Getty Images (US), Inc & Ors v Stability AI Ltd [2025] EWHC 109 (Ch)

[2025] EWHC 109 (Ch)

Case No IL-2-023-000007

IN THE HIGH COURT OF JUSTICE

BUSINESS AND PROPERTY COURTS OF ENGLAND AND WALES

INTELLECTUAL PROPERTY (ChD)

Venue The Rolls Building, 7 Rolls Buildings, Fetter Lane

London, EC4A 1NL

Date Thursday, 1st May 2025

Before

MRS. JUSTICE JOANNA SMITH DBE- - - - - - - - - - - - - - - - - - - - -

Between

(1) GETTY IMAGES (US), INC.Claimants

(a company incorporated under the laws of theClaimant

State of New York)Claimant

(2) GETTY IMAGES INTERNATIONAL UCClaimant

(3) GETTY IMAGES (UK) LIMITEDClaimant

(4) GETTY IMAGES DEVCO UK LIMITEDClaimant

(5) ISTOCKPHOTO LPClaimant

(6) THOMAS M. BARWICK, INC.Claimant

STABILITY AI LTDDefendant

MS. LINDSAY LANE KC and MS. JESSIE BOWHILL (instructed by Fieldfisher LLP) appeared for Claimants MR. HUGO CUDDIGAN KC and MR. EDWARD CRONAN (instructed by

Bird & Bird LLP) appeared for Defendant

Approved Judgment

Transcript of the Stenograph Notes Marten Walsh Cherer Ltd.,

2nd Floor, Quality House, 6-9 Quality Court, Chancery Lane, London WC2A 1HP.

Telephone No: 020 7067 2900. DX 410 LDE

Email: info@martenwalshcherer.com

Web: www.martenwalshcherer.com

[1]This is the claimants’ application dated 15 April 2025 to rely on a statement of case on development and training, which I will refer to as “the SOCDT”, served on 4 April 2025, designed to particularise their case in so far as it relates to allegations made in paragraphs 43 and 47 of the Re-Re-Re-Amended Particulars of Claim.

[2]The application comes on before the court with 27 working days to go to trial. The draft SOCDT is 91 paragraphs and 28 pages long.

[3]Owing to the urgency, I agreed following a full day's hearing yesterday, to give an oral judgment today. This inevitably means that the judgment will not mention all the evidence and submissions, but I have taken them all into account in arriving at the decision that I have reached. Introduction

[4]The background to these proceedings is amply set out in my earlier reported judgments on case management issues and I do not need to repeat it here. Suffice to say that this is a highly complex and technical case raising numerous novel issues for consideration by the court at trial which is due to commence on 9 June 2025.

[5]The current draft list of issues runs to some 63 issues for trial and the parties should by now be well into their preparations for that trial. The PTR is due to take place on 22 May and the skeleton arguments are due to be exchanged in 21 working days from today.

[6]The nature of the case and the issues arising have required the court to devote unprecedented amounts of court time to dealing with various disputed issues that have already arisen. This is the fourth hearing to take place this year in as many months.

[7]This application was first made on 20 March 2025 and raised before me at a case management conference on 28 March 2025. However, in the absence of a draft SOCDT at that time, I declined to grant an open-ended permission to the claimants but indicated instead that the application would need to be renewed once a draft SOCDT was available. This appeared to me to be particularly important in circumstances where the trial in these proceedings is fast approaching and it was essential that the defendant and the court should have a proper opportunity to consider the potential consequences of the provision of a draft SOCDT before addressing the question of whether permission should be granted.

[8]The application is supported by the third witness statement of Mr. McKenna, who characterises the SOCDT as a helpful pleading, prompted by a suggestion from the defendant made on 26 February 2025. He contends that it raises no new allegations. Mr McKenna also provided a fourth witness statement in response to the fifth witness statement of Mr Aries served by the defendant in opposition to the application.

[9]Consistent with Mr. McKenna's evidence, it is Ms. Lane KC's submission, on behalf of the claimants, that none of the rules applicable to late amendments apply to this proposed pleading, because it is not an amendment but a voluntary pleading provided purely to be helpful, in circumstances where the defendant had asked for such a pleading and had even gone so far, in a letter of 11 March 2025, as to identify what should be included in that pleading.

[10]Ms. Lane accepts that the SOCDT is a new pleading, but she says that it falls within the scope of the existing pleaded case and that she could just as easily have cut and pasted this document into her opening or closing submissions at trial. Accordingly, she submits that the primary question for the court on this application is whether this pleading and any response to it will be of utility to the parties and the court at trial. If so, then she says that permission should be granted.

[11]Given her stance as to the nature of the proposed pleading, Ms. Lane did not address the court on the rules in relation to late or very late amendments and she barely touched on the issue of prejudice to the defendant that may be caused by reason of the SOCDT being permitted. There was no discernible attempt to gainsay the defendant's evidence as to the steps that it would want to take if this pleading were now to be permitted and no attempt to suggest that those steps could in fact be undertaken in the time between now and the start of the trial.

[12]Not only do the claimants adopt this position because they consider their pleading to be voluntary and not strictly necessary, but also because they say that the defendant should have seen all of these allegations coming. I shall return to that point shortly.

[13]The defendant strongly opposes the application and relies upon the detailed and lengthy fifth witness statement of Mr. Aries in so doing. In short, it contends that the proposed SOCDT is replete with material facts which must be pleaded if they are to be relied upon at trial and which therefore require permission to amend. It says that the rules as to late and very late amendments apply, that there is no good reason for this proposed late amendment and it contends that it would suffer serious prejudice if the majority of the amendments are admitted.

[14]Neither party suggests that the June trial should be postponed in order to accommodate the work that would be required on the part of the defendant if the SOCDT is permitted. It appears to be accepted on all sides that the trial date must be maintained, a view which I heartily endorse.

