Detecting AI music after the most important event has already happened.

On August 6, Suno published How We're Building the Future of Music Responsibly, outlining new principles on artists, creativity, transparency, and generative AI.

A new download policy aimed at making mass distribution of generated songs harder. Transparency tools so Suno-generated music can be identified outside the platform.
Audio watermarking and fingerprinting.

Useful measures, and a line drawn at the same time. A watermark that applies only going forward marks the exact point where one company's accountability starts. Everything behind that line remains exactly as unlabelled and unprovable as it was yesterday, because closing that gap was never part of the problem anyone required them to solve.

Starting at the wrong end

Watermarking can tell a DSP that a recording came from an AI system. Fingerprinting can help identify generated material. Detection can tell us that something on Spotify or Deezer was probably generated by AI. All of it operates after the model has been built and the music has been generated.
It tells us what came out.
It tells us almost nothing about what went in.

Bolting the barn door after the horse has bolted, then advertising the lock.

The training happened first.

Suno says it chose not to use artist names in its training metadata, so users would create original music rather than imitate a particular artist. A reasonable product safeguard. Not the same thing as establishing that an artist's music was never present in the training material.

Sophisticated detection is arriving after enormous generative systems have already learned from human music. Deezer reported roughly 90,000 fully AI-generated tracks arriving daily by July, exceeding half of new daily uploads at points in June 2026. Detecting those tracks says nothing about where the musical knowledge behind them came from.

The liars' dividend

The dividend isn't collected by whoever fakes something. It's collected by everyone else, once convincing fakes exist at scale, because now every genuine thing can also be waved away as maybe fake.

Watermarking outputs in the future doesn't close that gap. It ratifies it for everything that came before. A platform looks responsible about the music it makes from today onward, while the music it already learned from stays unverifiable, undated, and someone else's problem to prove.
Every artist whose work predates this year's watermarking policy has just had their claim to it demoted to a matter of opinion.

Provenance has to start before AI use.

ProofProfile is not a tool for looking at a track afterward and deciding whether a machine made it. It establishes a structured, persistent, independently verifiable record of the human artist and the recording before that question becomes necessary.

An independently verifiable timestamp establishes that specific information existed before a specific point in time. It cannot reach backward and prove that Suno, Udio, or any other model trained on a given recording. It establishes something narrower: the human record existed first.

Detection tells us what happened later.
Provenance tells us what existed earlier.

A watermark added to a recording in 2026 identifies that output. It cannot reconstruct the provenance of the millions of human recordings that were sitting online during the formative years of generative music. That window does not reopen. A company can ship better watermarking next quarter. Nobody can manufacture an earlier, independently verifiable point in time after the fact.

Suno's principles are a real step, and a measurement of how late the industry started asking the question. The companies training the models were never going to be the ones building the record that predates them.
That isn't their liability to close.

Knowing that something was generated by AI is one problem. Having a reliable, machine-readable record of what existed before AI generated it is a different, earlier, and so far unaddressed one.

That is the problem ProofProfile is built to answer.