The Invisible Value of AI-Generated Music and the importance of prompts.

After many insightful conversations with researchers, songwriters, publishers, collecting societies, rights holders and AI companies, one observation has become increasingly clear: we need a better understanding of where value is created throughout the AI exploitation chain, and how that value should be divided and compensated.

The emerging consensus from our discussions is that there are three distinct value pools to consider: Training, Prompting and Outputs.

Understanding the mechanisms behind each of these pools, and the economic value associated with them, is fundamental to establishing a sustainable licensing framework for generative AI.

Each deserves its own discussion. In this article, we focus on the prompt pool, an area we believe has received far too little attention.

Understanding what happens between a prompt and an output

Firstly, it is important to recognise that AI models do not all operate in the same way. Their architectures, training methods and generation processes can differ significantly. This is precisely why establishing common principles for determining value is essential.

When a user prompts a generative AI model, whether to produce text, music or another form of content, the model generates a response based on patterns learned during training. Depending on the system, several underlying models, components or processing stages may contribute to the result.

Why is this important?

Because a prompt may explicitly reference one song, one songwriter or one artist, but that does not necessarily mean the resulting generation draws exclusively from that reference.

This is where we need to distinguish between what the user prompts, what the system uses to generate a response, and what eventually appears in the finished output.

These are three very different things. The prompt itself therefore cannot be treated as the single source of truth for attribution.

Consider something most of us have experienced. You ask an AI model to draft an email. It produces a first version, but you are not entirely happy with it. You adjust the prompt, change the tone, ask it to rewrite certain sections and continue until you have something you are satisfied with.

Generative music follows a comparable iterative process.

A user may generate several versions of a song, adjusting melodies, instrumentation, vocals, structure or style before arriving at the final recording.

For the purposes of this discussion, we will refer to these intermediate generations as archetypes.

Each archetype represents another stage in the generative process, another instance of the model's capabilities being used, and potentially another point at which economic value is created.

But not every archetype becomes a fully formed output.

And that distinction matters.

One prompt pool, two different forms of value

We believe the prompt pool should be divided into two distinct sub-pools:

1. Archetypes — the value generated during the creative and iterative process.

2. Outputs — the value associated with the resulting content that is selected, used, distributed or commercially exploited.

One way of conceptualising this distinction is to compare private, non-commercial activity with commercial exploitation.

An archetype may be generated, listened to, modified and discarded without ever being publicly distributed. The final output, by contrast, may be uploaded to a streaming service, synchronised with audiovisual content or monetised through social media.

This does not mean that archetypes are legally equivalent to private use, or that all outputs are commercial. Rather, the distinction helps illustrate that different activities may justify different licensing structures and economic valuations.

Both stages are connected, but they do not necessarily carry the same value.

More importantly, the references, influences and potential attribution identified in the final output may differ considerably from those involved in earlier generations. Some may no longer be detectable at all.

If we only measure the final output, we risk overlooking a substantial part of the value chain.

And if we cannot identify every reference point involved in the generation process, does that mean rights holders should not be compensated?

Absolutely not.

We will explore the challenges of attribution, and the limitations of relying on output similarity, in a separate article.

What happens when attribution disappears?

During our panel discussions in Palma, a colleague from Sureel AI shared an observation from a controlled testing environment that fundamentally reinforced the importance of examining the entire generation process.

The team had been studying the creation of AI-generated music, following the process from the initial prompt through to the finished output.

In one example, their analysis identified references associated with songwriters 1 and 2 during the earlier stages of generation.

However, towards the end of the process, when the song was approximately 97% complete, the identifiable references had shifted towards songwriter 3.

Think about what that means.

If we examine only the final song, our attribution analysis may identify the strongest similarity with songwriter 3.

But what happened to songwriters 1 and 2?

Now imagine that the user is not entirely satisfied with the result and prompts the system again.

The next generation begins from the previous output or its associated instructions, but the resulting music evolves further. New similarities may emerge with songwriters 4 and 5, before the final song develops a particularly strong resemblance to the work of songwriter 6.

The user loves the result and uploads the recording to a streaming platform or uses it in social media content.

When rights holders subsequently attempt to establish attribution by analysing that final recording, songwriter 6 may emerge as the strongest match.

