Content Provenance Is a Product Surface, Not a Label
2026-06-30 · 5 min read · Janaina Maia
TIDAL just announced that fully AI-generated music on its platform will no longer be monetisable. Tracks identified as 100% AI will get a visible AI badge, lose royalty eligibility, and be excluded from direct-to-fan sales. Impersonation of real artists will be removed entirely using automated detection.
TechCrunch reports that TIDAL's policy joins Spotify, Apple Music, and Deezer in addressing AI-generated content. Deezer, which says 44% of songs uploaded daily are AI-generated, has gone furthest: it removes AI tracks from recommendations and editorial playlists entirely, offers its detection tech to rivals, and just released a consumer tool that scans playlists on competing services for AI content.
TIDAL's EVP Tony Gervino framed it as protecting organic creativity. He also said the policy is a living document, open to change. That framing, a clear stance with built-in room to evolve, is more honest than most platforms have been.
But the product challenge is not whether to label AI content. It is what happens after you label it.
Labels are the easy part. Provenance is the hard part.
An AI badge on a track tells you that a machine generated it. It does not tell you what kind of machine, what data it was trained on, whether a human shaped the output, whether the person whose voice is imitated consented, or whether the track was composed entirely by an AI or composed by a human and arranged with AI assistance.
The badge is binary: AI or not AI. Reality is a spectrum. Some tracks are fully AI-generated. Some are human-written and AI-produced. Some use AI for one element, a drum pattern, a background vocal, a mix decision. The badge collapses all of that into a single tag, and the tag does not carry enough information for a listener, an artist, or a rights holder to make a good decision.
What people actually need is provenance: where did this come from, what tools were used, what human involvement existed, and what rights attach to each element.
What is content provenance?
Content provenance is the full history of how a piece of content came to be. In the physical world, we have strong norms around this. Wine labels show region, grape, and vintage. Art provenance traces ownership back to the artist. Restaurant menus name the farm. None of these systems are perfect, but they carry meaning that a binary label does not.
In digital content, provenance means knowing whether a track was composed by a person or a model, whether a vocal was sung or synthesised, whether a mix was curated by an engineer or optimised by an algorithm, and what training data contributed to the tools that produced each element.
Some of this is trackable now. Adobe's Content Credentials initiative embeds cryptographic provenance metadata in images. The C2PA standard, which Adobe helped create and which is now supported by Microsoft, Getty, and the BBC, allows a chain of authorship and modification history to travel with a file. Music does not yet have an equivalent, but the need is the same.
Why this matters beyond music.
Every platform that hosts AI-generated content faces the same structural question. It is not a music question. It is a trust question.
When GitHub gets flooded with AI-generated pull requests, the problem is not that the code is AI-written. The problem is that the provenance signal is missing, so reviewers cannot triage by confidence. When search results mix AI-generated articles with reported journalism, the problem is not the AI article itself. It is that the reader cannot tell which is which. When a customer service chatbot gives advice, the user needs to know whether a human verified it.
In every case, the product question is the same: what does the user need to know to make a good decision, and how does the interface surface that information without overwhelming them?
What product teams should build.
- Provenance, not just labels. A badge that says AI tells you almost nothing. A provenance trail that says composed by a human, produced with AI mixing tools, vocal performed by the credited artist tells you something useful. Design the surface to carry the nuance, not flatten it.
- Graduated trust, not binary gates. Demonetising fully AI content is a reasonable starting position. But the product should be able to handle the middle ground: content that is partially AI, AI-assisted, or AI-enhanced. A binary system will either over-penalise legitimate creative tool use or under-penalise fully synthetic content. Build for the gradient.
- Detection as infrastructure, not announcement. TIDAL says it will use automated detection to remove impersonation. Deezer is building detection tools as platform infrastructure. The right move is to invest in detection as a continuous, improving system, not a one-time policy announcement. Models for detecting AI content will get better. So will models for generating it. The detection layer needs to be a living system, not a flag.
- Artist consent as a first-class product concern. The deepest trust issue in AI music is not whether a track sounds like a real artist. It is whether a real artist agreed to have their style, voice, or likeness used. Consent management is a product surface: artists should be able to declare what AI use they permit, and that declaration should travel with their work.
- Living policies, not stone tablets. TIDAL's framing of its policy as a living document is the right instinct. Any platform building AI content governance should design the policy as a versioned, evolving system, not a permanent rule carved in stone.
My take.
Deezer's 44% figure is a warning shot. When nearly half of new uploads to a music platform are machine-generated, the volume problem is real. But the deeper problem is not volume. It is that most platforms are trying to solve a provenance question with a labelling answer.
Labels tell you what something is. Provenance tells you where it came from, who touched it, and what rights attach to it. Labels are cheap to add and easy to ignore. Provenance requires investment in infrastructure, standards, and product surfaces that carry meaning.
TIDAL is right to take a stance. Deezer is right to build detection as infrastructure. But both are still working with the word AI as if it is a single, knowable category. It is not. The next generation of content governance products will not sort content into AI and not-AI. They will trace provenance, manage consent, and give people the information they need to decide what they trust and what they do not.
The badge is the start of the conversation, not the end of the product design.