Is Your Music in AI Training Datasets? How to Check AI Watchdog & Document the Evidence
Share
Creator Rights · Dataset Evidence Workflow · Deep-research update September 5, 2026
Is Your Music in AI Training Datasets? How to Check AI Watchdog & Document the Evidence
The Atlantic's AI Watchdog gives musicians a rare way to search several large public music datasets for their own artist names and songs. The useful question is not immediately, “Did Company X train Model Y on my recording?” It is first: what exactly is the record I found, what kind of dataset is it in, and what can that evidence honestly establish?
The rule that governs this page
A dataset match is evidence of dataset inclusion. It is not, by itself, proof that a named company used your specific work to train a named commercial model.
The reverse matters too: a non-match only tells you that you did not find the work in the datasets and searches you checked. It does not prove the work was absent from every private, historic or undisclosed training corpus.
What AI Watchdog is actually searching
The four datasets should not be treated as interchangeable. They contain different things and were built for different purposes. A row of Spotify metadata is not the same kind of evidence as a downloadable audio research corpus; a YouTube link is not the same thing as the dataset publisher distributing the audio itself.
| Dataset | What it contains | Why creators should care | What not to assume |
|---|---|---|---|
| LAION-DISCO-12M | Metadata for 12,648,485 songs paired with links to publicly available YouTube samples. LAION says the dataset itself contains links and metadata, not the original audio files. | It is explicitly positioned for machine-learning research including audio foundation models, music information retrieval, generation and related tasks. | A match does not mean LAION distributed your master recording, or that every model developer downloaded or trained on the linked audio. |
| SLEEPING-DISCO 9M | A roughly 9.7-million-row music/song dataset presented by its authors as a large-scale pre-training dataset for generative music modeling. | Its stated purpose makes a match particularly relevant when documenting the public research ecosystem around generative-music training. | Its purpose still does not establish that a particular commercial company or deployed model used your specific row or recording. |
| Free Music Archive (FMA) | 106,574 Creative Commons-licensed tracks with audio, metadata, features and genre information; the full research collection includes downloadable audio. | This is materially different from a link-only or metadata-only collection because actual audio was assembled for music-analysis research. | “Creative Commons” is not one universal permission. Licence terms differ, and dataset inclusion alone does not answer every later use or commercial-training question. |
| Spotify Tracks Dataset | A tabular CSV-style collection of Spotify track IDs, artist/title/album fields and audio-feature metadata across 125 genres. | It can show that a track's metadata appears in a widely reused data collection. | Do not describe a metadata-row match as proof that the dataset contains a copy of the Spotify audio file. |
Audio, links and metadata are different evidence
Metadata
Artist, title, album, track ID, duration, genre or audio-feature fields can identify a work without containing the recording itself.
Links
A YouTube or other source URL can point researchers toward public media while leaving the original file hosted elsewhere.
Audio
A dataset that distributes audio presents a different technical and rights question from one that publishes only identifiers or links.
Documented training use
Evidence that a dataset was actually used in a training run is a further step beyond finding a work inside the dataset.
The JR evidence ladder: five levels of certainty
Use this ladder before deciding how strongly to describe what you found.
You found an artist/title record that appears to correspond to your work. Verify identity before going further.
The title, artist, identifier, source URL or other fields line up with your release records strongly enough to identify the work.
You can establish what the dataset record points to or contains: metadata, a URL, an audio sample, a full file or another asset.
A paper, repository, company statement, court filing or authenticated record connects a researcher or developer to use of that dataset.
Evidence connects your identified work, not merely the dataset generally, to the training process for the model or system at issue. This is a much stronger claim and often requires records unavailable to an ordinary public search.
Most creators using AI Watchdog will begin at Levels 1–3. That can still be valuable evidence. It simply should not be described as Level 5.
How to search your catalog properly
Do not search one spelling of your artist name, see nothing, and stop. Catalog metadata often contains duplicates, alternate credits, compilation listings and inconsistent punctuation.
Try punctuation, spacing and capitalization variants if your branding has changed.
Include producer names, featured-artist credits and older stage names that have appeared in distribution metadata.
This can surface a work even when the artist credit is inconsistent or the track appears through a compilation, collaboration or re-upload.
Where an ID, URL or platform track identifier is exposed, compare it with your distributor dashboard, ISRC records and public release pages. Do not rely only on title similarity.
Write down which of the four datasets produced the result and whether the record is metadata, a link or an audio-bearing collection.
The same recording can appear under multiple YouTube uploads, albums, playlists or metadata rows. Multiple rows are not automatically multiple independent uses.
If you find a match, preserve the evidence before you argue about it
The first useful action is documentation, not a social-media accusation.
- Take screenshots showing the result and enough surrounding context to identify the search tool and dataset.
- Record the exact artist name, track title, identifiers and source URL shown.
- Record the dataset name and, where available, row or record information.
- Record the date and time of the search. Public datasets, links and interfaces can change.
- Save the AI Watchdog result URL where the interface provides one.
- Compare the result with your master, distributor, ISRC, publishing and registration records.
- Save a short note identifying what the record actually contains: metadata, link, audio or another form.
- Store the result with your existing creator-rights and chain-of-custody records.
JR rule: document first, characterize second
“My song appears in LAION-DISCO-12M as a YouTube-linked record” is more precise and more defensible than “Company X stole my song to train Model Y” when you do not yet have evidence connecting that company, your particular work and that model's training run.
