In this introductory in-depth analysis, we report six key findings on the use of Artificial Intelligence (AI) in the cultural and creative sectors (CCS).
Finding 1: AI challenges the creative value-chain in two ways: shifting services performed by humans to algorithms and empowering the individual creator.
Finding 2: AI-generated content challenges authorship, ownership and copyright infringement. New exclusive rights on datasets must be designed in order to better incentivise innovation and research.
Finding 3: European cultural institutions have rich datasets of cultural artefacts that could be made accessible to a larger audience. AI has the potential to create rich ways for users to navigate through cultural content. Good practices in AI for cultural heritage accessibility need to be formalised and shared among the European cultural networks.
Finding 4: The use of AI for media content brings up issues regarding cultural and linguistic diversity. Public policies and measures are required to prevent discrimination in AI-based distribution platforms.
Finding 5: AI governance is centralised, which has an impact in the CCS. Funding instruments are needed to support less-centralised, human-centred AI.
Finding 6: The Union supports a rich environment for AI-Art, resulting in the development of critical discourse on technology and AI by the public, which should be sustained in the long run.
In an increasingly digitised world, the quantity and heterogeneity of data grow exponentially. Data capture Internet browsing activity, money transfers, energy consumption, health diagnostics, media creation and consumption, to name a few. AI designates the set of machine learning tools that are able to crunch this massive quantity of data, extract underlying patterns, and make predictions of future events and behaviours. The capacity of these tools made a significant step forward in 2012 when deep learning based algorithms reached human-like abilities in image recognition tasks. Since then, AI has spread at a tremendous pace across research disciplines, solving several problems such as speech recognition, and then evolving from academic research to consumer services used by billions of people.
In this context, AI is often erroneously considered neutral as it appears to be no more than a set of sophisticated optimisation mechanisms used to achieve a task, e.g. classifying images, generating sounds or texts, with the best performance. However, AI builds on data that capture socio-cultural expressions represented by music, videos, images, text, and social interactions, and then makes predictions based on these profoundly non-neutral and context-specific data.
Culture therefore plays a central role in the use of AI at scale. Culture needs to be addressed in the general discourse and public policies about AI, and this has not systematically been the case so far. The CCS are not among the priorities of numerous recent white papers and reports presenting policy options and recommendations on AI in society. In the white paper published in February 2020 by the European Commission “On Artificial Intelligence – A European approach to excellence and trust”, the Commission supports the development of human-centric AI. A human-centric perspective on AI should embrace cultural diversity and should support human creativity, critical discourses, and artistic idiosyncrasies.
The objective of this introductory briefing is to examine the role and impact of AI in the CCS, by reporting on AI use-cases in the CCS. The methodology used in this report consists of interviews with actors in the field. We, then, conducted desk research on published reports and white papers relevant to the topic written by several stakeholders: the European Commission2, UNESCO1, and a report on AI in the Media and Creative Industries. Finally, we completed the survey with articles from digital humanities and AI, and selected media coverage from the past five years.
 Kulesz (2018) Culture, platforms and machines: the impact of artificial intelligence on the diversity of cultural expressions. UNESCO, Paris.
 Reported in the acknowledgment section
 Caramiaux (2019). AI in the Media and Creative Industries. White paper, NEM initiative. https://hal.inria.fr/hal-02125504/document
AI in the creative value-chain
Recent reports4 have shown that AI has entered the creative value-chain at every level: creation, production, dissemination, and consumption. AI can automate tasks within this pipeline that were thought only feasible by humans not long ago. Research on certain tasks is mature. Examples include image discrimination and generation as well as audio source separation and mastering. In addition, AI use-cases within this pipeline have direct market applications and have created incentives for the private sectors to embed this technology into their products and services.
