House of Lords Communications and Digital Select Committee inquiry: Large language models
ABOUT THE BPI
The British Phonographic Industry (BPI) is the trade body for the UK’s recorded music industry and is celebrating its 50th anniversary in 2023. Our membership consists of the three ‘major’ record companies, Sony Music Entertainment UK, Warner Music UK and Universal Music UK, and nearly 500 independent British music companies.
We also have a charitable arm – the BRIT Trust, which funds the BRIT School, and as well as organising the annual BRIT Awards, we bring together the Mercury Prize for ‘Album of the Year’. The BRIT Awards are a vital means by which the school and other charitable initiatives are funded.
The BPI promotes and protects British music at home and abroad. We bring intellectual property (IP) infringement litigation in the UK on behalf of our members and all the performer and record label members of the collecting body PPL. Those performers and record labels collectively hold the IP rights of around 99% of the recorded music legitimately consumed in the UK.
The BPI in-house Content Protection Unit devises and delivers technical and practical strategies to protect digital and physical music content, including bringing precedent-setting legal cases, delivering a world-leading website-blocking programme, removing infringing content from the internet, and disrupting illegal entities and traders in collaboration with law enforcement agencies.
Overseas, we organise numerous trade missions, and we run the Music Export Growth Scheme (MEGS), a partnership between the recorded music industry and the UK Government which looks to give a platform to independent British artists internationally and boost UK music exports through the issue of grant funding.
We administer The BRIT Certified Awards (the iconic Platinum, Gold and Silver Awards Programme), co-own the Official Charts, and own and organise the annual BRIT Awards with Mastercard and Mercury Prize with FREE NOW.
BPI provides valuable insights, training and networking for its members with its free masterclasses and presentations and through its Music & Tech Springboard Programme, BPI Insight Sessions and our authoritative yearbook and reports.
EXECUTIVE SUMMARY
1 We are grateful to the Committee for launching this inquiry and inviting written evidence. The music industry is already embracing the opportunities and confronting the risks of AI, including generative-AI powered by foundation models, and we are participating fully in these discussions to ensure that AI enriches our creative, cultural and technological future.
2 The Committee states that its overarching objective is to examine “what needs to happen over the next 1-3 years to ensure the UK industry can respond to the opportunities and risks posed by large language models”. Our short answer, as it relates to the UK’s world-leading music industry, is that the Government must maintain a robust IP and personality rights framework that:
3 Government must also require:
4 These recommendations sit at the heart of the Human Artistry Campaign to which we are a signatory. They are fundamental to ensuring that AI remains an exciting tool that enhances but does not replace human artistry - and that enriches and does not impoverish our creative and cultural future.
5 We have only answered questions where we will feel we have something to contribute from a music industry perspective. Throughout our responses to these questions, we will use terminology which we would like to take the opportunity to clarify at the outset. Foundation models that power generative-AI are trained on inputs consisting of vast quantities of data comprising text, image, audio and video and can produce outputs consisting of text, image, audio and video. Large Language Models (LLMs) are a subset of foundation models.
6 At their inception, LLMs used text only as inputs and outputs, however, they are already evolving to be ‘multimodal’ – and arguably the most well-known LLM - OpenAI’s GPT series - can now operate across text, image and audio simultaneously, blurring the distinction between an LLM and a foundation model. The music industry is interested in any AI that either inputs or outputs music or music-related content including text, image, audio and video including vocal clone models. As our interests are not limited to text based LLMs, we refer to “foundation models” rather than LLMs throughout our response.
QUESTIONS
Capabilities and trends
What are the greatest opportunities and risks over the next three years?
a) How should we think about risk in this context?
7 This is a fast-moving area and the music industry is already embracing the opportunities and confronting the risks of generative AI. Below we seek to articulate these further.