What a match does—and does not—tell you
| Finding | What you can reasonably say | What still needs proof |
|---|---|---|
| Your work appears in a named dataset | The identified work appears in that dataset in the form shown by the record. | Whether a particular developer accessed it, whether audio was obtained, whether a specific model trained on it and what legal consequence follows. |
| The record contains a public-media URL | The dataset points to that public source at the recorded URL. | Whether the linked media was downloaded, retained, processed or selected for a later training run. |
| The dataset distributes audio | The dataset includes audio material associated with the track under the dataset's documented structure. | Whether a downstream commercial use complied with the applicable licence and law. |
| Your work does not appear | You did not find it in the datasets and searches checked. | Whether it appears in a different dataset, private corpus, historical copy or unindexed record. |
| A developer is documented as using the dataset | There is evidence connecting that developer or research project to the dataset generally. | Whether your specific record entered the training subset for the particular model you care about. |
How this fits with Suno, Udio and the training-data lawsuits
Jack Righteous already has a separate cornerstone tracking the legal cases. That page owns questions such as alleged training copies, acquisition methods, derivative outputs, anti-circumvention claims, licensing settlements and jurisdiction. This AI Watchdog page owns the individual creator evidence workflow.
The distinction mirrors what courts and technical investigations increasingly have to separate: what was collected, what was retained, what entered training, what a model later generated and what a creator can prove about a specific work.
Read the Suno & Udio Copyright Lawsuits 2026 tracker →
For a different kind of evidence—the leaked Suno collection scripts and internal training-data references—read Suno Data Breach Explained: What the Leak Reveals →. That investigation uses the same discipline: collected data is not automatically retained data; retained data is not automatically training data; training data is not automatically memorized output.
Pre-release checking and post-release dataset checking are different jobs
JR's AI Music Copyright Checks Before Release asks whether your planned output creates identifiable rights problems before distribution. AI Watchdog reverses the direction: after music exists publicly, can you find evidence that your work appears in public datasets?
Before release
Check your sources, licences, uploaded audio, voice/identity risk, suspicious similarities, platform terms, distributor rules and human contribution.
After release
Keep registration and distribution records, monitor where the work travels, and preserve credible dataset evidence if your catalog later appears in a searchable collection.
Add the result to your creator proof record
Your rights folder should not contain only proof of how you made a song. It can also contain evidence about where the finished work later appeared.
For each important release, keep the master, project files, lyrics, contributors, distribution metadata, ISRC, registrations, release date, public URLs, platform terms, relevant licences and credible dataset-search results. A clean record is much more useful than trying to reconstruct the chain months or years later.
Continue with Music Registration & Royalty Collection in 2026 →.
If you find a match, choose the next step by your objective
| Your objective | Best next move |
|---|---|
| I'm simply checking my catalog | Save the result, classify the dataset record correctly, and stop there unless a stronger reason to investigate appears. |
| I want a clean rights record | Attach the result to your master/ISRC/registration file and record when and where it was found. |
| I am evaluating a lawsuit headline | Use the JR lawsuit tracker to distinguish platform acquisition, model training, outputs and creator conduct before connecting your result to the case. |
| I found a commercially important work in a dataset | Preserve the evidence in original form and consider sharing it with the publisher, label, collecting society or rights professional who represents the relevant rights. |
| I am considering a legal claim | Do not rely on this article or a search result as legal advice. A qualified copyright or entertainment lawyer can assess ownership, jurisdiction, evidence and available claims. |
Should you contact a lawyer if you find your music?
A dataset match is information, not an individualized legal conclusion. Whether further action makes sense depends on who controls the composition and master, the dataset, what form of the work appears, the evidence connecting that dataset to any downstream use, the jurisdiction and the value at stake.
If a work has meaningful commercial value or the evidence becomes relevant to an active dispute, preserve the original result before contacting a qualified copyright or entertainment lawyer. If you have a publisher, label or collecting society, ask what evidence they want you to retain rather than assuming a screenshot alone is sufficient.
The Jack Righteous position
Creators deserve better visibility into where their work travels. AI Watchdog does not settle the AI-training legal debate, but it does something useful: it moves ordinary creators from abstract speculation toward inspectable evidence.
Use it as an evidence check, not a verdict machine. Search carefully. Identify the dataset. Determine whether you found metadata, a link or audio. Preserve the record. Keep your language no stronger than the evidence. Then connect the result to the rest of your rights documentation.
Choose your next route
Do something useful with the result.
If you found your catalog, preserve it. If you are preparing a release, check the work itself. If you need the bigger legal story, use the lawsuit tracker. If you are building a durable catalog, organize the rights record.
Rights & Ownership Guide Pre-Release Copyright Checks Lawsuit TrackerPrimary sources and dataset documentation
- The Atlantic — AI Watchdog / music-dataset investigation
- The Verge — searchable AI-music training database coverage
- LAION — LAION-DISCO-12M documentation
- Ahmed, Radonjic & Rabby — SLEEPING-DISCO 9M paper
- FMA — dataset documentation and research repository
- Spotify Tracks Dataset — dataset card
Editorial note: Deep-research review completed September 5, 2026. Dataset availability, public records and legal claims can change. JackRighteous.com has not independently audited every record in these datasets. This page provides creator education and evidence-preservation guidance, not legal advice.
