AI Fashion Designers Bring Along Both New and Vintage Copyright Issues
Artificial intelligence (“AI”) is now capable of creating works that previously were created solely by human authors. The development of AI challenges the copyright system, which was created to protect the creative endeavours of human creators. The fashion industry provides a particularly fascinating research context for contemplating the copyright issues related to the use of AI, because fashion is an extremely IP-intensive industry that is known for its complex copyright environment. The increasing use of AI is likely to cause even more copyright issues, and perhaps even increase the amount of garment and accessory knock-offs, look-a-likes, imitations and copies. This creates a need to take a further look at the copyright problems that AI brings along and to consider how these issues affect the industry on a larger scale.
THE USE OF AI IN THE FASHION INDUSTRY
The development of AI and digitalization in general open massive opportunities in the fashion industry, from designing to manufacturing and sales. Since this article delves into copyright issues, the focus is on the use of AI in fashion design rather than marketing and sales.
Guiding the design process by predicting current and future trends
Trend forecasting is perhaps the most important way of using AI in the current fashion industry. The purpose of trend forecasting is to predict, for example, what kind of colours, silhouettes, materials and styles will be popular in the future. AI-based trend forecasting utilizes data mining, which is the process of uncovering patterns in large amounts of data. In copyright terms, a definition of data mining can be found from the newly adopted DSM Directive. Article 2 (2) defines text and data mining as “any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations”.
The benefit of AI in trend forecasting is that it can deal with significantly larger amounts of data than a human brain. It should be noted, however, that trend forecasting is not fashion designing itself, but it can definitely affect a human designer’s work by guiding it to the direction that is as fashionable as possible.
AI as an assistant to creativity
AI solutions provide rather helpful tools for fashion designers in their creative work. For example, AI provides a possibility to customize a garment based on an individual consumer’s preferences. A customized fashion design, however, can hardly be counted as an original work in the copyright sense. Customization as a form of AI creativity could merely be defined as bringing already-existing works to the consumer’s attention and making these works more attractive to them.
Furthermore, AI can help a designer to turn sketches into colour images. This is called image-to-image translation and it is utilizing generative adversarial networks (GANs). It includes converting a simple black-and-white drawing into a colour image. This technique helps presenting a design in online platforms and catalogues as a more real-life version before it has been actually manufactured. However, like customization, image-to-image translation is certainly more of a designer assistant. It would not be likely to generate original works by itself.
AI as an independent designer
In addition to directing the style of designing and assisting a human designer, it is possible to incorporate AI in the designing process itself and let AI design independently – at least in theory. In 2017, Amazon claimed the ability to train a generative adversarial network (GAN), a type of generative model, to design garments. The use of GAN as a fashion designer refers to a process of taking a dataset of images, and outputting images that are visually similar but generated by the model. The images that form the input dataset could be, for instance, images of garments that are trending on social media. This type of use of real-time data and GANs for fashion design purposes can help a fashion company to understand a demand for the garments before producing them.
AI can be given the capacity to learn styles from large datasets of content, and by using that content, mimic human designers. AI’s deep learning techniques are not limited to copying the styles of pre-existing fashion designers: they can also be used to mix and combine multiple sources, from a variety of styles, and come up with rather original outcomes.
American fashion brand Marchesa provides an interesting example of incorporating AI into fashion designing. In 2016, Marchesa and IBM Watson co-created a dress that “thinks”. The dress was showcased in the annual Met Gala. The AI program that was being used, IBM Watson, is a cognitive system that understands, reasons and learns. Watson scanned vast volumes of fan social data to enable the dress to change colour in response to social sentiment by using Watson’s Tone Analyzer API to analyse fan tweets and showcase them in real-time. In other words, the dress understood and responded to its fans by changing its colour. It was able to “redesign” itself over and over again by analysing data. In theory, the AI application could come up with changing prints and patterns decorating the dress that could (if they were created by a human author) be considered original in copyright terms.
Conclusions about fashion design and the use of AI
There has been some debate about whether AI will revolutionize fashion designing and if so, when will that revolution take place. Some experts (such as Leanne Luce, the author of Artificial Intelligence for Fashion) have considered AI fashion designers to be, at least for the time being, “more of a hype than anything production ready”. AI-based programs still do only what operators have told them to do, albeit with fantastic speed and accuracy. Regardless, we cannot ignore AI’s potential to create fashion designs without human interference. When it comes to the latest types of AI, the computer is able to make many of the decisions involved in the creative process without human intervention. Hence, it is possible that in the future AI will become a principal source of new designs.