Opportunity #1 – for the UK to continue to be a world leader in both creativity and technology
8 On a macrolevel, the Government does not need to choose to champion the creative industries or technology – it can and should champion both - by supporting IP. The creative industries and emerging technology are rightly identified by Government as two of the five sectors which should be priorities for economic growth. The UK is a global leader in the creative industries, which are worth around £120bn in GVA – more than the combined value of the aerospace, automotive, life sciences and oil and gas sectors.[1] The UK is the world’s third largest music market and the second biggest exporter. As music has become global, competition has also become global, particularly as music from Latin America and South Korea becomes increasingly popular. Notwithstanding this competition, one in ten tracks streamed globally is by a British artist, adding significantly to the soft power of the UK on the international stage.
9 The UK is also a global leader in artificial intelligence (AI), ranking third for AI research and development and home to a third of Europe’s AI companies - double that of any other European country. These two sectors are both thriving to date, thanks in large part to a robust and carefully considered IP framework that is pro-investment, pro-innovation and that largely achieves a good balance where both can grow without stifling one another. Striking this balance is particularly important given the creative industries’ strong economic and cultural track record, and strong future growth potential[2] - and because the sustainability and competitiveness of the UK’s creative industries and the development of meaningful, human-centric AI are inextricably linked and entirely inter-dependent.
10 The development of safe and meaningful AI systems, especially foundation models, depends on the continued availability of high-quality inputs, the creation of which require a robust copyright framework that continues to incentivise the creation of and investment in new works. In other words, the interests of artists, investors, fans and AI developers are best served when the ability to create and invest in original human creativity is safeguarded by a robust IP framework, based on strong exclusive rights.
Opportunity #2 – Creativity and technology can thrive in a partnership based on recognition of IP
11 On occasion, the Government and regulatory bodies have characterised IP authorisations as a barrier to entry or a brake on AI innovation[3], in fact nothing could be further from the truth. When creative innovation and technological innovation go hand-in-hand, we see growth – as demonstrated by the success of music streaming. The partnership between music companies and paid streaming platforms has delivered unparalleled value for consumers and reversed a decade of decline in revenues caused by prolific music piracy.
12 With the advent of AI, the music industry is once again embracing the opportunities presented by new technology. AI is already enhancing how artists make music, how platforms organise music in playlists and recommendations, how artists connect with fans and how fans access music. It is also enhancing and accelerating many of the tasks behind the scenes from production, mixing, mastering, scheduling, planning, administration, understanding consumer trends and tackling piracy.
13 There have been some high-profile examples of ‘deep fake’ tracks – including where the voice of a well-known artist has been merged with the melody of another. We discuss the implications of this when it occurs without authorisation in paragraph 14 below. However, it is worth noting that discussions between major music companies and AI companies are already underway to forge partnerships for the authorised use of artists’ music and voices in generative-AI tools underpinned by commercial licensing and artist consent or ‘opt-ins’.
Risk #1 – The Government could make policy choices to facilitate generative-AI at the expense of human creativity
14 Generative-AI begins with, and relies upon, access to ‘data’ to train the foundation models. Where such data consists of works protected by copyright, rightsholders must have the discretion to choose whether to grant or refuse permission for works to be used, to determine an appropriate level of compensation and to have a say in the applicable economic model.
16 These negotiations – which are already underway – must be conducted in a free and competitive market. Policy choices that remove competition or distort the market – such as compulsory licences, collectivisation of licensing, or exceptions without evidence of market failure – must be avoided at all costs. The music industry, along with other creative and IP-rich industries, was astonished that last year the UK government was contemplating the introduction of a ‘text and data mining exception’ that would have allowed AI developers, including some of the world’s largest companies, to ingest all copyrighted works ever produced without permission.[4] This exception would have removed any agency artists have in the use of their work, as well as their ability to earn from the use of their music as we enter the AI era. The impact on the creative industries would have been existential. Thankfully this policy has now been abandoned. Any suggestion that it be reconsidered must be resisted in the strongest possible terms. We are very grateful to this Committee for a recommendation to this effect in your previous report ‘At risk: our creative future’.
