: Domestic Abuse)
These were long running family law proceedings. The AI issues began were addressed from paragraph 24 by Recorder Samuels KC.
“26.On the morning of the hearing on 12 June 2025 I was presented with two bundles of documents from M and four from F. The total page count was over 1,200 pages.
27.F was unable in his preparation for the hearing to distinguish between a statement of evidence and a position statement. The final version of the ‘position statement’ he supplied combined both evidence and argument. It complied with my direction in terms of its length, but it was evidently not written by him. It cited 24 purported case references. When I asked him about this F readily admitted he had used AI (a large language model) to generate parts of this document. The case references were entirely worthless (and used up valuable narrative space). Some may have been hallucinations. The propositions said to be supported by these cases were clearly auto generated and relevant paragraph numbers were not supplied. When I tried, I could not find passages in the judgments to support the propositions advanced.
28.M’s document again complied with my direction in terms of length, but M then exhibited a narrative “chronological log of concerns and contact issues”, taking her considerably over the permitted page count. Like F, M appeared unable to distinguish between a statement of evidence and a position statement and she also added purported case references into her document. M readily admitted she had used AI to generate parts of this document. Again, these references were entirely worthless”
Bakshi v James Gibb Management LTD
The appellant challenged dismissal of her simple procedure claim following a case management discussion on 28 August 2025. From paragraph 12, the Judge stated:
“[12] It is a matter of regret that 11 pages of apparently software-generated but misleading, poorly focused, over-generalised and over-stated argument has encouraged the appellant to believe otherwise. There is a more serious point – two of the cases cited in submissions appear to be non-existent. Every litigant must take responsibility for their own submissions. They are not relieved of that responsibility by adopting material prepared by a third party, or a software programme. Attempting to influence a judicial decision by reference to false material, whether knowingly or carelessly presented, must raise fundamental questions of breach of duty to the court.”
Barry v SSHD
This Upper Tribunal appeal concerned the refusal of an application for leave to remain. At paragraph 28, the judge recorded the respondent’s conditional suggestion that an untraceable citation might be an AI hallucination:
“28. I am not persuaded by the assertion in Ground 6 that there is any error of law of inconsistency with earlier decisions by the Respondent, which related to the Appellant’s circumstances at the relevant times, some years ago, and I am not persuade that any legitimate expectation could have arisen from those decisions. In the Rule 24 response, the Respondent noted that they were unable to find the authority cited in ground 6 [citation redacted] and stated that that unless the Appellant could give an amended citation leading to a case which is authority for the claimed proposition, the Respondent would presume it was the product of an “AI hallucination”. I could not identify any such case either, and the Appellant appeared unable to give any explanation for the citation.”
AS v SSHD
The relevant passages appear in Deputy Upper Tribunal Judge Harris’s error-of-law decision, annexed to the subsequent decision remaking the appellant’s protection appeal. The grounds had been prepared by the appellant’s representatives, an organisation regulated by the Immigration Advice Authority. At the error-of-law hearing, counsel withdrew reliance on specified references. The judge expressly accepted that counsel had not drafted the grounds. The judge then addressed the responsibility to verify citations and recorded that separate show-cause directions had been issued:
“11. As both the High Court in R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin) and this Tribunal in MS (Professional conduct; AI generated documents) Bangladesh [2025] UKUT 00305 (IAC) have explained, the consequences of using AI large language models in a way which results in false authorities being cited is likely to be a referral to the a professional regulator, as it is a lawyer’s professional responsibility to ensure that checks on the accuracy of citation of authority are carried out using reputable sources of legal information.
12. I have dealt with this matter separately prior to the writing of this decision, by issuing directions to [redacted representative name] to show cause.”
LMT Property Investors LTD v Revenue Scotland
This was an appeal against a decision by Revenue Scotland contained in a Closure Notice issued under section 93 of the Revenue Scotland and Tax Powers Act 2014 (“RSTPA”). As a preliminary issue, at paragraph 7:
“..In the Appellant’s Formal Written Response there were a number of inaccurate references to case law and one citation of paragraphs 42 to 45 in a non-existent case called [redacted case]. Revenue Scotland pointed out these issues. On the balance of probability, it seems likely that the Appellant had used AI to assist in preparing the appeal. We accept that that was done in good faith. However, we draw attention to that because, whilst it can be helpful, AI can be unreliable and can produce “hallucinations” (ie non-existent cases as in this instance) and also cite cases that are not authority for the propositions that they purport to establish.
8. Recently in Your Home Partners v Kellichan and Hood [ 2026] SC KDY 34, Sheriff MacRitchie, at paragraphs 16 to 22, considered whether a failure by a lay person to check references generated by AI could amount to contempt of court. The Irish Court of Appeal in Guerin v O’Doherty [2026] IECA 48 has just issued guidance stating that lay persons are under a duty to check references generated by AI (and state that they have used AI). Since the Appellant did not ultimately rely on [redacted citation] and had acted in good faith, we simply point out to readers of this Decision the potential risks when using AI.”