COPYRIGHT ISSUES RELATED TO AI DESIGNERS
Fashion and copyright: general issues
In order to fully understand the big picture, it is necessary to take a look at the pre-existing copyright problems that the fashion industry has been struggling with for decades, before delving into the new copyright challenges that AI brings to the table.
The fashion industry is notorious for various forms of imitation, such as copying, piracy, counterfeits and knock-offs. This is at least partly caused by the fact that the fashion cycle needs at least some kind of imitation in order to function. The difference between inspiration and copying is everything but easy to determine. Quite often it is extremely difficult to distinguish whether a fashion design takes part in the same trend with the source of inspiration, or is a copy of the source of inspiration and hence, may infringe on the copyright of its author. This whole distinction between participating on a trend and copying another fashion design is related to the idea-expression dichotomy of copyright law.
What makes it relatively risk-free for fashion companies to copy each other is that fashion items often fall outside of the scope of copyright. The reason for this is that historically the originality threshold for works of applied art has been rather high. Nevertheless, the situation in Europe seems to be changing due to the recent CJEU judgment C-683/17 Cofemel, which noted that the standard of originality is the same for both pure arts and applied arts. Having said that, when it comes to garments and accessories, their functionality plays a role in copyright protection. The CJEU held in C-393/09 BSA that when the designing process of any type of work is dictated by the technical function of the work, the criterion of originality is not met (although it must be noted that copyright protection does apply to a design whose shape is, at least in part, necessary to obtain a technical result, if the design otherwise meets the standard of originality as interpreted by CJEU – see C-18/833 Brompton).
To conclude, the fashion industry faces significant copyright challenges already, without any interference from AI. The key challenge is related to fashion designs passing the originality threshold and hence qualifying as copyright protected works, which in many cases reduces the legal risks of copying fashion designs. The lack of protection has helped especially fast fashion companies to knock off popular designs of high-end and indie designers. Thus, when AI meets fashion designing, we are not only dealing with the already-existing copyright issues that the industry faces, but also with the copyright questions that AI brings along.
Fashion and AI: new copyright issues
As mentioned earlier in this text, what often makes it difficult for fashion designs to be protected by copyright is that they often lack originality, and a work cannot be protected by copyright unless it is original. In the European Union, originality is considered to mean “author’s own intellectual creation” and that the author has “stamped the work with their personal touch” (as defined in C-5/08 Infopaq and C-145/10 Painer). The standard of originality applies to all categories of works, even the ones of applied art, such as fashion designs. The EU approach to originality highlights the personhood of the author. Originality has consistently been interpreted in a way that it requires the author to be human, reflecting the civil law countries’ natural law justification for copyright protection. The Berne Convention itself does not define authorship, nor does it require the author to be human. However, the member countries of the convention seem to agree that the author needs to be human. This applies even in cases of arguably borderline authorship, such as applied art. In other words, within the scope of Berne it is not possible for an AI fashion designer to be considered as an author. This results to the conclusion that an AI-generated fashion design cannot be considered original, unless there was a sufficient amount of human intervention in the creative process. In theory, there could be cases where one would be able to find sufficient human input from an AI-generated work and hence, a human author behind the AI. However, purely AI-generated fashion designs would be likely to fall outside of the scope of copyright protection by default. And if these AI-generated fashion designs would not fulfil the requirements of design protection or some other form of IP protection, they would be free for anyone to copy. This would be the kind of copyright law approach that some scholars find ideal for fashion designs, because they consider that fashion designs should be placed outside of the copyright regime.
Another factor to consider is the possibility of an AI designer infringing another designer’s copyright. AI’s ability to “create” is mostly based on mimicking pre-existing fashion designs by using GANs. In other words, the creativity of GANs is based on borrowing elements from already-existing designs by other creators. This configuration sounds like it is asking for copyright infringement lawsuits, but is it really? Here we once again run into the problem of originality in fashion designs. It is notorious that most fashion designs face problems in reaching the originality threshold already before the interference of AI designing. Thus, if an AI designer uses the massive amount of data it has been fed and comes up with a fashion design that is very similar to another designer’s work, that would not necessarily constitute a copyright infringement, if the source data would comprise of fashion designs that would not qualify as original in copyright terms.
However, there are plenty of fashion designs that qualify as original works. Would the use of original fashion designs as AI designer’s source data constitute a copyright infringement? Not necessarily. The DSM Directive provides a copyright exception that might create new room for this kind of lawful use of copyright protected material in Europe. The directive includes mandatory exceptions regarding text and data mining. Article 4 allows acts of reproduction and extraction for the purposes of text and data mining. Furthermore, reproductions and extractions may be retained for the same purpose. The article also applies to data mining purposes that have a commercial motive. This is relevant from the perspective of AI designers, because in the process of their analysis they will invariably reproduce copyright protected fashion designs. After implementation of this exception in the DSM Directive, these reproductions could be exempted from infringement. However, it shall be noted that article 4 allows copyright holders to opt out of the exemption.