[15]Against that introductory background, I observe that the positions of the parties to this application are quite plainly worlds apart. Owing to the claimants' position, it appears to me that I must begin by grappling with the status of the proposed pleading and thus the appropriate approach to be taken to this application. The status of the SOCDT

[16]In my judgment, the claimants' Re-Re-Re Amended Particulars of Claim pleaded, as Mr. Cuddigan KC for the defendant submitted, little more than a placeholder case in relation to the claim on training and development. Thus paragraph 43 of the Re-Re-Re Amended Particulars of Claim reads as follows:

"The Defendant's development team includes a significant number of individuals who reside and work in the UK, who will be identified upon receipt of appropriate confidentiality undertakings, including the Defendant's lead generative AI developer. It is to be inferred that, during the development and training of Stable Diffusion, visual assets and associated captions, used for the development and training of Stable Diffusion .... including Visual Assets (including Copyright Works) and Captions from the Getty Images Websites, were downloaded on servers and/or computers in the UK. Pending the provision of disclosure and/or evidence, the aforesaid inference is made based on the following facts and matters."

[17]Pausing there, the particulars that follow then focus on the location of the defendant’s human and computer resources, including Russian and Ukrainian developers employed by the defendant in the UK but, as is accepted by the claimants in their skeleton for this hearing, say nothing whatever about any specific activities of the defendant. Essentially this paragraph asserts an inferential case of downloading of images by reason of the defendant employing developers within the UK.

[18]Paragraph 47 provides particulars as to the training of Stable Diffusion more generally but there are no similar particulars in relation to its training and development in the UK, and I emphasise in particular for these purposes the word "development". There are no particulars whatever as to what is meant by “development” in this pleading.

[19]The pleading at paragraph 43 was entirely justified prior to disclosure because all the relevant information as to the defendant’s activities was held by the defendant. But, following disclosure, if there were further material facts arising out of that disclosure on which the claimants wished to rely in support of their case on training and development, it seems to me that they were obliged to identify those facts in a properly particularised pleading.

[20]I reject Mr. McKenna's evidence that the SOCDT contains only matters that could properly have been addressed in opening or closing submissions at trial. This evidence was patently mistaken, as the claimants have effectively recognised by the service of a revised draft SOCDT in advance of the hearing which removes from the original draft obviously new allegations as to the viewing and caching of images said to amount to acts of infringement. The pleaded case on development and training has always concerned only the downloading (and thus also storage) of images in the UK.

[21]Although it is true that a couple of sections of the SOCDT, to which I shall come in a moment, may be characterised as largely helpful, regrettably, I consider Mr. McKenna's evidence to be inaccurate in other respects also.

[22]As the defendant correctly points out, the SOCDT seeks to introduce reference to entirely new datasets never previously identified in the pleaded case (which refers only to LAION datasets), new technical issues regarding the use of Notebook software packages, allegations regarding the infringing activities of new individuals never previously identified, and allegations of copyright infringement by a third party, apparently known as Encord.

[23]While in one sense Ms. Lane is right that these are matters which fall within the very broad scope of the development allegation, they are not matters which could have been anticipated by defendant or which it could have been in any sense expecting, having regard to the very general pleading as it currently stands. Furthermore, they are not matters which amount purely to a plea of evidence or to a general factual narrative which need not be pleaded.

[24]In my judgment, even without the allegations of viewing and caching, the SOCDT plainly contains allegations of material fact, to which I shall return in more detail in a moment, which could not fairly be dealt with without warning, without disclosure and without the opportunity for responsive, factual and expert evidence from the defendant. The pleading of these material facts is, in my judgment, essential for the orderly and fair progress of the case. See in particular Boake Allen Limited & Others v HMRC

[2006] EWCA Civ 25

, at [131], where Mummery LJ said this:

"While it is good sense not to be pernickety about pleadings, the basic requirement that material facts should be pleaded is there for a good reason, so that the other side can respond to the pleaded case by way of admission or denial of facts, thereby defining the issues for decision for the benefit of the parties and the court. Proper pleading of the material facts is essential for the orderly progress of the case and for its sound determination. The definition of the issues has an impact on such important matters as disclosure of relevant documents and the relevant oral evidence to be adduced at trial. In my view, the fact that the nature of the grievance may be obvious to the respondent, or that the respondent can ask for further information to be supplied by the claimant are not normally valid excuses for a claimant's failure to formulate and serve a properly pleaded case setting out the material facts in support of the cause of action."

[25]Given the approach taken by the claimants, I should add that the entirety of the case that is now sought to be advanced in the SOCDT could not possibly have been relied upon at trial purely on the back of an allegation that the pleaded case was advanced "pending the provision of disclosure and/or evidence" as I understood the claimants' case to be, at least at the last hearing. That is no more than a plea that at an early stage in the proceedings it has not been possible to provide the particulars that might later be available. It is not a "get out of jail free" card which enables a party to run whatever case it likes on the evidence at trial.

[26]Equally, I reject the suggestion that the entirety of the case as set out in the SOCDT could have been deployed in opening or closing submissions without any further pleading. In so far as it seeks to identify material facts, this would have been little more than an attempt to ambush the defendant, which is entitled to understand the case it must meet at trial and to have a proper opportunity to answer that case.

[27]In the circumstances, in my judgment, the usual rules as to amendments, including as to late and very late amendments, must therefore apply to this application in respect of the SOCDT, at least in so far as it seeks to plead material facts, and I reject the claimants' submissions to the contrary.

[28]In pleading out the particulars of their case on development and training in the UK, as currently identified in broad terms at paragraph 43 of the Re-Re-Re Amended Particulars of Claim, the claimants cannot rely on the service of an entirely new pleading to escape the usual rules on amendment, nor can they realistically contend that this pleading in its entirety is purely helpful but not, in any sense, necessary.

[29]In substance what the claimants are doing is seeking to amend their inchoate inferential case. In so doing, they do not of course need to plead inferences that might be drawn at trial (although Mr. Cuddigan has put down a clear marker by reference to authority as to the need for notice of any such inferences) and they do not need to plead each and every disclosure document on which they intend to cross-examine witnesses or to rely upon at trial.