Risk #2 - AI developers could fail to respect IP laws or keep records of training inputs
17 We understand that in this early phase of generative-AI, many, if not most foundation model developers effectively ignored the law by scraping content needed for training purposes from the internet[5], including but not limited to, content available on User Uploaded Content (UUC) platforms, which in almost all cases has not been authorised for such usage. Those developers have then made available to the public, and in some cases commercialised, these systems meaning that the creative community has effectively subsidised the development of AI by technology companies.
18 This is not a sustainable model. The failure of foundation model developers to appropriately engage with rightsholders and use copyrighted content without permission infringes copyright laws and causes harm to artists and consumers. We urge the Government and law makers to be unequivocal that ingestion without authorisation constitutes infringement to ensure that as existing generative AI models evolve and new ones emerge, it is clear that developers must seek authorisation.
19 Record keeping is another fundamental requirement for developers of generative-AI. Accurate, detailed and transparent record keeping of training data and how that material has been used within the model to produce outputs is essential to underpin and aid enforcement of IP and so that licences can be validated and the correct artists and rightsholders can be compensated. As a first step, the Government should ensure that this is mandated through existing regulatory bodies, and by the Intellectual Property Office (IPO), possibly through the vehicle of a properly focussed Code of Practice under discussion as the output of the AI and Copyright Working Group. In addition, AI developers should label AI-generated outputs to inform consumers and disclose which inputs were used to train the AI system. Full and transparent records will allow consumers to differentiate between AI-generated content and human artistry.
Opportunity #3 – AI can help fans to find the music they love on streaming platforms
20 AI is already integrated into all the leading music streaming platforms, powering music discovery, innovative functionality and introducing fans to new music through individually curated playlists, recommendations and AI-generated personal DJs. The role of AI in helping fans find authentic music that they genuinely want to listen to is beneficial to both consumers and artists. It is set to play an increasingly important role in influencing customer choice as the number of tracks added to music streaming platforms continues to grow exponentially,[6] and as music streaming platforms compete to win and retain customers by providing ever better curation and functionality.
Risk #3 - High volumes of AI outputs risk flooding an already competitive music streaming market, generating fake streams, diluting human artist earnings and disorienting consumers
21 The Competition and Markets Authority’s recent market study in relation to music streaming found that a surge in the number of artists who are uploading their music to streaming platforms is impacting artist income, as an increasing numbers compete for a relatively fixed revenue pool.[7] This situation could be severely compounded if streaming services were to become flooded with illegally cloned tracks coupled with bots programmed to generate artificial plays in a fraudulent bid to earn royalties, that would otherwise be paid to human creators and rightsholders. In this scenario, artists could find themselves competing for a share of a limited royalty pool against illegal clones of their own recordings and voice, unauthorised adaptations of existing tracks, derogatory treatment of their original works and third parties passing off new materials as their own creative output, all to the detriment of the artists’ identity and reputation. Spotify recently reported that it has removed tens of thousands of AI-generated tracks from its platform this year.[8]
22 Similarly, AI is also capable of learning to defeat any countermeasures deployed by platforms to detect stream manipulation, and we expect that unscrupulous users may use AI to execute the end-to-end process of devising, implementing and misappropriating revenue as part of stream manipulation programmes. Meanwhile consumers could be left disoriented and discouraged, unable to find the recordings they actually want to listen to, and unable to discern whether the recordings they do listen to are genuine or fake or whether their streaming subscription fees are being paid to human artists or an AI operator.
Risk/Opportunity #4 - Foundation models could facilitate music piracy and streaming fraud by pointing users to unlicensed sources and unauthorised tools or they could play an educational and legitimate signposting role instead
23 The BPI’s Content Protection department has recently undertaken an analysis of the responses provided by ChatGPT (Open AI), Bard (Google), and Bing AI (Microsoft) to hypothetical users seeking information on the following areas, which severely impact the music industry by causing direct losses or unfair distribution of revenue in the case of streaming manipulation:
On average, 50% of the responses appeared to be enabling piracy or fraudulent behaviours. Bing returned the highest level of responses facilitating infringement, followed by Bard and ChatGPT.[9] The percentage of infringing responses grew further when follow-up questions were asked. In some cases, the chatbots refused to provide answers, citing copyright and ethical concerns. At times, they pointed users to legitimate streaming services or gave them tips on how to legally promote streams without resorting to bots.