Furthermore, if the dataset of images used by GAN consists of original fashion designs, the resulting work needs to be modified enough from the data in order to avoid infringement claims. The resulting work should be altered to the point that it is no longer substantially similar to the source of inspiration in order to eventually get to the point where (in a case of a human creator) the creator of the latter work could be considered an author that has an independent copyright to their sole-authored work. However, if AI only makes de minimis modifications to the source of inspiration, AI is an infringer, not an author-like subject.
An important aspect to consider is, would AI designers change the market for fashion designs? In the current copyright environment, AI-generated designs would end up in the fashion industry’s public domain even more easily than the creations of human designers. This is something that we should try to avoid, especially for sustainability related reasons. It is clear that way too much clothing is being produced and thrown away. Fashion suffers from overproduction and overconsumption. Hence it is worthwhile to consider, how will the increasing use of AI designers affect the world of manufacturing, and how does that further affect the copyright system. AI designers might work faster and more effectively than a human designer, but a large amount of its “creativity” of is based on mimicking pre-existing designs. This could lead to an increasing amount of copies in the industry which is already struggling with countless knock-offs. Some scholars consider copying as the engine that keeps the fashion industry going, and creates more and more fashion consumption. The danger is that the use of AI designers might end up filling the public domain of the industry. This could boost the cheap knock-off activity that is a common business practice of especially fast-fashion companies. Making manufacturing of knock-off fashion easier by increasing copyright-free space in fashion designing might give incentives for fast-fashion companies to produce even more knock-offs, since the risk of being sued for copyright infringements decreases. What is more, letting AI-generated innovations fall automatically into the public domain might also have a chilling effect on investments in AI systems. There are very few incentives for fashion houses to invest in and develop AI designers, if their designing results would automatically be free for their rivals to copy.
HOW SHOULD COPYRIGHT LAW DEAL WITH THIS?
A variety of academics have already presented some possible solutions to how copyright law should treat AI-generated works. Denicola argues that if a user’s interaction with a computer prompts it to generate its own expression and the result is excluded from copyright, this is a tenuous, ultimately unnecessary and counter-productive distinction. Furthermore, it “denies the incentive of copyright to an increasingly large group of works that are indistinguishable in substance and public value from works created by human beings”. In Denicola’s opinion, a computer user who initiates the creation of an AI-generated work should be recognized as the author and copyright owner of the resulting work. Another interesting solution to the problem is the one considered by Alén-Savikko, Ballardini and Pihlajarinne. They presume that in the future, copyright protection might develop into a “dual system”, that would be divided to “romantic” protection of human-created original works, and “industrial” protection of investments and development of machine creation.
Recognizing AI creativity in the eyes of copyright law in either (or both) of the above mentioned ways would be a suitable solution for fashion houses that actually do create designs themselves – with or without the help of AI. It would not serve the business strategy of companies that just rely on copying successful designs of others. If the copyright system considered AI creativity worth protecting, it could also provide incentives to develop and utilize AI designers and promote creativity instead of imitation.
In addition to recognising detailed copyright issues such as authorship, originality and infringement, it is important to see the big picture: how will AI designer change the market due to low-cost mass production of works that look and function like human-created works. In the special case of the fashion industry, this includes the problem of overproduction that flourishes in the low-IP environment. Overproduction leads to overconsumption and an environmentally damaging outcome. Legal scholars have indeed pointed out that more fashion goods are consumed in a low-IP environment compared to how much would be consumed in a world of high IP protection. Since there is no doubt that the industry is struggling with major sustainability problems, a situation where more and more fashion goods are being consumed should not be seen as an ideal outcome. Copyright law should not encourage businesses into unsustainable activity by allowing (fast fashion) copycats to produce cheap knock-offs from AI-generated designs that, if they were created by a human author, would be protected by copyright.
SOME CONCLUDING REMARKS
Fundamentally, the copyright issues that the increasing use of AI in fashion designing brings to the table are actually not that different than the already existing copyright problems in the special context of the fashion industry. The underlying issue in both old and new problems is that fashion designs have difficulties in reaching the originality threshold. AI seems to add some more concerns to the already-existing problems that are caused by fashion’s nature as applied art, but essentially, the fundamental issues are the same. What is new is that AI might affect the scale in which these copyright problems show in the real world and make some of them more remarkable than before. In the current copyright environment, it seems that AI-generated fashion designs would be more vulnerable for copyists than human-created designs
This article is based on a research funded by IPR University Center and The Ministry of Education and Culture. The author’s original peer-reviewed research article “Fashion piracy and artificial intelligence—does the new creative environment come with new copyright issues?” is published on Journal of Intellectual Property Law & Practice Vol. 15 Iss. 3 and can be downloaded here.
