AI Copyright Is an Infrastructure Problem.... Not a Policy One

Joe Wee · 2026-05-27

£8 billion and 220,000 jobs

The UK music industry alone generates around £8 billion in GVA and supports roughly 220,000 jobs [1]. Ministers call the creative sector a priority. And yet, over the last 12 months, the debate about AI and copyright slipped into a false choice: let technology hoover up creative work for free, or miss out on innovation.

Lord Ed Vaizey, former UK Minister for Culture and the Digital Economy, put it plainly in City AM: "Copyright is not a brake on innovation. It is the market mechanism that turns British creativity into investment and the route to responsible AI" [2].

He's seen this before. In the 2010s, the same chorus said copyright was outdated and infringement was the price of modernity. It wasn't true then. It isn't true now. What happened instead: commercial licensing gave us Spotify, Amazon Prime, and Netflix.

I've sat in the licensing chair

I ran a mobile games publishing business. We published Angry Birds, Cut the Rope, and hundreds of indie titles across 600+ creative studios. I negotiated the first mobile Superman game with DC Comics. Ice Age and Predator with 20th Century Fox. Stan Lee's Verticus and John Woo's Bloodstroke via CAA. He-Man: The Most Powerful Game in the Universe with Mattel. Every one of those deals came down to the same four questions: who owns the IP, what are the usage terms, how do you enforce them, and how do you prove compliance.

The enforcement part was always manual. Lawyers reviewed builds. QA checked assets against licence scope. Somebody kept a spreadsheet of what was approved and what wasn't. It worked because humans were in the loop and the scale was manageable.

AI agents don't work that way. They access content at machine speed, at scale, without asking. The four questions are the same. The enforcement infrastructure doesn't exist.

I'm also an indie musician. I built BandSaaS for independent artists.... same problem from the other side. Artists creating original work, worried about AI scraping it, with no way to track what's being used or enforce their rights.

When Vaizey says "commercial licensing at speed and at scale".... that's not abstract to me. I've done the deals. I've been the rights holder. The missing piece was always enforcement infrastructure that works at runtime, not after the fact. That's what A2A is.

The licensing market already exists. It's tiny.

A new WPI Economics / BPI report, "Driving UK Growth: the role of licensing music in the age of AI," found that nearly 80% of independent record companies see licensing music for AI as key to future growth [1]. But less than 20% are actively licensing. That gap is the opportunity.

The report profiles companies already building licensed AI tools: Human Native, Klay Vision, Voice Swap, Sound Patrol.... all fully and ethically licensed by major copyright holders [1]. These are not theoretical. They are shipping products with licensed content.

But they represent the beginning, not the norm. Most AI agents accessing creative content today have no licence check, no audit trail, and no provenance tagging. The infrastructure to enforce licensing at scale does not exist for most companies.

Two golden keys

Vaizey identifies exactly what's needed to accelerate the licensing market [2]:

  1. Legal certainty.... standing firmly behind UK copyright law
  2. Meaningful transparency.... about what is used in AI training

"These are golden keys to accelerating dialogue and deal-making between creative and tech companies. Get those foundations right and Britain can lead this global transformation, not by picking a side, but by making the market work for all" [2].

Legal certainty is a government job. Meaningful transparency is an infrastructure job. That's where we come in.

What "meaningful transparency" actually requires

When rights holders ask "what did your AI use?" they need more than a policy document. They need a tamper-proof record. Every time an AI agent touches a piece of content, the system should log:

Everyone reads the same file. No ambiguity. No "we think we might have used it."

The New York Times lawsuit against OpenAI [3] turned on exactly this question: what did the model consume, and can you prove it? The Getty Images case against Stability AI [4] asked the same thing for images. In both cases, the defendants had no granular audit trail. The arguments became "we trained on the whole internet" versus "you trained on our copyrighted work." Nobody could produce a record of what was accessed and when.

Three things rights holders actually need

1. Proof of access. A permanent, tamper-proof log of every piece of content your AI agent touched. Not "we have a content policy" but "here are the 14,000 content interactions from last month, with timestamps and licence status for each one."

2. Prevention of reproduction. A safety gate between the agent and copyrighted content. Known works get blocked instantly. An independent AI judge catches the subtler cases.... paraphrasing, style copying, near-verbatim reproduction that simple pattern matching can't catch.

3. Licence-check-before-use. Not an honour system. A real checkpoint. The agent asks for permission before using licensed content. The system checks whether the licence is valid. Allowed or blocked. Logged either way. The music industry has been doing this with tracking codes and collection societies for decades [5]. AI agents need the same infrastructure.

The EU already legislated this

EU AI Act Article 50 mandates that AI-generated content must be identifiable as AI-generated [6]. That means labelling. Watermarking. If your agents produce content and there's no way to tell it came from AI.... you're already non-compliant as of the December 2026 deadline.

The US has the COPIED Act proposal [7]. The UK consultation is heading the same direction [8]. And in April 2026, the Global Creative Economy Council (a joint initiative of Creative PEC and the British Council, representing five continents) published a Global AI Agenda for Cultural and Creative Industries [9]. Action 4 calls explicitly for "clear authorisation and licensing mechanisms for AI training," "support for IP marketplaces that facilitate lawful access to content," and "transparency requirements to ensure the use of creative content for AI purposes is properly disclosed."