[30]In this regard, Ms. Lane was correct to draw my attention to the Chancery Guide at paragraph 4.2(e) as to the need for special care in setting out only those factual allegations which are necessary to establish the cause of action. However, it is my firm view that this proposed pleading clearly contains material factual allegations which must be pleaded if they are to be relied upon in connection with the claimants’ case on development (and perhaps training) in the UK.

[31]Having said all of that, there is no doubt that by reason of the somewhat hybrid approach that the claimants have taken to the SOCDT, they have pleaded both allegations of material fact, which need to be in a pleading, and in some cases evidence, such as the content of disclosure documents and their detailed case as to the identity of the specific datasets used for the training of Stable Diffusion, which does not need to be in a pleading. To this extent the claimants have in fact deviated from the paragraph of the Chancery Guide to which Ms. Lane drew my attention, albeit, as she says, in an attempt to be helpful in advance of the trial.

[32]Nothing I say in this judgment should be taken to indicate that the claimants may not use the disclosure documents available to them at trial to cross-examine witnesses or to advance arguments consistent with their pleaded case. The law on amendments

[33]The claimants did not seek to challenge the propositions of law advanced by the defendant in relation to the approach to be taken by the court to amendments, including late and very late amendments. As to the law on amendments generally, I was referred to Kawasaki Kisen Kaisha Limited v James Kemball Limited

[2021] EWCA Civ 3

per Popplewell LJ at [17] and [18(1)]-[18(3)], to the effect that the court will refuse permission to amend to raise a case which does not have a real prospect of success, that the claim must carry some degree of conviction, that it must be coherent and properly particularised and that it must be supported by evidence which establishes a factual basis which meets the merits test. To like effect, I was also referred to Elite Property Holdings Ltd v Barclays Bank Plc

[2019] EWCA Civ 204

, per Asplin LJ at [41] to [42].

[34]I accept the defendant's submissions, which were not challenged, that the requirements for coherence and particularisation are of particular importance close to trial when there is no time for the pleading to be queried or probed. I also accept that the requirement for support in the evidence is of importance in this application, where Getty seeks no disclosure in respect of its new case and should be able to plead a case that meets the merits test on the available materials now.

[35]The law on late amendments is well known. The defendant relied on Quah v Goldman Sachs International

[2015] EWHC 759 (Comm)

, per Carr J, as she then was, at [38]. It also referred me to CIP Properties v Galliford Try

[2015] EWHC 1345 (TCC)

, per Coulson J, as he then was, at [19] and Steenbok Newco 10 Sarl v Formal Holdings Ltd

[2024] EWHC 1160

, per Bryan J at [8] to [31].

[36]Where these principles were not challenged, and subject to the points I raise immediately hereafter, I take them all as read and do not need to set them out in this judgment. I bear them well in mind in exercising my discretion in dealing with this application.

[37]The additional point I make, however, is necessary because of the content of Mr. McKenna's statement in which he suggests that the defendant has had several weeks since service of the SOCDT during which it could and should have investigated the matters raised and made a start on a substantive response. This point is, in my judgment, misconceived.

[38]Steenbok makes clear at [22] that a responding party is not obliged to divert themselves from their trial preparation to meet a case which is the subject of a contested application for permission to amend. I am inclined to agree with the defendant that Mr. McKenna makes this point because it is obvious that the allegations in the SOCDT cannot be addressed from a standing start in the time to trial. Indeed, the claimants have not suggested otherwise. Furthermore, the fact that the claimants recognise a right to investigate and respond to these allegations appears to me to support the proposition that these are not merely particulars that could have been raised, without further notice, at trial. Is this a late or very late amendment?

[39]In my judgment, this is plainly a very late amendment, in that were I to permit it in its entirety, the trial would be jeopardised in that the work that the defendant would be required to undertake in order fairly to respond could not be completed prior to the trial. This is addressed in detail in the fifth witness statement of Mr. Aries, whose evidence as to the work that would now be required by the defendant if this pleading is permitted and the impossibility of completing that work in time for trial I accept.

[40]As I have said, I did not understand this evidence to be resisted by the claimants. Their point was a different one, namely that the defendant should have seen this coming. Mr. McKenna's evidence is that: "There is nothing in the SOCDT that the defendant was not aware of and able to deal with as it saw fit in evidence prior to it being served".

[41]I shall come in a moment to various aspects of the proposed SOCDT, but in general terms I reject this evidence, which appears to me fundamentally to misunderstand the extent to which the scope of litigation is governed by the pleadings.

[42]There is no reason whatever that a party must be familiar with every detail of its own disclosure, in this case running to approximately 9,000 documents and 78,000 pages, or that it is to be expected to anticipate a case that the other side may seek to advance based on the content of a few pages within that disclosure. All it need do in its witness evidence is address the allegations that are in fact pleaded against it. There is no need to spend time and money trying to anticipate what may be pleaded against it in due course. The witness evidence in any case is determined by reference to the pleaded case, as is plain from Practice Direction 57AC.

[43]Absent a pleaded case on the new material facts and technical issues now set out in the SOCDT, the claimants are wrong to suppose that there was any reason for the defendant to address them in its evidence.

[44]Similarly, and contrary to assertions made in the evidence of Mr. McKenna, there is no need to deal with matters raised in an expert report which are not pleaded but should have been - see Lenovo v Ericsson

[2025] EWHC 613 (Pat)

, per Meade J, at [12] to [13]. As he said:

"The pleadings are there to avoid surprise and to manage the proceedings efficiently in a cards-on-the-table way, and it is completely unacceptable to put in evidence which is unsupported by the pleading without any reason or any explanation."

[45]That Professor Farid's first report provided at the time of the summary judgment application referred to a "wide range of activities that are capable of falling within the definition of training and development" does not, in my judgment, assist the claimants, contrary to Mr. McKenna's evidence, in circumstances where that wide range of activities is not identified in the existing pleading and no allegations are made as to which of the activities falling within this wide range is in fact relied upon.

[46]As far as their case on development of Stable Diffusion is concerned, the claimants have never pleaded which of these wide range of activities is said to comprise an infringement of UK rights.

