The use of protected content to “train” AI systems such as chat GPT is increasingly widespread and such content may pollute the streaming platforms with productions generated by artificial intelligence to the detriment of the creativity of authors and artists.
Music: ChatGPT won’t kill creativity, but it will be very good for marketing
The alarm was raised a few days ago by Financial Times, which gave the news of an initiative undertaken by the largest company in the record industry, Universal Music (UMG), the label of Taylor Swift, U2, Katy Perry and in Italy of artists such as Vasco Rossi, Tiziano Ferro, Elodie and Blanco, with the sending of a series of communications to the main streaming platforms asking “to stop access to their music catalog for developers who use it to train AI technology”. “We will not hesitate to take steps to protect our rights and those of our artists,” UMG wrote to online platforms in March, in emails seen by the FT. “This next generation of technology poses significant problems,” said a person familiar with the situation. “Much (of generative AI) is trained by ingesting data from the most popular global pop music,” UMG added.
Legal uncertainty about text and data mining
The topic of machine learning is related to mining protected content which is the foundation on which every AI application is built. Against this background, one can clearly associate this type of activity with the so-called text and data mining which was the subject of the recent Copyright directive of the EU, the 790/2019.
The directive established that, in addition to being relevant in the context of scientific research, the text mining techniques and data are widely used by both public and private entities to analyze large amounts of data in different areas of life and for various purposes, including public administration services, complex business decisions and the development of new applications or technologies. Rightholders should retain the possibility to grant licenses for uses of their own works or other subject-matter outside the scope of the mandatory exception provided for in this Directive on text and data mining for scientific research and existing exceptions and limitations set out in Directive 2001/29/EC. At the same time, it should be taken into account that users may find themselves in a situation of difficulty in text and data mining. legal uncertainty as to the possibility of making or not, on works or other materials to which they have had legal access, reproductions and extractions aimed at extracting text and data, in particular when the reproductions or extractions made for the purposes of the technical process may not satisfy all the conditions under the current exception for acts of temporary reproduction under Article 5(1) of Directive 2001/29/EC.
In order to increase legal certainty in such cases, and to encourage innovation also in the private sector, this Directive should provide, under certain conditions, for an exception or limitation for reproductions and extractions of works or other subject-matter , for text and data mining purposes, and allows you to keep copies made long enough for text and data mining purposes. Such exception or limitation should apply only in case of legal access to the work or other materials by the beneficiary, even when the work or other materials have been made available to the public online, and to the extent that the right holders have not properly reserved the rights to make reproductions and extractions for the purposes of text and data mining.
The footholds of the EU Directive
In the case of content made publicly available online, it should be considered appropriate to reserve such rights only through the use of tools that enable automated reading, including metadata and terms and conditions of a website or service. The reservation of rights for text and data mining purposes should be without prejudice to other uses. In other cases it may be appropriate to reserve rights by other means, such as contractual arrangements or a unilateral declaration. Rightholders should be able to apply measures to ensure that their reservations in this regard are respected. Such an exception or limitation should leave unchanged the mandatory exception for text and data mining for scientific research purposes set out in this Directive, as well as the current exception for acts of temporary reproduction set out in Article 5(5). 1 of directive 2001/29/EC.
As can be clearly seen from this provision, while not mentioning the theme of artificial intelligence and even less the generative one, the European legislation offers an evident foothold to protect the contents from an unlimited mining activity.
Italian law
In Italian legislation, this provision has found its definition in art. 70 ter lda which defines text and data mining, the extraction of text and data, as “any technique automated aimed at analyzing large quantities of text, sound, image, data or metadata in digital format with the purpose of generating information, including patterns, trends and correlations”.
At the art. 70 quater allows the extraction of text and data in general, by anyone, even for the mere purpose of profit.
Therefore, enterprises and developers intending to use copyrighted works to train a generative AI system will have to meet three criteria:
- obtain legitimate access to the data;
- verify that the right holders have not reserved the right to make the Reproductions for TDM purposes;
- keep copies made only for as long as necessary for TDM purposes.
The uses of machine learning (ML) associated with training and text and data mining
Let’s see in detail what are the most widespread uses of machine learning (ML) systems connected with training and text and data mining at the moment. In regards to musical compositions the generative AI can build a musical composition, with the output composition being further edited by an artist or turned into a recording rendering the composition through different types of instruments and MIDI synthesizers (Musical Instrument Digital Interface ). Where these AI compositions use ML, the model would be trained on pre-existing compositions in notation or MIDI form. Generative composition models could be co-trained on composition data, along with text tags, to allow text-based terms to drive the generative process.
Music recordings
On the music recordings front, they are created directly as an output of the AI. One method is through tight integration between the AI compositions and computer-played instruments (see above). But where sounds are synthesized directly as AI output, this requires training ML models on multiples of existing recordings.
Generative recordings can often capture the timbre and expressive feel in a way that is not accessible in generative compositions. This is because a model trained on existing recordings is able to “absorb” the sonic qualities of the training material. In some cases, models can also be co-trained on registrations and text tags, to allow text-based terms to drive the generative process.
Additional uses of AI include sampling, transcription, beat or instrument generators, remixing, etc.
Other uses of AI
Also not to be overlooked are software tools for segmenting an existing recording intro with separate plans, such as a vocal plan and a plan for instrumental or even individual instruments. which in return can be used for remixing or creating inputs for the AI, for example for allow training of a speech model on the voice of a particular singer. Training templates created using artist voice, extracted from the above process, can be used to create new artist tracks and new artist content and are gaining more and more popularity.
Finally, other uses involving AI are the Beat Matching which is a process used by DJs to blend tracks. Intelligent tempo matching tools are available, although this isn’t really a form of creative AI.
Graphics/Videos
Graphics/Videos. The advent of tools like Dall-E, Stable Diffusion, Midjourney and others, is paving the way for the creation of artwork such as album covers or NFT graphics. There are many copyright issues involved, and several stock image providers do not distribute AI-generated images due to open questions and risks. Artificial intelligence tools are emerging to generate high frame rate videos.
The lyrics would be an integral part of the composition process. Algorithms like GPT3 are highly capable of generating coherent text, including poetry.
Deepfakes
Deepfakes. In a musical context this refers to the creation of sounds, usually vocals, based on deepfake models of a specific artist’s voice or style. This technology can also be used to simulate an artist’s likeness.
Conclusions
We are therefore faced with many innovations that will upset the music sector and which must find, on the one hand, as we have seen, a legislative regulation to protect those entitled and on the other a content protection garrison to prevent derivative works generated by AI without authorization pollute the original creative process.
As a Universal Music executive commented in an interview with the FT: “We have a moral and commercial responsibility to our artists to work to prevent the unauthorized use of their music and prevent platforms from ingesting content that violates the rights of artists and other creators. We are confident that our platform partners will want to prevent their services from being used in ways that harm artists.”