Creation has been facilitated by advances in algorithmic generation of new media content with impressive quality (see more specific details in Appendix). AI-based generative models are applied to music, text, images, or videos. A driving force behind the development of content generation is to help automate time-consuming uncreative tasks that may sidetrack creators from their main task, which consequently increases costs. Automated journalism is an example and consists in automatically collecting data feeds from online content providers and populating templates, which are usually made by human journalists, with these data. Automatic journalism is primarily applied to routine stories, such as sports reports. Automatic journalism is already widely used, but different strategies can be observed throughout Europe. In Finland, the majority of media outlets have preferred in-house development of technology responsible for automatic generation of content, by investing in human resources. In the UK, the BBC uses an external platform where journalists can configure, to some extent, the generated reports. In France, Le Monde or France Bleu have chosen to completely out-source the process to an external company called Syllabs. Using out-sourced solutions is understandable but can raise questions when content generation algorithms are used to write more complex and less supervised reports. Reuters just released a prototype that creates sports reports generated directly from video content, without human supervision. How can quality assessment be implemented? What is the level of human control on such generated content? Will human intervention eventually no longer be required for automated content generation? The frontier between AI used to assist or AI used to replace content creators is fuzzy. Certain players are positioning themselves explicitly. Antescofo, a Paris-based start-up, proposes a tool for automatic accompaniment in classical music, which does not intend to replace orchestras but rather simulates them in pedagogical or rehearsal scenarios.
Reducing costs through automation also appears throughout the audiovisual production chain, raising important questions. For example, music mastering usually occurs in professional studios and can be expensive for artists that are not yet established. AI-based tools can help these artists create high-quality musical productions that they can then use to approach labels. Europe is attractive for AI-driven audio engineering with hotspots in Berlin (Landr), Barcelona (Dolby labs) and Paris (Spotify France). We also found similar solutions in the movie sector, such as automated editing, although most of the companies proposing these services are located outside Europe. Here, disrupting the creative value-chain means that algorithms internalise steps that were previously handled by experts. Editing, producing, and mastering are tasks requiring specific skills and equipment, such as, for instance, a professional music studio. Automation in the creative process can therefore reduce dependency on external expertise, providing creatives and artists with “low entry fee” tools. However, in doing so, it can collaterally damage expertise that was initially needed to create the datasets from which AI-based systems were built. The impact of this disruption on Cultural and Creative Industries (CCI) is unclear. To what extent will the algorithms performing expert tasks incentivise research and innovation? What is the expected deskilling within the creative value-chain? A recent report by the World Economic Forum (WEF) forecasts a growing number of jobs in CCI, facilitated by increased access to technology. On the other hand, calling upon expert skills within the creative process may remain the preferable choice for bespoke demands.
 Audionamix is a French company proposing solutions for audio source separation for creation and production https://audionamix.com
 Danzon-Chambaud (2020) https://larevuedesmedias.ina.fr/journalisme-automatise-robot-media
 WEF (2018). Creative Disruption. http://www3.weforum.org/docs/39655_CREATIVE-DISRUPTION.pdf
AI and copyright
In a time when AI is used to generate content, new questions also arise with respect to rightsholding. On the one hand, copyrights for AI-generated outputs (music, images, videos) put into question the existing notion of authorship. This has been illustrated in recent projects using AI-powered algorithms to generate paintings. The Next Rembrandt project produced a painting generated from Rembrandt’s body of works. The impressive result is a speculation about what could have been Rembrandt’s next painting. Similarly, the Portrait of Edmond Belamy is a painting created by an algorithm called Generative Adversarial Network (GAN, see appendix). The artist’s signature is the equation identifying GAN-like algorithms. Is the author of the painting the AI-powered algorithm, the team(s) putting together the system, or the author of the original paintings which were used as a training dataset?
If the author of an AI-generated work cannot be legally identified, the work may not be protected by copyright. In the UK, the Copyright, Designs and Patent Act defines the author of a work as follows: “In the case of a literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken”. In Germany, however, “only the author’s own intellectual creations constitute works within the meaning of this Act”, therefore AI works cannot be protected. Alternatively, may the system be identified as the author? Under the current legal framework, an AI-based system may not be considered as an author. However, recent innovations challenge this state of affairs. One example is AIVA, a start-up company and the name of a music-generating AI-based system that the company developed, which has recently been given the status of composer by the SACEM, France’s authors’ rights society. Then, depending on the type of subscription, an AIVA user can own the copyright of the work generated with AIVA.