This isn't one country or one industry lobbying. It's a global consensus: AI companies will need to prove what their systems consumed and produced. The infrastructure to enforce it doesn't exist for most companies. Yet.

So how does this actually work?

Before any AI agent accesses or uses content, it checks in with A2A Infrastructure. Think of it like a security desk at the front of a building. The agent says "I want to summarise chapter 3 of this book." A2A checks four things in sequence:

  1. Is this a known protected work? Pattern matching catches recognised content instantly.... books, tracks, articles. If it's on the blocked list, the agent is stopped before it touches anything.
  2. Does the output look like reproduction? An independent AI safety judge reviews whether the agent is copying, paraphrasing too closely, or reproducing the original. The judge has no memory of previous conversations.... it can't be talked into making exceptions.
  3. Is there a pattern of scraping? The system watches for agents that access lots of copyrighted content in sequence.... the kind of behaviour that looks like systematic harvesting rather than legitimate use.
  4. Is this agent allowed to access this type of content? Each agent has a defined scope. A summariser can read. A content generator might be blocked from accessing certain catalogues entirely.

Every interaction is logged with a tamper-proof digital fingerprint. The full record is exportable as a single file.... hand it to a rights holder, a regulator, or a court. That's the "meaningful transparency" Vaizey is asking for.

The Spotify analogy is exact

Vaizey draws the parallel explicitly: in the 2010s, commercial licensing of copyright gave us Spotify, Amazon Prime, and Netflix [2]. The same pattern is starting in AI. Human Native, Klay Vision, Voice Swap, Sound Patrol.... these are the Spotifys of licensed AI tools [1].

But Spotify needed infrastructure. It needed licensing systems, royalty tracking, usage metering, catalogue lookups. The music industry built that infrastructure over a decade.

AI needs the same infrastructure. 80% of independent labels want to licence for AI. Less than 20% are doing it. The gap isn't willingness. It's plumbing.

The good news? We don't have to wait.

A2A Infrastructure already does this. An agent wants to use licensed content. A2A checks the licence. Valid? The agent proceeds. Expired? Blocked. Usage cap reached? Blocked. Every decision is logged. The rights holder gets a monthly report showing exactly how their content was used, by which agents, how many times.

The music industry waited a decade for Spotify's plumbing. AI doesn't have to. The plumbing exists now. That's how you close the 80/20 gap.

Let me show you, not tell you

I said BandSaaS was built for independent artists worried about AI scraping their work. It now runs on top of A2A. Every track an artist registers on BandSaaS becomes a protected work with a licence gate in front of it. This isn't a slide.... here it is, running on real repertoire.

My own band, Crosswinds, registered our catalogue. Take the track “All Good Things” (ISRC USCGH1915861).... a real recording, co-written and co-owned by its two writers, with the royalty split recorded on the work itself. An AI company holds a summarise-only licence to it. Watch what the licence scope actually enforces:

Both attempts.... the one that was allowed and the one that was stopped.... are written to an immutable access log the rights holder can see and export. Nobody has to take anyone's word for it. Everyone reads the same file.

That's the whole argument, made concrete. A registered work. A licence with a defined scope. In-scope use allowed, out-of-scope use blocked, and a tamper-proof record of both.... the exact "meaningful transparency" Vaizey is asking for, working today, on a track I helped write. The infrastructure the New York Times and Getty couldn't produce in court is a licence key and an access log.

Want to know how A2A can help your business?

Tell us about your content and your AI challenge. We'll show you how we can help.

Find out how we can help Creative Industries Use Case
References
  1. WPI Economics / BPI (2026). "Driving UK Growth: the role of licensing music in the age of AI." Profiles Human Native, Klay Vision, Voice Swap, Sound Patrol. 80% of independents see AI licensing as key to growth. BPI
  2. Vaizey, Lord Ed (2026). "Copyright isn't dead in the age of AI, it's key to growing UK creative industries." City AM
  3. The New York Times Company v. Microsoft Corporation et al. (S.D.N.Y. 2023). Copyright infringement complaint against OpenAI. NYT
  4. Getty Images v. Stability AI (D. Del. 2023). Copyright infringement for training on 12M+ images. Copyright Lately
  5. IFPI (2025). "Global Music Report 2025." Revenue and licensing infrastructure. IFPI
  6. EU AI Act, Regulation (EU) 2024/1689, Article 50: Transparency obligations for AI-generated content. artificialintelligenceact.eu
  7. US COPIED Act proposal (2024). Content Origin Protection and Integrity from Edited and Deepfaked media. Congress.gov
  8. UK Government (2023-2026). AI and Copyright Consultation. GOV.UK
  9. Creative PEC / British Council / GCEC (2026). "A Global AI Agenda for the Cultural and Creative Industries." 11 key actions including licensing mechanisms, transparency requirements, and IP marketplace infrastructure. PEC
  10. Baysal, H. (2026). "Asimov Safety Architecture." IETF Internet-Draft
Help

Help

Need help? Here are some quick links: