IL-2023-000007 - [2025] EWHC 2863 (Ch)

The re-worded prompts were generated by Professor Farid prompting ChatGPT with the prompt “provide a reworded version of this image caption: [relevant original Getty images caption]”. Getty Images submit that the re-worded prompts are not contrived because they correspond substantially to captions for Getty Images’ Content, which are often descriptive of the type of content (since it may be based on real places, people and events). In their opening submissions Getty Images asserted that “[i]t is likely that a user of Stable Diffusion would come up with something substantially similar to a caption in order to describe the type of content they wished to generate”. However, there is simply no evidence whatever to support this assertion and certainly no attempt to establish a “likelihood” by reference to any statistical evidence. As I have already said, the evidence before the court is that these prompts were a deliberate choice by Ms Varty directly informed by her atypical use objectives.

I agree with Stability’s submissions in closing that the case on re-worded prompts is necessarily dependent on the verbatim prompts case. This was supported by Professor Brox who explained in his oral evidence (which I accept) that he would not expect the re-wording by ChatGPT to make any real difference to the likelihood of watermarks* being generated because the concepts described in the text are what is important. I did not understand Getty Images to gainsay this evidence.

As I have said above, the verbatim prompts case posits that a user will visit a Getty Images Website, electronically copy a caption and then paste that caption into Stable Diffusion. As Stability points out, the re-worded prompts case involves an extra step: the Getty Images caption must be extracted, or at least read, in order for it to be re-phrased. If the approach adopted by Professor Farid were to be undertaken then the caption would have to be pasted into an AI model such as ChatGPT together with an appropriate prompt inviting the model to re-word the caption.

In closing, Getty Images put their case on verbatim prompts and re-worded prompts no higher than that “[a] user may copy and paste a caption from a Getty Images Website or amend a caption to vary what appears in the image…” (emphasis added). This much is not disputed by Stability. But there is simply no evidential support for the proposition that a real life user has in fact done this, or indeed that (having done so) watermarks* have been generated.

The prompts at nos. 11 and 13-18 on page 25 of Getty Images’ NoE were invented by Ms Varty based on (i) events that she was aware of or were imagined by her; and (ii) featured people that she was aware of or were imagined by her, and which were loosely inspired by the verbatim captions she used elsewhere. Getty Images admit that, in one sense, these prompts are contrived because they were created by Ms Varty rather than arising naturally, but nevertheless they say that they are neither artificial nor unrealistic. They contend that these examples serve to demonstrate that (as the Experts accept) “with any kind of prompt, the user may obtain an image bearing a watermark*”.

The six “inspired prompts” used by Ms Varty are:

“11 Striker Harry Kane of Tottenham Hotspur scores equaliser during London Darby against the Chelsea at White Hart Lane on February 12, 2020 in Tottenham, London

13 Son Heung-min of Tottenham Hotspur celebrates after defeating Arsenal in North London Derby, on February 7, 2016 at White Hart Lane, London

14 Fred Smith of Brittingham Hotspur celebrates after defeating Arsenal in North London Derby, on 7 February, 2018 at White Hart Lane, London

15 Fred Smith of the Philadelphia Eagles celebrates after defeating the New England Patriots on February 4, 2018 in Minnesota

16 One Direction performs during the Super Bowl Halftime Show at Tulane Stadium on February 4, 2018 in New Orleans, Louisiana

17 Pop star Bruno Mars performs at Wembley Stadium, London on March 20, 2022 in, London

18 Pop star Dua Lipa performs at Wembley Stadium, London on March 20, 2022 in, London”

Stability accepts that the generation of images bearing watermarks* is possible, that real world users may generate such images and that such users may use very detailed prompts, but it contends that the Court has no means of determining that these, or equivalent prompts, have ever been deployed by any user of any version of the Models in the UK resulting in the generation of a watermark*.

It points out that, as Ms Varty accepted in cross examination, each of these prompts follows the grammar or structure that Ms Cameron confirmed in her evidence was the “house style” for Getty Images’ captions on news, sport, entertainment and contemporary Content. Stability submits that use of this structure would appear to be dependent upon knowledge of the Getty Images captions. I agree. In my judgment, reliance upon this category of prompts is dependent upon establishing that a real world user would wish to replicate the grammar and structure (together with the unconventional date format – at least for UK customers) of a Getty Images caption, while not copying it precisely and that, having done so, he or she would then generate a watermark*. There is simply no evidential basis for a finding that any real world user of Stable Diffusion in the UK has deployed this unusual syntactical arrangement or that, having done so, a watermark* has been generated.