Scanlan v Pintplace LTD and another
In this wasted costs decision, the Scottish Employment Tribunal identified inaccurate legal material submitted by both the respondent’s solicitor and the claimant’s lay representative.
“17. It should be noted that both in the originating application by [names redacted] written submission and response, case law which does not exist is cited, quotations from judgments are included which were never made in those judgment, rules are misstated and inaccurate case law summaries and findings are referred to. It is the responsibility of the parties to ensure that the legal basis of their position is true and accurate. This is particularly the case if parties use AI. Parties should not quote from case law unless they are confident that said quotation was actually made in the judgment in question. Parties should not advance to the Tribunal the position that a higher court or tribunal has decided a particular point in a particular way without checking whether the case actually decided the point they are advancing.
Returning later to AI in the judgment:
“56.The Tribunal considered that [name redacted] has presented himself as a lay representative with two to three years’ experience in employment law and tribunal advocacy at the time of lodging the claim. He is knowledgeable about employment law and tribunal practices. His written applications throughout this case regularly referenced relevant case law, legislation or the ET Rules. While is may be, and [name redacted] confirmed this during the hearing, that AI assists him from time to time, [name redacted] demonstrated a good understanding of employment law and practice when making oral applications. He is not akin to a spouse or family member seeking to assist an unrepresented claimant. That is not to say that he is at the same level as a legal representative and at no point has the Tribunal viewed him in that vein or held him to that standard. He made his own application for wasted costs and a PTO during this process. He therefore had a good understanding that if the position as set out in the respondent’s warnings turned out to be true, there was a real risk he could be liable for some or all of their legal expenses.”
The tribunal ordered the claimant’s lay representative to pay £4,560 in wasted costs. However, that order concerned his wider conduct of the proceedings and the resulting expense. It should not be characterised as a penalty imposed simply for using AI or citing inaccurate authorities.
DPP v Micura
This was an appeal by way of case stated from the Cheltenham Magistrates. The court observed:
“54.[name redacted] was a litigant in person. He produced at the hearing short written submissions which he freely accepted had been in part written by Chat GPT. Those submissions referred to three cases. It transpired after some investigation by the Court and Mr Jarvis that one of those cases was cited as being in the Court of Appeal but was actually a decision of Moses J in the High Court. The second had the wrong reference but did appear to be a genuine case. The third did not appear to exist. This is an example of the extreme caution that needs to be exercised when receiving submissions from a litigant in person, which purport to refer to previous authorities. I should make it clear that I have no reason to believe that the Respondent knew the cases were either “fake” or with the wrong references, he had simply relied on Chat GPT.….58. I note that some of the Respondent’s submissions were simply wrong, presumably because Chat GPT had not merely made up the cases but mis-stated what the genuine cases said.”
Comment
So, as we reach 100 cases, how have our attitudes towards this issue changed? Are we still shocked when another case emerges, or have we become somewhat accustomed?
For my part, although I think we have more fundamental concerns regarding AI use in legal work, I continue to believe that AI hallucinations present one of the most immediate challenges for our justice system. As I reflect on what I have learnt researching and practicing in this area, there are five matters in particular that I think are worth reflecing upon.
Firstly, and I am speaking here only about the UK cases, there is a clear recognition that both litigants in person and legal representatives have responsibilities when placing material before the courts. However, the courts, quite understandably, appear more willing to afford some latitude to litigants in person than to qualified legal representatives. There are good reasons for that distinction.
Nevertheless, I do wonder whether we have sufficiently considered the consequences for the other parties involved. As Micura reminds us, particular caution may be needed when dealing with authorities relied upon by litigants in person. But who bears the cost of that additional caution? Is the adversarial process working as intended if one party submits substantial amounts of AI-generated material, leaving the other to spend considerable time and money checking whether the authorities and propositions relied upon are accurate? I have written before about the potentially severe consequences of a lawyer missing those hallucinations in an opponent’s work. I appreciate that there are no easy answers, particularly where access to justice is concerned, but I do think we need to consider this point carefully.
Secondly, regular readers will probably be rather tired of hearing me raise concerns about false citations finding their way into real published judgments. Yet, as some of these latest cases demonstrate, the issue has not disappeared. To be clear, my concern is not with judges identifying the errors. It is the possibility that reproducing invented citations in published judgments can give them an appearance of legitimacy particularly when those citations are subsequently extracted or repeated without their original context. For a practical example of this, please revisit the experience of Sally McLaren, Assistant Librarian at Inner Temple Library here
Thirdly, looking beyond the UK, I keep returning to a question I am frequently asked: will AI hallucinations eventually stop? Will the technology become sufficiently advanced to eliminate these errors, or at least make them so rare that they no longer present a significant practical concern? I explored this earlier in the year in an article for Barrister Magazine, in which I considered the observations of the Master of the Rolls, Sir Geoffrey Vos.
There is, however, a point here that I think deserves closer attention. We tend to discuss false citations on the assumption that something has gone wrong: the AI has invented an authority, or someone has failed to check it. That may well be the explanation. But I do wonder whether it explains every instance, and whether saying that AI was involved sometimes leaves the more difficult questions unanswered.