[47]The same point may be made about the content of Professor Farid's latest report for trial, which refers to the use of Notebook software now raised for the first time in the SOCDT. It is the pleading which determines the scope of the evidence, not the other way around.

[48]I agree with the defendant that the claimants' suggestion that there is no real need for the SOCDT and that it is purely voluntary appears at odds with the requirements of the overriding objective, the requirement that parties should properly plead out their case and relatedly the need for parties to put their cards on the table.

[49]Furthermore, as I observed during the hearing, if it is genuinely the claimants' position that they can run all the points pleaded in the SOCDT at trial without a pleading (and without giving the defendant any opportunity to respond to those points), then one wonders why they have incurred the time and expense in making this application for permission to rely on the SOCDT. The logical consequence of their position is that there is no need for permission from the court for reliance on the SOCDT.

[50]As I understood the claimants' submissions, they reject the suggestion that there has been any delay in making this application and so presumably they reject the contention that the SOCDT is provided late or very late on the grounds that the pleading of the SOCDT could not have been undertaken until Fieldfisher had received all of the defendant's disclosure and completed their review and analysis of the totality of that disclosure, following which counsel needed to provide input and then draft the SOCDT. Mr. McKenna also says that there has been no delay because the defendant only suggested a need for this pleading on 26 February this year. I reject this case.

[51]As to the suggestion that this pleading was prompted by the defendant and that its timing is purely a function of the claimants' attempt to assist by addressing the request made by the defendant for a better understanding of the case that was being advanced against it, the way in which the claimants choose to run their case is a matter for them. If particulars of their case are required in order to alert the defendant to material facts on which they intend to rely at trial, then it is incumbent upon them to ensure that those particulars are provided to the defendant with sufficient alacrity in advance of trial.

[52]That the defendant was asking to understand whether there was anything more to the claimants' case on training and development, given its very broad pleading in paragraph 43 of the Re-Re-Re Amended Particulars of Claim, is perhaps unsurprising, but it is neither here nor there. Similarly, the fact that the defendant has never made a request for further information in relation to paragraph 43 again takes the claimants no further. If there was a need to plead new material facts in support of the case in paragraph 43, that should have been identified by the claimants as quickly as possible upon review of the disclosure and it should have been treated as a priority. Mr. McKenna's attempt to suggest that there is no delay in the provision of this draft pleading because the concept of a SOCDT has only very recently been raised by the defendant is for this reason wholly misconceived.

[53]As for the point on disclosure, the defendant's disclosure was first received on 4 November 2024 in redacted form and then in unredacted form on 15 January 2025 (after a dispute which could have been resolved in December 2024 so as to enable the claimants to see the unredacted disclosure then). Some further disclosure on one issue was subsequently provided on 6 March 2025.

[54]I agree with the submissions made by Mr. Cuddigan that, given the impending trial date, the claimants should have appreciated that if they wished to provide particulars of their training and development case, they must do so as soon as possible, including, if necessary, by putting together a pleading which they knew might subsequently need amendment. Yet they appear not to have acted with any real sense of urgency. Mr. McKenna says that the documentation was more extensive than expected, but 9,000 documents were disclosed in contrast to the claimants' costs budget which provided for 30,000. Furthermore, Fieldfisher engaged the assistance of Weil, Gotshal & Manges LLP in the task of disclosure and the evidence in Mr McKenna’s fourth witness statement confirms that the review of the unredacted disclosure documents was completed by mid-January 2025 and that the review of the previously redacted disclosure documents was completed by mid-February 2025.

[55]In any event, various aspects of the new proposed pleading are not in fact dependent upon any documents while others are dependent upon documents received in November or, at the latest, January (in respect of which review was completed by mid-February). Why the SOCDT could not be served until 4th April, despite relevant documents relied upon in the SOCDT having been provided to the claimants' expert for inclusion in his report served in early March, is not adequately explained, particularly given that the claimants were aware that the case was being rigorously case managed in February and March of this year, thereby facilitating the making of an application (if necessary) at the earliest possible opportunity. I do not accept that the fact that the claimants were heavily involved in many other workstreams (as is of course also the case for the defendant) is a legitimate reason for delay.

[56]I agree with the defendant that the lack of any proper explanation for the delay may well be down to the fundamentally misconceived approach that the claimants took to this proposed pleading, effectively taking the view that it sets out voluntary allegations which do not need to be pleaded. However, a misunderstanding of the procedural rules of this court is not a good reason for delay.

[57]I consider the lack of any good reason for the delay in bringing this application forward to be an important factor in considering whether to permit the proposed SOCDT.

[58]In a nutshell, the claimants should have appreciated that in the event of further material facts becoming evident following the disclosure process, those facts would have to be pleaded by way of further particularisation sufficiently in advance of trial to ensure no unfairness to the defendant, specifically to ensure that the defendant had sufficient time in which to respond to them and adduce any further evidence as it saw fit. Individual sections of the SOCDT

[59]I now turn to consider individual sections of the SOCDT which raise different considerations. Section A, paragraphs 2 to 10

[60]Section A contains definitions and abbreviations and is itself unobjectionable. The defendant made no submissions about it and, in so far as I decide to permit any other sections of the SOCDT, there can be no real objection to granting permission for section A (always assuming that it remains relevant).

[61]Importantly, it now identifies a difference between training and Final Training of Stable Diffusion. Footnote 1 to the SOCDT concedes that the claimants do not allege that Final Training of Stable Diffusion using the datasets identified in section C, to which I shall return in a moment, took place in the United Kingdom. This appears to put the focus in this pleading on the issue of development.

[62]In this context, it occurs to me that for good order the claimants probably now need to consider whether further amendments to their Re-Re-Re Amended Particulars of Claim are required in order to ensure that any allegations which are no longer pursued are crossed out and that their case on training in the UK, in so far as it continues to exist, is clear. Section B, paragraphs 11 to 16

[63]Section B is entitled "The defendant's failure to maintain, preserve and disclose relevant documents and information". It pleads various forms of misfeasance on the part of the defendant, including a failure to keep detailed records, a failure to preserve documents, a failure to provide adequate disclosure and a failure to produce adequate witness statements.