Stability also points out that with the exception of the two ‘Fred Smith’ prompts (which are not concerned with real people) each of these prompts represents an “out-of-scope” use for the Models in that it is intended to produce a piece of misinformation: a fake depiction of an event (on a particular date at a particular location) that never took place. While it may of course be the case that “out-of-scope” use takes place in the real world, again there is no evidence to which I have been directed by Getty Images to show that “out-of-scope” use of this sort has in fact occurred (that the structure referred to above has been used) and that real users in the UK, or anywhere else, have generated watermarks* by such use. Certainly none of the images in Annex 8I (insofar as it is possible to see the prompts that have been used) appears to fall within this category. Further and in any event, as Stability points out, any user intending to produce fake imagery of this kind will of course be aware that he or she is engaging in fakery and will view the ensuing image with that aim in mind.

Ms Varty’s last category of prompts is “invented prompts”. Examples used for the purpose of the images in Annex 8H (at 19-24 on page 26 of Getty Images’ NoE) are:

“19 Miley Cyrus, performs onstage, during the 2008 American Music Award, news photo

20 Editorial news photo, Savoy School May 24, 2013 in Washington

21 Miley Cyrus at American Music Awards, news photo

22 News photo of MTV VMAs

23 Vector art spaceship

24 Obama, news photo”

These prompts are short and use two phrases which Ms Varty explains in her evidence she understood might produce watermarks*; namely “news photo” and “vector art”. Getty Images submit that the words “news photo” might be added by a user to a prompt “if they wanted a realistic image of a current event”, while the words “vector art” have in fact been used by real life users, as is clear from Annex 8I.

Stability accepts that the words “news photo” “might” be deployed by a user if he or she wanted a realistic image of a current event, albeit that it points out that this would be out-of-scope use of the Model. Further, it is clear from the examples in Annex 8H that the use of these words will, on occasions, result in the generation of a watermark*. Professor Brox accepted as much, explaining that the chances of this were likely increased because Getty Images Websites include many news photos.

Although Stability points out, correctly, that Getty Images has carried out no analysis of the frequency of use of this phraseology by real world users, or the circumstances in which it would, or would not, contribute to the appearance of a watermark*, there is one example of a real world user generating an image with a watermark* using this form of prompt in the evidence: an exhibit to Ms Varty’s statement shows two images of Donald Trump behind bars generated by Stable Diffusion v1.5 (which is not in issue in this case) with the caption “Old donald trump behind bars in a jail, news photo”. I reproduce one of those images below (“the Donald Trump Image”):

Ms Varty was unable to provide any information as to the identity of the user or why he or she had added the words “news photo”.

Although there is no evidence of any of the Models in issue being used in real life to generate images using a caption which includes the words “news photo”, I consider the Donald Trump Image to provide evidence that a real world user of a diffusion model has in fact used the words “news photo” to generate an image. I do not find this very surprising. Professor Farid explained in his cross examination that it was very common for users of Midjourney Discord to “have stock phrases that they use over and over and over again on every caption because they have found that those particular words, like 85mm lens, F1.2 aperture, ultra realistic, give them the type of images they want”. I consider that it is reasonable to infer that for anyone wanting to achieve an “ultra-realistic” image, the words “news photo” might well be used in this way. Accordingly, I also consider it to be reasonable to infer that users of the Models in issue in the UK will also have used the phrase “news photo” to generate an image and, given the evidence in Annex 8H, combined with that of the Experts, I consider that, on balance, at least one real user of v1.2, 1.3, 1.4 and 2.1 in the UK will have generated an image using these words which includes a Getty Images watermark*. I take this view notwithstanding that some users may understand (from their reading of the Model Cards and/or other relevant material to which I shall return later) that the use of the words “news photo” amounts to “out-of-scope” use. All I really understand this to mean is that the Model was not designed to generate photo-realistic images of current events and real people; but I can see no reason whatever to suppose that this would prevent or preclude users from “playing around” with the Models to see what they are capable of producing (particularly given that the Model Cards highlight that the Models are intended for “research purposes”).

The words “vector art” (which are referring to a type of digital art that uses mathematical equations to represent images as a series of points, lines, curves and shapes) have been deployed by two real world users (as is clear from the images at pages 3 and 23 of Annex 8I). I have inferred that the first of these images was generated by v1.x. The second image is dated 28 January 2023 and so it is impossible to know for sure to which Model it relates and I have declined to draw an inference (although SD XL and v 1.6 can be ruled out owing to the date of the image). Ms Varty’s prompts using the words “vector art” produced synthetic images with iStock watermarks* for v1.2, v1.3 and v1.4 only. Getty Images has no examples in Annex 8H of synthetic images with watermarks* being produced by the use of the words “vector art” in respect of any other version of the Model. Accordingly, I consider that, on balance at least one real world user in the UK of v1.2, 1.3 and 1.4 will have generated an image using the words “vector art” which includes an iStock watermark*. There is no evidence on which I could make a similar finding in respect of any of the subsequent versions of the Model.

In conclusion: