Copyright in the outputs of generative AI

A new Italian Supreme Court judgment on copyright in an image created using software is a good jumping off point for a discussion of copyright in images produced by “generative AI”.

The Italian Supreme Court (in Rai Radiotelevisione Italiana S.P.A. v Biancheri Chiara No. 1107/2023) has held that the use of software per se does not exclude the human creativity necessary for copyright protection but, where this is challenged, a finding of fact will be needed to check whether and to what extent the use of the tool absorbed (“avesse assorbito”) the creativity of the artist who made use of the software (section 5.3). The facts were not considered in the case because the plea was only raised during the appeal and was inadmissible (section 5.2).

Similar questions have been raised in various jurisdictions, e.g.: Design Data Corp. v. Unigate Enter., Inc., 847 F.3d 1169, 1173 (9th Cir. 2017) and Reardon LLC v. The Walt Disney Co., 293 F. Supp. 3d 963 (N.D. Cal. 2018) (USA); Express Newspapers Plc v Liverpool Daily Post [1985] 1 W.L.R. 1089 (UK); and Telstra Corp Ltd v Phone Directories Co Pty Ltd [2010] FCAFC 149 (Australia).

The same question applies to generative AI: does copyright subsist in works created by the new breed of systems, such as DALL·E 2, Midjourney and Stability Diffusion? Here there is the extra complexity of whether a user’s text prompt is sufficient to meet the originality threshold for copyright in an artistic work.

In the UK, the closest cases are Cala Homes (South) Ltd v Alfred McAlpine Homes East Ltd (No.1) [1995] 7 WLUK 62, [1995] F.S.R. 818, [1996] C.L.Y. 3632, in which a person giving detailed verbal instructions was held to be a joint author of artistic works, and Kenrick & Co v Lawrence & Co, (1890) 25 Q.B.D. 99 (under the old UK law, the Copyright (Works of Art) Act, 1862).

These cases foreshadow the key question of fact of generative AI: whether a specific text prompt passes the threshold for originality. But both cases need to be approached with caution since they pre-date the harmonised approach to originality for the EU (then including the UK) in a number of EU Directives and CJEU judgments.

In Cala Homes the instructions “took a whole day in which the design of each of the first group of 12 house designs were considered in detail” and that “much if not all of the design features … were insisted upon … during the briefing and vetting sessions …”. By contrast, the judge noted (though without any express endorsement) that counsel supporting the case for joint authorship “conceded that mere instructions such as ‘paint me a yellow flower’ would not make the person who gives those instructions a joint author of the resultant painting. In that case all the skill and labour in composition would come from the painter.”

In Kenrick & Co v Lawrence & Co, the judge suggested that a person who, unable to draw, told an artist to draw a hand carrying a pencil in the act of finishing a cross enclosed in a square (to represent voting) was not the sole author of the artistic work but “possibly” a joint author with the artist.

It therefore seems possible that, in the UK, textual instructions can give rise to a claim of authorship in an artistic work generated by AI, especially where they specify elements of the expression protected by the artistic copyright, such as the composition, tone and visual style. Choosing the subject matter alone seems unlikely to suffice, particularly in the case of broad subject matter (such as “a flower”), as this contribution may fall the wrong side of the ideas vs expression distinction repeatedly seen in copyright treaties, laws and case law. This is consistent with Kenrick & Co v Lawrence & Co, in which the judge said “I think that I am upon very safe ground in saying that the mere choice of subject can rarely, if ever, confer upon the author of the drawing an exclusive right to represent the subject”.

In practice, it seems successful generative AI results can involve significant creative choices by users, which should make authorship more likely.

An interesting practical example is the selection of the subject and style for an image for a magazine cover, described in Gloria Liu, “The World’s Smartest Artificial Intelligence Just Made Its First Magazine Cover”, 21 June 2022. The image was created from the prompt “wide-angle shot from below of a female astronaut with an athletic feminine body walking with swagger toward camera on Mars in an infinite universe, synthwave digital art” and then further manipulated using “an experimental feature, not yet available to users, that draws on the context of the image to ‘extend’ it to the correct cover proportions” (but which, from an animation included in the online article, appears to have involved choosing from a range of alternative design elements). Similar examples are discussed in a recent article in Wired magazine: “Engines of Wow”.

It should also be remembered that the UK, unlike most jurisdictions, also provides for protection of artistic works generated by computer where (by definition) there is no human author. In that case the work is owned by “the person by whom the arrangements necessary for the creation of the work are undertaken”. That is going to be another interesting factual question in the case of generative AI!

There are at least two disputes over authorship of AI generated works to watch, both in the US. First, the DABUS team’s appeal of US Copyright Office’s refusal to register A Recent Entrance to Paradise, an image created by the DABUS AI. Second, whether the US Copyright Office cancels its registration of Zarya of the Dawn, a graphic novel created/generated by Kristina Kashtanova using MidJourney. The USCO has given notice of potential cancellation and requested details of the human involvement in the creation of the work (see here) (it currently appears the registration has been cancelled but only, at this stage, because of system error).