The question of ownership must also be addressed. Does an AI-generated output belong to the person who designed the dataset used to train the system, the person who trained the system, or the person who built and implemented the system? Ownership attribution depends on stakeholders (human creators, technology or data providers). As an example, the Google Arts & Culture in Paris proposes art residencies where artists use the company’s AI technology in their projects. Artists are assisted by Google engineers, or freelance artists contracted by the company, to help them handle the technology. After the residency period, technological outcomes (algorithms, applications, or user interfaces) are owned by Google, while the created works (images, videos, or music pieces) are owned by the artists in residency. In this context, AI-based systems are seen as tools, regardless of level of human involvement. Another example can be taken in the music industry. Endel, a Berlin-based start-up company and the name of an app that generates personalised soundscapes to enhance focus or encourage relaxation, is the first AI-based system signing a major label distribution deal with Warner Music Group. Warner is not in control of Endel’s algorithm, but shares half of the royalties with the company.
AI-generated creative content also raises new questions regarding copyright infringement, which “is the use of works protected by copyright law without permission for a usage where such permission is required”. As pointed out in Frosio (2019), deepfakes, artificially generated media in which a person’s likeness in a video or image is replaced by someone else’s, challenge the current legal framework. The notion of copyright infringement is also discussed in the music industry. An AI-based system could be trained on an artist’s music tracks so as to be able to generate new tracks that resemble that artist’s music. Would it be a case for copyright infringement? The question is fraught. The way an AI-based system uses training data to generate content is neither copy-pasting nor sampling. These systems involve learning mechanisms that could provide the machine with creative skills (as commented about AlphaGo). Consequently, it becomes difficult to prove that an algorithm was designed to copy an artist, as well as that an algorithm was trained on an artist’s music it infringes on.
These examples put into question the notion of originality. If someone in a video is replaced by someone else’s likeness, is the new work original? If an algorithm generates a music track that resembles an existing artist’s music, is the AI-generated work original? This is an open question directly stemming from recent advances in AI and their applications to the CCS. Since most AI-based systems involve training procedures relying on fixed datasets, it is not obvious to what extent the generated content can be construed as original with respect to these datasets. These questions still need to be addressed in legal terms. In that regard, the EU regulation for AI-generated content is at its infancy, and last year the Commission issued a literature review on Intellectual Property and AI.
On the other hand, copyrights for the inputs of those algorithms pose a very important, although less addressed, challenge. AI algorithms are based on machine learning techniques that require processing datasets for classification, recommendation, or generation, but these datasets are often proprietary. What type of Intellectual Property (IP) rights for data would allow machine learning and data mining?
There is an ongoing discussion in Europe about how copyrighted content, sound and movie catalogues for instance, can be used by non-profit and for-profit stakeholders in order to enable research and foster innovation. Should data be protected with new exclusive IP rights? Currently, Union law allows fewer exceptions on copyrighted content than laws in the USA, China, or Australia. It has recently evolved towards exceptions for non-profit purposes, meaning that research centres and universities can apply data mining techniques i.e. the crunching of massive datasets of music, videos, or images to look for patterns, on copyrighted content, with the aim to incentivise research and innovation. However, stakeholders argue that there is still a need to allow for-profit stakeholders to have access to proprietary datasets in order to efficiently foster innovation. The music industry is a telling example. In a 2016 article, an investor in the music industry counted the number of music companies that have succeeded in achieving venture returns for their investors. The success rate was about 4%. The author wrote that this “is a direct result of the high royalty rates incumbent upon startups who wish to license digital music for use in their apps”. This observation is all the more relevant today, when the majority of newly created music start-up companies involve AI, for which the necessary datasets have expensive licences, or require to negotiate specific agreements.
 Frosio (2019) Artificial Intelligence and IP – Mapping Legal Challenges for the European Digital Single Market. http://www.iprhelpdesk.eu/sites/default/files/newsdocuments/AI%20and%20Legal%20Challenges%20in%20the%20DSM%20%20%28EU%202019%29.pdf