How to Verify Music Authenticity Before Licensing

When you license a track, you are paying for rights the seller may not actually hold. Purely machine-generated music has an unsettled copyright status in most jurisdictions, and the EU AI Act makes labeling of AI-generated content mandatory in 2026. A verification pass that takes minutes protects the sync fee, the placement, and the entire indemnification chain behind it.

Provenance is now a contract problem, not a taste problem

Whether an AI track sounds good is irrelevant to the licensing question. What matters is whether the licensor can grant the rights the agreement describes. Four exposure points come up in practice:

• Copyrightability. Works generated entirely by a machine, with no meaningful human authorship, generally cannot be protected by copyright in the United States. If the underlying composition is unprotectable, the exclusivity you are paying for may be unenforceable against anyone.
• Warranties and indemnities. Standard license agreements have the licensor warrant that the work is original and does not infringe third-party rights. Generative models trained on unlicensed catalogs put a question mark over that warranty, and you inherit the dispute if a claim lands after the campaign airs.
• Disclosure obligations. The EU AI Act requires AI-generated content to be labeled starting in 2026. A brand campaign running in Europe with an undisclosed AI track becomes a compliance issue on top of a rights issue.
• Platform policy. Streaming services are building their own screening. Deezer, for example, runs an in-house AI-music detection tool. A track flagged after release can be demoted or removed, taking your placement's value with it.

The five-step pre-licensing check

Run this sequence on every track before signature, whether it is a one-off sync deal or a catalog buy:

1. Ask for provenance in writing. Request the DAW project file, stems, and session dates. A licensor who composed the track can produce these in minutes. Evasiveness here is itself a data point.
2. Check embedded credentials. Some generators embed provenance metadata. AI Song Checker reads both C2PA credentials and SynthID watermarks, so a labeled file is caught immediately, before any acoustic analysis runs.
3. Run a forensic scan on the delivered master. Watermarks can be stripped by re-encoding; acoustic fingerprints cannot. This is the step that catches laundered files (details below).
4. Scan the seller's adjacent catalog. One clean file proves little. If the rest of the licensor's recent output scores as Suno or Udio, treat the track you are buying with the same suspicion.
5. Paper the result. Add an explicit representation that no generative-AI system produced the composition or master, tied to indemnification. If the licensor balks at signing that sentence, you have your answer.

What a forensic scan actually measures

Human ears stop being reliable around the current generation of models, a problem covered in detail in why modern AI beats are hard to detect. Signal analysis does not have that ceiling. The ASC v8.3 engine evaluates 82+ forensic signals and combines them with Bayesian inference scoring; on a holdout set of 50,000+ tracks it reaches 99.1% accuracy with a 0.4% false-positive rate. The signals that matter most in licensing disputes:

• Frequency cutoff at 16 kHz where a studio master keeps energy above it: a common trace of generation pipelines rather than recording chains.
• Neural codec artifacts in the 5-8 kHz band, characteristic of EnCodec-based models such as Meta's MusicGen.
• Spectral flatness and phase coherence entropy that sit outside the ranges produced by real rooms, real microphones, and real mixing decisions.
• Micro-timing, tempo drift, and attack variance. Human performances wobble in structured ways; generated ones are either too perfect or wrong in statistically detectable patterns.
• Checkerboard deconvolution artifacts and resampling notches left behind when a file is re-rendered to hide its origin.

Match verification depth to deal size

Not every deal justifies the same diligence. Scale the check to the money and the reuse risk:

Scenario	Risk profile	Recommended check
One-off sync placement	Moderate: single work, single campaign	Single-track scan plus C2PA/SynthID check
Sample pack or beat lease	High: one flagged loop contaminates every derivative work	Scan every stem and file, keep the reports
Library / production music intake	High volume, repeated exposure	Automated screening via the REST API (free tier: 100 requests/day)
Catalog acquisition or label signing	Highest: valuation depends on rights being real	Full-catalog batch analysis with CSV export and per-track PDF certificates (Pro)

Generator-specific checks worth running

Most AI material circulating in commercial pitches today comes from a handful of tools, and each leaves a distinct signature. If a track pattern-matches a specific generator, confirm with the dedicated model: the Suno detector covers v3.5 through v5 output, and the Udio detector handles v1.0 and v1.5. Riffusion, ElevenLabs Music, Stable Audio, and Mureka are covered as well. Note that newer model versions consistently fool listeners who could spot the previous generation by ear, which is exactly why the check needs to happen at the signal level, not in a listening session.

A note on unreleased masters

Licensing diligence usually involves files under NDA, so where the audio travels matters. AI Song Checker extracts audio features directly in your browser via the Web Audio API; only the numerical features are sent to the server, and no audio is stored without an account. You can verify a confidential pre-release master without uploading the recording itself anywhere. You get 3 analyses per day without an account, unlimited with a free email signup, and URL-based analysis works for anything already public on YouTube, Spotify, or SoundCloud.

Where to go from here

Make verification a standing line item in your licensing checklist: provenance documents, credential check, forensic scan, catalog spot-check, contract language. It adds minutes to a deal and removes the single largest new rights risk in music licensing. If you want to sharpen your own ear as a first filter, start with the five audible signs of an AI-generated song, then let the signal analysis do the part ears cannot.