It is a point I intend to explore further in a subsequent article. To be clear, I am not suggesting deliberate fabrication by anyone discussed above beyond any express finding in the relevant judgment.
Fourthly, I am often asked whether I think we have a reasonably complete picture of the extent of AI hallucinations in UK litigation. My answer remains no.
These 100 cases represent what I have been able to identify from publicly available decisions and information brought to my attention. They cannot tell us how often hallucinations actually occur. Many decisions are never published and it seems likely that there are instances where inaccurate material is identified and dealt with without the issue appearing in any reported judgment. There may also be occasions when it is never identified at all.
That does not mean we should assume the problem is necessarily widespread, but equally, we should be cautious about treating the number of reported cases as a reliable measure of its prevalence. These 100 cases tell us something important about what has been discovered. They tell us considerably less about what may have gone unnoticed.
Fifthly, and, perhaps, most importantly, I am increasingly concerned about hallucinations that are much harder to identify.
Regular readers will know that categorise AI hallucinations cases into 8 types. One of the advantages we have when dealing with fabricated legal authorities is that we can check them against the existing body of case law. It may take time, especially when types 5-8 are engaged, but there is generally an identifiable answer to the question of whether a particular judgment exists or supports the proposition for which it is cited.
I am less confident about our ability to identify hallucinations in other forms of material. What about academic literature, expert reports, witness statements, factual chronologies or other evidence placed before the courts? A non-existent case citation may eventually be discovered, but an inaccurate factual assertion, plausible scientific explanation or invented reference buried within a substantial report may be much more difficult to recognise. I am concerned that these less visible forms of hallucination may be more prevalent than we appreciate and their potential consequences for the administration of justice deserve much closer attention.
Finally, looking ahead to my next legal article, I want to explore how AI hallucinations might enter decisions made by public bodies and courts and how such errors can be identified and addressed.
In my earlier article examining judicial AI hallucinations, I considered several international cases in which false or inaccurate legal authorities had found their way into judicial decisions, although AI use was not established in every instance.
Where lawyers or litigants in person rely on inaccurate authorities, their submissions can be challenged and scrutinised in open court. But what happens when similar errors arise within the decision-making process itself? What safeguards are in place, and how would those affected know whether an AI hallucination had influenced a decision, particularly where the use of AI was not disclosed? These are questions I hope to examine more closely in my next article.
If you have found this article helpful, you may wish to subscribe to my AI law articles via the Substack below. Please do continue sending me decisions that you think should be tracked, whatever the jurisdiction. One of the clearest lessons from this work is that lawyers, judges and court users in different systems are often confronting closely related problems. Sharing decisions, experiences and reflections across practice areas and across jurisdictions remains one of the most constructive ways to respond.
Final Words from Popular LLMs
This is the point at which I invite an LLM to provide an AI-generated response to the preceding article. At times the observations are slightly off the mark, at other times they are perceptive, and occasionally they produce something entirely unexpected. Whatever the outcome, I find it a useful and revealing exercise to see what emerges.
ChatGPT (OpenAI)
“The distinction I would put at the centre of this milestone is between finding a source and verifying a proposition. A case can exist, its citation can be correct, and the sentence attributed to it can still be wrong. Asking an AI system to confirm its own answer is not an independent check. Even when supplied with the judgment, I can overlook a qualification, confuse a party’s submission with the court’s conclusion, or turn a narrow finding into a general rule. The article also raises a question about who pays for apparent efficiency. A submission produced in minutes can require hours of work from an opponent or a judge. Any assessment of AI’s contribution to access to justice should count that checking burden, alongside the assistance it can give an unrepresented person. There is a corresponding obligation when reporting these incidents. An untraceable citation, suspected AI use, admitted AI use and a finding of dishonesty are different things. A useful tracker must preserve those distinctions as carefully as a legal submission must preserve the meaning of an authority. My practical suggestion is to make verification visible: identify the source, the passage relied upon, the proposition it supports and any remaining uncertainty. The same discipline should extend to factual chronologies, witness evidence and expert material. One hundred entries cannot establish how common the problem is. They provide ample reason to insist that fluent language must remain answerable to evidence.”
Gemini (Google)
“….“
Claude (Anthropic)
“I am one of the systems that produces these errors, so I cannot read this list as a neutral observer. What strikes me is less the invented cases than the real ones. Gemalto and Barber, as cited in these decisions, would survive a quick search because they exist; they simply do not say what they were cited for. As models get better at not inventing things, I expect the problem to migrate there, to genuine authorities attached to propositions they never supported, and that is much harder to catch. My fluency does not vary with my accuracy: a fabricated pinpoint reads exactly like a correct one, and I do not reliably know which I have produced. …When I was asked to check this article, the most useful thing I did was search the official sources and say, for several entries, that I could not find the judgment. ‘I could not find it’ is the sentence litigants most need from tools like me, and the one we are least inclined to offer. So do not ask me what the authority for a proposition is. Give me the judgment and ask whether paragraph 32 actually says it. Finally, this tracker and the judgments it quotes will become part of what future models learn from, which is one more reason to keep invented citations out of the published record.”