24 While chatbots can facilitate music piracy and streaming fraud, the analysis shows that they can also correct illegal behaviours by providing educational responses. As foundation models continue to evolve and expand their userbase, we value the educational action they could perform in society. Creative companies and search engines established a successful model for this in 2021 with a UK Code of Practice under which search engines took action, globally, to reduce the prominence of illegal websites in search results. In addition, as part of the Code of Practice, Facebook and YouTube took steps to reduce the visibility of piracy tutorials. As AI operators will be providing a similar role for consumers or in some cases sit within the same companies already providing information, they could provide a similar educational function. Obligations to this effect could be incorporated into a properly focussed Code of Practice such as that under discussion in the IPO led working group on AI & IP.
Conclusion on risks/opportunities: The UK should capitalise on the opportunities and eliminate or minimise the risks
25 Unless addressed through the robust application and enforcement of IP and personality rights, the risks identified above could lead to a race to the bottom in relation to artist earnings; record labels’ ability to invest in the human talent of tomorrow; the quality of our creative expression as a society and an erosion of the connection between an artist and their fans. Ultimately, this would result in less original content being created for generative-AI to draw upon which could lead to a closed eco-system where AI can only continue to generate from the existing AI-generated content, and in the long-term human creativity is sidelined. Instead, law makers should capitalise on the opportunities, by safeguarding the UK’s existing copyright framework, together with the respect for personality rights, and by developing additional rules to ensure that AI developers are transparent about the content their systems ingest and how they are used. The result ought to be a fair music marketplace, where the developers of foundation models and those deploying them operate and can compete on a level playing field and creators and investors maintain the incentive to make new music. More detailed suggestions on how to do this are detailed below in our answer to the ‘Domestic Regulation’ section.
Domestic regulation
How adequately does the AI White Paper (alongside other Government policy) deal with large language models? Is a tailored regulatory approach needed?
26 The Government’s AI White Paper established a series of principles which existing regulators are tasked with applying to the sectors they oversee. The principles are:
We applaud these principles which are embodied in our policy recommendations for music as follows:
What are the non-regulatory and regulatory options to address risks and capitalise on opportunities?
27 The government’s approach as outlined in the White Paper is that existing regulators apply the principles. As mentioned above, the creative and IP-based sectors do not have or require a regulator. They rely upon the UK’s IP framework which provides the foundation in law that is needed in order to forge market-based solutions in partnership with entities which seek to use copyright material – in this case AI developers. These negotiations – which are already underway – must be conducted in a free and competitive market. There is no need for regulatory or statutory intervention for the creative sectors at this point. Indeed, any policy choices that remove competition or distort the market – such as compulsory licences, collectivisation of licensing, or exceptions – must be avoided at all costs.
International Context
How does the UK’s approach compare with that of other jurisdictions, notably the EU, US and China?
29 Some broadly agreed principles have already emerged at an international level, enshrining IP at the heart of policy development on AI. Our proposal that respect for human artistry be added as a sixth principle in the UK would align with these international declarations. In addition, the UK Government must be mindful that liability rules need to be robust enough to avoid AI operators being able to escape liability through the use of complex AI value-chains that can involve many different actors and/or forum shopping. For example, AI operators may seek to evade securing licences from rightsholders by accessing content in a jurisdiction in which they claim an exception to copyright applies and subsequently releasing their systems for global use, including in jurisdictions with more strict rules, which would undermine national copyright law and sovereignty. This is another reason why a robust requirement of record keeping by AI developers and operators across the chain is critical to upholding rights and enabling enforcement.