[64]These allegations are, as I understood Ms. Lane to acknowledge during her submissions, wholly unparticularised. For example, there are no particulars of missing materials or of failures to comply with disclosure obligations. Furthermore, these allegations are advanced prospectively in so far as they relate to allegations in the SOCDT itself, which is both illogical and procedurally unfair. In addition, paragraph 13 raises an allegation of knowledge, which again is entirely unparticularised.

[65]In my judgment, these are very fundamental problems and would likely result in the refusal of permission for any amendment but must plainly result in such a refusal in the case of a very late amendment.

[66]Section B contains allegations which cannot properly be responded to by the defendant. As they stand, they are inadequately particularised and insufficiently coherent.

[67]Aside from these fundamental problems, there is in any event no need for a party to identify the inferences it will seek to draw from the evidence at trial in its pleadings, as I have already said, and neither side suggests otherwise.

[68]That these are inferences that the claimants might seek to draw at trial, depending on the evidence, is not in question, but for present purposes the pleading at Section B does not satisfy the requirements for a coherent and properly particularised pleading, particularly given the late stage of the proceedings.

[69]Accordingly, it would not be consistent with the overriding objective or indeed the principles to which I have referred on late amendments to permit reliance on Section B. It is not even, in my judgment, of any real assistance to the court or the defendant in preparing for trial, given the wholly unparticularised nature of the assertions made. Sections C and D, paragraphs 17 to 36

[70]Sections C and D are concerned with datasets used for the development of Stable Diffusion and Visual Assets and Copyright Works used for the Final Training of Stable Diffusion.

[71]In my judgment, much of the evidence included in these sections is correctly classified by the claimants as evidence which could have been presented in submissions at trial or put to witnesses in cross-examination. The question of the identity of the individual datasets used to train the various different versions of Stable Diffusion is plainly in issue in the proceedings and the defendant has served evidence in the form of a statement from Mr. Wagrez dealing with this very point. In essence, his evidence says that he was generally aware that LAION datasets were used to train Stable Diffusion but he does not know the details and he cannot identify which datasets were used for the training of the various different models.

[72]Against that background, I consider that it is entirely legitimate and to be expected that the claimants would undertake their own investigations by reference to the disclosure documents to try to obtain further information. This they have done and sections C and D set out the results.

[73]During submissions at the hearing, it became clear that the defendant could not sensibly object to this approach and nor could it suggest that sections C and D were not “manageable” subject to a couple of key points. Those points were, first, that the new references to the LAIOCO dataset and the coyo700m dataset must be excised and, second, that the references in paragraphs 19, 21, 33, 34 and 36 to inferences to be drawn from the absence of proper records being maintained or disclosed (a reference back to section B) must also be excised.

[74]Taking these points in turn, I agree with the defendant that the references to the LAIOCO and coyo700m datasets must be removed from this pleading. These are entirely new datasets never previously pleaded. The existing pleading refers only, on numerous occasions, to LAION datasets having been used in the training of Stable Diffusion. LAION was accordingly a prominent search term in the DRD, but there were no searches against any other datasets and none is mentioned by any of the defendant's factual witnesses. Yet the claimants in this proposed pleading seek to invite adverse inferences by reason of the absence of any disclosure relating to these entirely new datasets.

[75]Reliance on a new dataset that is said to have been used in the training or development of Stable Diffusion appears to me to be quite obviously a new material fact that has to be pleaded if it is to be relied upon. It is a pleading that the defendant would be entitled to investigate and respond to in advance of trial and it is a pleading that would likely have resulted in additional search terms being used for disclosure.

[76]Mr. Aries says in his statement that no investigations have been made by Bird & Bird into the content or use of these new datasets and that it is far too late for such investigations to start now. I agree.

[77]Having regard to all of the factors relevant to late and very late applications to amend, I consider that no good reason has been provided for the late attempt to bring new datasets into the proceedings. Whilst I accept that the fact that these datasets were used in the training of Stable Diffusion has only become evident following disclosure, nonetheless I have already queried why the claimants did not act more quickly to set out its new case.

[78]While I accept that excluding reference to these datasets may well prejudice the claimants, as they will not be able to advance any case based on them at trial, none the less I consider that that prejudice has come about through their own conduct in raising this issue so late and that, accordingly, this is a less important element in the balancing exercise.

[79]Set against the prejudice to the claimants is the obvious prejudice and disruption to the defendant of being required to deal with an entirely new case with only weeks to go to trial at a time when all of the defendant's energies should be taken up with preparing for that trial.

[80]Having said all of that, I accept the claimants' submission that they can rely on LAION-2B and so they can also rely on material in the LAIOCO dataset that comes from, and essentially forms the make-up of, the LAION-2B dataset. I did not understand that to be controversial. What the claimants cannot do is rely upon the coyo700m dataset which was merged with the LAION-2B dataset to make the composite dataset called LAIOCO.

[81]In any event, as will become clear, there is now no reason to rely on coyo700m owing to my decision in relation to section E.

[82]As to the defendant's further complaints, I agree that, consistent with my decision in relation to section B, and for the same reasons, the proposed inferences in the paragraphs to which I have referred based on the absence of proper records etc. must also be excised from these sections. Section E, paragraphs 37 to 56

[83]Section E is concerned with the “Development of Stable Diffusion [and the] creation of the LAION-A dataset and filters”. At paragraph 37.1 it identifies four individuals in the DeepFloyd team, Mr. Konstantinov, Mr. Shonenkov, Ms. Ivanova and Ms. Bakshandaeva, whose identities are already clear from evidence served by the defendant in the case.

[84]At paragraph 37.2, the pleading states that three of these individuals lived and worked in the UK, a plea that is not controversial. At paragraph 37.3, the pleading contends that the remaining member of the team, Mr. Shonenkov, worked in Dubai but that he also worked in the UK, a pleading which is contrary to the evidence of the defendant's witnesses. Reliance is placed on three documents in disclosure for this assertion.