Songwriters 1, 2, 3, 4 and 5 may be entirely absent from the final similarity analysis.

What it demonstrates is something important, an analysis of the finished output may tell us very little about the intermediate stages of the song generation.

And if the commercial licensing model relies exclusively on the finished output, it may fail to capture value associated with the earlier stages.

This is why the archetype pool deserves serious consideration.

We cannot build an equitable licensing structure around the assumption that everything of economic relevance will remain identifiable in the final product.

From technological value to commercial value

Having established why intermediate generations matter, we need to turn our attention to the money.

How should rights holders participate in the revenues generated by these systems?

Some of the largest companies operating in generative AI have achieved valuations in the billions of dollars, not collectively, but individually.

Their business models may involve training on copyright-protected works, developing systems capable of generating new content, and enabling users to distribute or commercially exploit that content.

The legal questions surrounding those activities remain ambiguous, and the underlying rights and permissions vary across jurisdictions.

But from a licensing perspective, the commercial opportunity is unquestionable.

These platforms are not simply generating individual songs. They are building entire businesses around the ability to generate content at scale.

And that sounds rather familiar.

Haven't we crossed this bridge before?

Think back to Musical.ly and the emergence of social media platforms.

These services introduced new ways for users to interact with music, create content and distribute it to enormous audiences.

Copyright-protected works became central to the user experience, while questions surrounding synchronisation, licensing, monetisation and the allocation of value had to be addressed.

The music industry has been here before.

We reached that bridge, negotiated our way across it, and eventually developed licensing structures to accommodate new forms of exploitation.

Admittedly, we crossed it on rather wobbly legs.

Today, we find ourselves standing at yet another bridge.

Except this time, it feels like the longest suspension bridge in the world, with a seemingly endless free fall below.

Everyone is racing to reach the other side. Technology companies are moving at extraordinary speed, investors are eager to see returns, and rights holders are trying to establish where their rights and economic interests fit into the equation.

How should the prompt pool be monetised?

There is no single answer, nor should we assume that one commercial structure will suit every generative AI service.

There are several mechanisms worth exploring:

●         Minimum fees per prompt or generation, ensuring that value is recognised regardless of whether a song is ultimately leaves the platform.

●         Revenue-sharing arrangements, allowing rights holders to participate in the commercial success of generative AI platforms.

●         Greater-of structures, combining minimum guaranteed payments with a percentage of revenues.

●         Proxy-based calculations, using measurable indicators such as users, subscriptions, generation volumes or consumption of computational resources where direct attribution is unavailable or unreliable.

A particularly interesting approach would be to establish a defined value allocation between the archetype pool and the output pool.

For example, a platform could contribute a negotiated percentage of relevant revenues to compensate for the generative process, while a separate allocation addresses the outputs.

The precise allocation would depend on the service, the underlying rights, the available data and the negotiated licensing terms.

The point here is to ensure that the economic value created during the generation process does not disappear simply because an intermediate song doesn't reach its final form and then leaves the platform.

And critically, any such framework must recognise that training, generation and downstream exploitation can represent distinct uses and sources of value.

Establish the value collectively. License it individually.

The technology is advancing faster than the legal and commercial frameworks intended to govern it.

While policymakers, courts and industry participants continue debating the boundaries of copyright protection and AI exploitation, generative platforms are expanding their capabilities and commercial reach. For every second that passes, an AI model can analyse and create new language for generation on 44,000+ songs, this increases the urgency of establishing workable principles.

But speed should not come at the expense of understanding.

Before entering negotiations, we need to establish what is being exploited, where value is created, and how that value should be measured.

Training, prompting and outputs should not be treated as one indistinguishable activity.

Nor should we assume that a licensing agreement covering one necessarily addresses the economic value of the others.

At Rights Intelligence, we support a collective effort to establish the principles and value chain for generative AI licensing, followed by individual implementation through commercial agreements appropriate to each platform, repertoire and use case.

This requires collaboration between rights holders, publishers, collecting societies, technology companies, researchers and policymakers.

We need transparency into generation processes, commercially meaningful data, and licensing models that recognise value even where precise work-by-work attribution is technically unavailable.

The industry has navigated technological disruption before. The question is whether this time we can establish the value before the market decides it for us.