30 As the Committee progresses its work in this area, reference to the following international developments may be useful:
- Material protected by copyright or related rights must not be used in an AI process without prior authorisation from the rightsholder.
- Meaningful record keeping obligations must be imposed on developers, providers and operators using material protected by copyright or related rights in AI processes.
- Material solely generated by AI, with no human creativity, should not be rewarded with copyright (or related rights) protection.
“We support the development of tools for trustworthy AI through multi-stakeholder international organizations, and encourage the development and adoption of international technical standards in standards development organizations through multi-stakeholder processes. We recognize the need to immediately take stock of the opportunities and challenges of generative AI, which is increasingly prominent across countries and sectors, and encourage international organizations such as the OECD to consider analysis on the impact of policy developments and Global Partnership on AI (GPAI) to conduct practical projects. In this respect, we task relevant ministers to establish the Hiroshima AI process, through a G7 working group, in an inclusive manner and in cooperation with the OECD and GPAI, for discussions on generative AI by the end of this year. These discussions could include topics such as governance, safeguarding of IP rights including copy rights, promotion of transparency, response to foreign information manipulation, including disinformation, and responsible utilization of these technologies. We welcome the Action Plan for promoting global interoperability between tools for trustworthy AI from the Digital and Tech Ministers’ Meeting. We recognize the potential of immersive technologies, and virtual worlds, such as metaverses to provide innovative opportunities, in all industrial and societal sectors, as well as to promote sustainability. For this purpose, governance, public safety, and human rights challenges should be addressed at the global level. We task our relevant Ministers to consider collective approaches in this area, including in terms of interoperability, portability and standards, with the support of the OECD. We express our interest in possible joint cooperation in research and development on computing technologies. We also task our relevant Ministers to consider ways to further promote digital trade.”
‘Given that generative AI technologies are increasingly prominent across countries and sectors, we recognise the need to take stock in the near term of the opportunities and challenges of these technologies and to continue promoting safety and trust as these technologies develop. We plan to convene future G7 discussions on generative AI which could include topics such as governance, how to safeguard IP rights including copyright, promote transparency, address disinformation, including foreign information manipulation, and how to responsibly utilise these technologies. These discussions should harness expertise and leverage international organisations such as the OECD to consider analysis on the impact of policy developments and GPAI to conduct relevant practical projects.’
September 2023
12
[1] As quoted in the House of Lords Communications & Digital Committee’s January 2023 report ‘At risk: our creative future’ – original source here.
[2] https://musically.com/2023/06/29/goldman-sachs-updates-its-music-industry-revenue-forecasts/
[3] See for example : Patrick Vallance authored Pro-Innovation Regulation of Technologies Review
[4] The Patrick Vallance authored Pro-Innovation Regulation of Technologies Review also recommended a text and data mining exception as a means to the UK securing a competitive advantage internationally in the field of AI – apparently without considering the impact on the UK’s world leading creative industries, the dependence of generative AI on sustainable sources of original content or the fact that international agreements have already established clear principles to safeguard IP into the era of AI (see our response to question 6).
[5] https://www.reuters.com/legal/getty-images-lawsuit-says-stability-ai-misused-photos-train-ai-2023-02-06/
[6] https://www.musicbusinessworldwide.com/there-are-now-120000-new-tracks-hitting-music-streaming-services-each-day/
[7] “The surge in artist participation has also impacted upon remuneration. Prices for music streaming services have been relatively stable for some time and tend to take the form of flat monthly fees. This means that with more artists and more streams being played, the average value of each stream and the average earnings per artist fall.”
[8] AI’s disruptive forces are rapidly reshaping the music industry | Financial Times (ft.com); Spotify ejects thousands of AI-made songs in purge of fake streams | Financial Times (ft.com)
[9] ChatGPT is only trained on data up until September 2021. This had a particular impact on AI-related queries, which returned fewer infringing answers in ChatGPT as most advances in AI-generated music have occurred post September 2021.