[85]Paragraphs 38 to 43 need to be read in full, but in summary they then plead that the DeepFloyd team created the LAION-A dataset which was used to train certain Stable Diffusion models, that they also created filters, referred to as DeepFloyd filters, which were used for the LAIOCO and LAION-A datasets and that it is to be inferred that in the course of creating the LAION-A dataset and the DeepFloyd filters they downloaded and stored on UK computers Visual Assets and Copyright Works. This is alleged to be an infringement of the copyright in those Visual Assets and Copyright Works by reproduction and also an infringement of database right.

[86]Accordingly, it is said that the LAION-A dataset (which was used for Final Training) and the DeepFloyd filters were themselves created by infringing acts occurring in the United Kingdom.

[87]Paragraphs 44 and onwards then provide particulars of these activities, which I shall not set out in full but can broadly summarise as falling under the following headings.

[88]First, paragraphs 45 to 46, "DeepFloyd model and YouTube presentation". This is an allegation that Mr. Konstantinov and Ms. Bakshandaeva were involved in a presentation published in April 2023 during which they presented a screenshot of Notebook software which had been used to retrieve and display images taken from a LAION dataset. I understand this allegation (which is unobjectionable in itself in that it merely pleads evidence) to be simply setting up the allegation that then follows.

[89]Second, paragraphs 47 to 50, "Notebook software". In these paragraphs, technical assertions of material fact are made as to the nature and use of Notebook software (specifically in paragraphs 47, 49 and 50), including that the contents of Notebook are typically hosted and saved locally on a user's computer with the ability to connect to remote servers for file storage. It is inferred that at least Mr. Konstantinov and Ms. Bakshandaeva used Notebook software of this kind in their work for the defendant and it is asserted that if the contents of the Notebook are displayed locally on a user's computer, those images "are likely to have been downloaded and saved locally".

[90]Third, paragraphs 51 to 56, "DeepFloyd filters and other dataset cleaning". In these paragraphs (which again need to be read in full), it is alleged (in summary) that to create the LAION-A dataset, the DeepFloyd team filtered the LAION-5B dataset, de-duplicated images and created custom filters. It is then pleaded that it is thus highly probable and to be inferred that various identified acts were undertaken by each of the four members of the DeepFloyd team. These acts are said to involve the creation of what are called Trial Datasets, (not a term ever used in a pleading previously, and not a term supported by any documentary evidence), which it is said would have been made up of images selected from a repository held by the defendant comprising the LAION-5B dataset. It is then said, in summary, that it is likely that, to determine what should be included in a Trial Dataset, images stored remotely would have to be viewed locally and the Trial Dataset itself would be viewed locally before then being used to test the filter or deduplication program. The results of this process would then be viewed locally to assess its accuracy.

[91]It is then averred that the Trial Datasets were downloaded and stored on a local computer by the relevant individual as this would enable the test to run more quickly, thereby reproducing the images in material form on the local computer. Alternatively, it is averred that even if that was not the case, the act of running a program locally using Notebook software would itself download and store the images, thereby reproducing those images on the user's local computer.

[92]In any event, it is also averred that it is likely that the individual involved in this process would have downloaded and stored the Trial Dataset and its results locally and that the Trial Datasets would have been used for the purposes of testing the filter or deduplication program before it was deployed on the full LAION-5B dataset. It is alleged that the results of the filtering and deduplication process (i.e. the images retained and the images removed) would have been viewed by members of the DeepFloyd team on the screens of their local computers.

[93]At the heart of the allegations at paragraphs 47-50 and 51-56 lie a variety of important allegations of material fact, including: first, that the DeepFloyd team made use of Notebook software with particular technical properties, specifically that displaying that software locally would likely have the effect of downloading and storing images locally; second, that in the course of development, Trial Datasets would have been created and viewed locally and downloaded onto local computers by members of the DeepFloyd team (thereby causing images in the Trial Datasets to be reproduced in material form on the local computer); third, that even if Trial Datasets were not downloaded locally, the act of running a program locally using Notebook software would download and store the images held in the Trial Dataset thereby reproducing the images; fourth that whether the deduplication or filter program was run locally or remotely it is likely that the Trial Datasets would have been downloaded and stored locally; fifth (at paragraph 56) that it is likely that the Trial Datasets included a substantial number of Visual Assets, including Copyright Works.

[94]On the basis of these material facts, it is asserted that images must have been downloaded and stored on local computers. This is not, to my mind, mere evidence on which the claimants could have relied at trial without a pleading. It is material information which explains to the defendant the case that it must answer on the issue of development at trial.

[95]Interestingly, this pleading relies only upon one publicly available document in the form of the YouTube presentation which Mr. McKenna says was only discovered on 13th February, 2025. However, he does not say why steps were not taken immediately to plead that document, its relevance presumably being the reference to the use of Notebook software and the case that was then suggested to the claimants by reason of that use.

[96]Professor Farid deals with Notebook software in his report of 7 March 2025, but instead of prioritising the need to plead this new case as quickly as possible, the claimants appear to have thought it enough to include the point in his report and then to assert that the defendant will have an opportunity to respond to that in its own expert report in due course. It is unclear whether Professor Farid considers this is common software for data scientists such that he had identified it as a potential issue prior to the discovery of the YouTube presentation.

[97]I have already dealt with the fundamental procedural error in taking this approach, and it certainly does not permit the claimants to say that the defendant knows all about this point and should have dealt with it before. I reject the claimants' case that the use of Notebook software (given the technical allegations that are made as to the consequences of its use) is something that needs only to be addressed in expert reports and not pleaded, a suggestion which I consider to be fundamentally at odds with the most basic requirements of pleading.

[98]I consider that the same applies in respect of the new concept of Trial Datasets which does not appear to find support in any documentary evidence whatsoever and so, in theory, could have been pleaded at any time. There is no explanation whatever as to why it has not been pleaded earlier. Indeed, it is wholly unclear where the allegation as to Trial Datasets even comes from.