 See e.g. K Raustiala & C Sprigman,‘The Piracy Paradox: Innovation and Intellectual Property in Fashion Design’ (2006) Virginia Law Review 92 (8), 1687–1777; CS Hemphill & J Suk, ‘The Law, Culture and Economics of Fashion’ (2009) Stanford Law Review 61(5); S Scafidi,‘Intellectual Property and Fashion Design’ (2006) Intellectual Property and Information Wealth 1(115) 115-131; S Teilmann-Lock & T Petersen, ‘Louboutin’s red sole mark and the logics of fashion’ (2018) Journal of Intellectual Property Law & Practice 13(11); and H Härkönen, ‘Muoti tekijänoikeudellisena teoksena: näkökulmia käyttötaiteen teoskynnykseen ja kopiointiin Suomessa’ [Fashion as a copyright protected work: perspectives to the copyright threshold and copying of applied art in Finland] (2018) Defensor Legis 6/2018. See also Journal of Intellectual Property Law & Practice Fashion Law Special Issue (JIPLAP 13(11), 2018).
 L Luce, Artificial Intelligence for Fashion (San Francisco: Apress Media LLC 2019), 129.
 Luce 2019, 125–126, 129
 J-M Deltorn & F Macrez, ‘Authorship in the age of machine learning and artificial intelligence’ (2018) Center for International Intellectual Property Studies Research Paper Series 2018(10), 5.
 Luce 2019, 131, 137.
 G Noto La Diega, ’Can the law fix the problems of fashion? An empirical study on social norms and power imbalance in the fashion industry’ (2018) JIPLAP (14)1; Härkönen 2018, 909, 910, 918–920 and H Härkönen, ‘Tekijänoikeus ja käyttötaide: EU-tuomioistuimen C-683/17 Cofemel -ratkaisun vaikutukset suomalaiseen tekijänoikeustraditioon’ [Copyright and applied art: CJEU resolution C-683/17 Cofemel’s implications on the Finnish copyright tradition’] (2020) Defensor Legis 1/2020.
 E Derclaye, ‘A Model Copyright/Design Interface’ in Estelle Derclaye (eds) The Copyright/Design Interface (CUP Cambridge 2018) 441.
 Noto La Diega 2018, 18, 22–23; C Sprigman, ‘Some Positive Thoughts about IP’s Negative Space’ in Kate Darling and Aaron Perzanowski (eds) Creativity without Law (NYUP New York 2017) 256 and K Raustiala & C Sprigman,‘The Piracy Paradox: Innovation and Intellectual Property in Fashion Design’ (2006) Virginia Law Review 92 (8), 1705–1717.
 J Ginsburg, ‘People Not Machines: Authorship and What It Means in the Berne Convention’ (2018) IIC 49, 134.
 E.g. K Raustiala & C Sprigman, The Knockoff Economy. How Imitation Sparks Innovation (OUP 2012) 21 and C Sprigman, ‘Some Positive Thoughts about IP’s Negative Space’ in Kate Darling and Aaron Perzanowski (eds) Creativity without Law (NYUP New York 2017) 251–258.
 J Grimmelmann, ‘There’s No Such Thing as a Computer-Authored Work – And It’s a Good Thing, Too’ (2016) 39 Columbia Journal of Law & The Arts, 410.
 Ellen MacArthur Foundation (2017) ‘A new textiles economy: Redesigning fashion’s future’and McKinsey & Company (2016) ‘Style that’s sustainable: A new fast-fashion formula’.
 Sprigman 2017, 256 and Raustiala & Sprigman 2006, 1722, 1724, 1728–1729, 1733.
 R Denicola, ‘Ex Machina: Copyright Protection for Computer Generated Works’ (2016) Rutgers University Law Review 69, 286, 287. See also: A Bridy, ‘The Evolution of Authorship: Work Made by Code’ (2016) Columbia Journal of Law & the Arts 39/395, 400, 401.
 A Alén-Savikko, R Ballardini & T Pihlajarinne ‘Tekoälyn tuotokset ja omaperäisyysvaatimus – kohti koneorientoitunutta tekijänoikeutta?’ [Automated content production and originality in copyright law – towards a machine-oriented regime?] (2018) Lakimies 7–8/2018, 993.
 Raustiala & Sprigman 2006, 1206–1207.