[99]Given the stage of the proceedings, I agree with the defendant that one would expect some support for this case in contemporaneous documents, but there is none. I also agree that the idea that Trial Datasets existed but that they have not come to light in the disclosure exercise done to date appears fanciful, not least because the search terms for disclosure included key words such as "dataset".

[100]In her submissions, Ms. Lane focused on witness statements provided on behalf of the defendant from Mr. Konstantinov and Ms. Bakshandaeva, two members of the DeepFloyd team. These statements confirm the existence of the DeepFloyd team. They confirm the use of filters in connection with a different model (i.e. not Stable Diffusion) and they also confirm that the defendant has always known that the allegation against it concerns downloading in the UK. Both witnesses deny downloading any images from LAION datasets and deny working on Stable Diffusion. However, I reject the suggestion that these witnesses could and should have been asked about the YouTube presentation or the use of Notebook software (including as to its technical properties) or the suggestion that there existed Trial Datasets containing a substantial number of Visual Assets including Copyright Works from which images had been downloaded. These were plainly not pleaded issues and had never previously been raised. As I have already said, the defendant was under no duty to anticipate these allegations and I reject the suggestion that the existing particulars at paragraph 43.1-43.7 of the Re-Re-Re Amended Particulars of Claim (which include reference to a development team residing and working in London) are in any way sufficient to foreshadow new allegations of this sort. As I have already said, those particulars focus on the location of the defendant’s personnel but make no mention of the particular activities they are alleged to have engaged in which would have involved the downloading of Visual Assets and Copyright Works.

[101]To my mind, there is no good excuse for these important issues being raised so late in the day. Furthermore, it is clear that there would be significant prejudice to the defendant if they were to be permitted. As Mr. Aries confirms, it would be necessary to revisit Mr. Konstantinov and Ms. Bakshandaeva's evidence in circumstances where it has already been necessary to obtain second statements from them at the claimants' request in February of this year. Notably, the claimants did not inform the defendant at that time that there were other issues directly relevant to these witnesses that they intended to raise, notwithstanding that they must have been aware of the YouTube presentation at around this time.

[102]I accept Mr. Aries' evidence that, as a matter of fairness, if these amendments were now to be permitted it would be necessary to undertake a series of factual investigations and to seek further evidence, including as to the use of Notebook software and Trial Datasets by the DeepFloyd team.

[103]I also accept that it would be desirable to conduct further searches of the defendant's documents in circumstances where no search terms were directed at the use of Notebook software and Mr. McKenna has identified only ten documents which refer to it.

[104]I also accept Mr. Aries' evidence that paragraphs 49 to 51 of the SOCDT advance a series of technical assertions in relation to Notebook software which require expert evidence if the defendant is to respond properly. I agree with the defendant that such expert evidence does not naturally fall within “the field of artificial intelligence models”, which is the field of expertise of the existing experts, and I accept that the defendant may feel that there is a need to obtain evidence from someone with IT expertise. In any event, Mr. Aries also says that even assuming this issue to be within his expertise, the defendant's expert, whose lengthy report in this matter was served on the day of the hearing, has no time available to extend his involvement in this case in the run-up to trial. He is already required to engage in the preparation of a joint report in conjunction with Professor Farid which Mr. Aries says, and I accept, is likely to involve a lengthy process. This would leave the defendant in a position where it would have to find and instruct a new expert and obtain a report in the run-up to trial at the same time as seeking further factual evidence which might be directly relevant to the expert.

[105]To my mind, this would be a wholly unsatisfactory and extremely unfair state of affairs. I accept Mr. Aries' evidence that there is simply insufficient time to undertake all of this work whilst at the same time preparing for this highly complex and technical trial.

[106]Having regard to the balancing exercise to be applied in the circumstances of a very late amendment, I do not consider that these proposed amendments in Section E should be permitted. The claimants have not advanced a good reason for their failure swiftly to plead this new case upon identifying the alleged use of Notebook software and Trial Datasets, and they have not satisfied the heavy burden which rests on them in seeking to advance a very late amendment. Although the claimants will be prejudiced by my refusing permission for this pleading, I consider that their prejudice must therefore carry a lesser weight in the overall balance.

[107]On the other side of the equation, this is not merely a situation in which the defendant's lawyers would be mucked around in the run-up to trial if these amendments were to be permitted. In my judgment, they would be subjected to enormous and unfair disruption and additional pressure, together no doubt with some duplication in costs, in having to explore issues further with pre-existing witnesses. They would be likely also to wish to explore further obtaining evidence from Mr. Shonenkov, although in light of the strength of the defendant's position on the new allegations of material fact there is no need for me to deal with the additional potential issue caused by the new allegation that Mr. Shonenkov, who was not included as a custodian owing to the fact that the defendant says that he was based in Dubai, carried out work in the UK. He is just one of the members of the DeepFloyd team against whom these very late allegations are made and there is no need to lengthen this already long judgment by dealing with his position. Section F, paragraphs 57 to 91

[108]Section F is entitled, "Further use and reproduction of Visual Assets and Copyright Works in the United Kingdom in the course of the development of Stable Diffusion". This section pleads a somewhat miscellaneous collection of allegations about downloading and storage of images in the UK which originally appeared to be made against 31 individuals, but which Ms. Lane clarified as being made only against the individuals specifically named in these paragraphs against whom it is said that they would have had the technical ability to carry out downloads. For reasons to which I will return, this clarification does not entirely assist in relation to a number of the existing allegations.

[109]These allegations appear from paragraph 58 to be heavily reliant upon the drawing of a proposed inference as to the failure to maintain or disclose proper documents, which is entirely unparticularised and bad for all the reasons I have already given.

[110]Paragraphs 63 to 64 make allegations as to the creation by an unnamed individual of new training datasets from old training datasets based on a Slack conversation between two individuals not previously named. Mr. Aries explains that the individual referred to in the conversation detailed in the relevant document is a Mr. Kumar, an existing custodian whose relevant repositories have been searched for relevant key words but have turned up nothing favourable to the claimants. This pleading is based on a disclosure document received in November 2024 and provided in unredacted form in January 2025. For reasons I have already given, I do not accept that it could not have been pleaded earlier.

[111]Once again, the claimants seek to suggest that this allegation was foreshadowed in Professor Farid's first report, but I have already dealt with why that contention is misconceived. I accept Mr. Aries' evidence that the introduction of this allegation, involving a previously unpleaded individual, requires further enquiries to demonstrate the work that Mr. Kumar undertook in the relevant period. I also accept that this could not be done in the time available. Paragraphs 65 to 76

[112]In these paragraphs, the claimants allege that certain named individuals downloaded data on their local computers. They rely on two documents, one from the 8 February 2023 referring to Mr. Phillips as a "dataset collector" and referring to Ms. Logeswaran and Ms. Patel as having been hired out of the UK to “curate data”, and the other from the 19 August 2024 referring to Mr. Phillips being involved in "dataset preprocessing" and "collecting datasets".

[113]In light of these documents, the claimants seek to infer,(i) that Mr. Phillips' work involved downloading images;

(ii) that Ms. Logeswaran and Ms. Patel were involved in downloading images as part of their curation of datasets;

(iii) more generally, that unnamed individuals in the UK were instructed to, and did, download data in the UK. It is unclear how Ms. Lane's concession applies in relation to that latter allegation.

[114]In my judgment, these allegations are all tenuous in the extreme. They carry no degree of conviction and they appear to depend in part on the court being invited to draw adverse inferences which are inappropriate. They are also insufficiently particularised. There is no specific identification of the datasets on which the claimants rely in respect of any of these allegations, notwithstanding the need for them to establish downloading by reference to the LAION datasets (i.e. the only datasets currently pleaded). There is no identification of the unnamed individuals referred to in paragraphs 73 to 74 or, indeed, further at paragraph 75, and no explanation as to how that pleading now fits with the concession that this Section only applies to named individuals.

[115]Indeed, the pleading in relation to Ms. Logeswaran and Ms. Patel expressly confirms that the claimants do not know the nature of the datasets that they were curating, or whether they were used for the training and development of Stable Diffusion. This is hopeless at this stage and certainly insufficient to meet the merits test for an amendment, aside from the lack of any proper particularisation.

[116]On close analysis of the document relied upon by the claimants, it is plain that very far from supporting the claimants' case, the references to Ms. Logeswaran and Ms. Patel concern their work on entirely different datasets belonging to the defendant's paying clients, not connected to LAION at all. This pleading also suffers from the same points on lateness and the claimants only have themselves to blame for the position in which they now find themselves.

[117]Having regard to the balancing exercise, I do not consider that this pleading should be permitted. Aside from the fact that it does not satisfy the most basic requirement for a coherent and properly particularised pleading with a real prospect of success, it would plainly create considerable additional work for the defendant, as Mr. Aries confirms, if it has to be dealt with, including the need to seek evidence from various new witnesses not previously identified. Paragraphs 77 to 81

[118]These paragraphs concern allegations relating to Mr. Tom Mason, the defendant's CTO in the UK at the relevant time. The claimants seek to rely upon a Slack conversation between Mr. Mason and Mr. Ruhl which refers to a comparison between “alpha and beta”. They infer that to make this comparison it is likely that Mr. Mason and others, who are unidentified, would have downloaded and viewed a selection of Training Datasets. Once again, however, this is highly speculative and fails entirely to meet the requirements for a properly particularised or coherent pleading. There is no attempt to identify the others who are referred to and no explanation as to why it is thought that "alpha" and "beta" is in fact a reference to datasets. Indeed, these are referred to earlier in the SOCDT as being models.

[119]For similar reasons to those given already, this pleading should not be permitted at this late stage. There is no good reason for it having been pleaded so late and the prejudice to the defendant is again substantial owing to the enquiries that would need to be made, as Mr. Aries confirms. Mr. Ruhl was not even an original custodian. Paragraphs 82 to 91 - annotation of datasets

[120]This allegation relates to a proposed outsourcing project regarding data annotation. Once again, it is on its face a highly speculative pleading. Paragraph 84 pleads that the claimants do not know whether the outsourcing proposal was implemented beyond trial, and paragraph 91 pleads that the claimants do not know the outcome of communications with the third party to whom the outsourcing was to take place, namely, Encord. There is no basis pleaded on which the defendant could be held liable for the acts of this third party. These paragraphs thus fail at the first hurdle. They are insufficiently particularised and they fail to meet the merits test.

[121]Paragraphs 85 to 89 rely upon documents for the inference of infringement, but these are not convincing.

[122]In respect of the first document referred to at paragraphs 85 to 86, there is no reference whatever to whether any downloading was carried out in the UK, an essential element in the claimants' cause of action.

[123]In respect of the second document referred to at paragraphs 87 to 89 sent by Ms. Davydova of the defendant to Amazon, there is no more than a reference to “dataset images” without reference to the identity of the dataset concerned. There is thus no attempt to link this to the pleaded case on LAION datasets and thus the allegation is unconvincing. To respond to these new allegations would require, as Mr. Aries confirms, new evidence involving a new witness, Ms. Davydova, and third parties, including Encord and Amazon. I accept that this would not be viable in the time allowed. Ms. Davydova is not an existing custodian and so her documents have not even been collected for searching.

[124]In the circumstances that I have identified, the balancing exercise in respect of these new allegations in section F appears to me to fall firmly in favour of refusing permission.

[125]Finally I repeat, however, that in so far as the claimants wish to cross-examine by reference to any of the documents on which they rely in Sections E or F at trial, in order to seek to make out their pleaded case, then there is nothing to stop them from doing so. If that cross-examination were to result in information which affected that pleaded case in some material way then an application to amend might be needed at that time, as Ms Lane appeared to accept during her submissions. Conclusion

[126]In conclusion, I shall permit the claimants to rely upon the SOCDT in respect of section A in so far as it remains relevant and serves any purpose, and sections C and D, subject to the excisions to which I have referred.

[127]I shall not grant permission in respect of sections B, E and F of the SOCDT for the reasons I have given. - - - - - - - - - -