16 November 2026 · LicenseTrack

How Independent Composers Can Verify AI Training Use of Their Work: Auditing, Disclosure, and Remedies

A practical evidence trail for composers checking possible AI-training use: organise rights records, review public disclosures, ask what providers and societies can confirm, and understand the limits of identifiers and legal routes.

By LicenseTrack

16 November 2026Last updated: 16 November 20269 min read

Question: What verification and remedy mechanisms can independent composers use to audit whether their work has been used for AI training without consent, and what disclosure or enforcement steps are currently available?

There is no identifier or catalogue search that, on its own, confirms whether a particular recording or composition was used to train a model. A careful review can create a stronger record and help you ask focused questions, but a provider’s training records may not be public and a lead is not proof. This guide uses U.S. examples where noted. Rights, procedures and deadlines vary by jurisdiction. The sources and legislative status below were checked on 16 November 2026.

1. Make a rights file for each work

Start with the thing you may need to identify and the rights you actually control. Keep the composition and each recording as separate records. For the composition, collect titles and alternate titles, writers, split sheets, publisher or administrator details, agreements, amendments, assignments and registration references. For each recording, keep the master owner or licensee information, release and version details, distributor or label statements, and the agreement that explains your authority to act.

Keep dated drafts, session files, stems, exports, lyric or score versions, emails and delivery records where you already have them. Record who created or supplied each item and when. If there are co-writers, a publisher, label or other rights holder, confirm who can make an inquiry or take a step for the relevant right; do not assume that control of a recording also means control of the composition, or vice versa.

2. Use identifiers to make your catalogue easier to describe

Use an ISRC for a sound recording when one has been assigned, and an ISWC for the underlying musical work when one has been assigned. The International ISRC Registration Authority says ISRC identifies sound recordings and music videos, not compositions; CISAC says ISWC identifies musical works by reference to their title and creators, not particular recordings or instances of use. An ISRC does not itself establish ownership. These codes can help a provider or society locate the right catalogue entry, but neither is an AI-training disclosure flag or evidence that a model ingested the work. A missing search result also does not prove that the work was absent from a dataset. (IFPI, ISRC FAQ, checked 16 November 2026; CISAC, International Identifiers, checked 16 November 2026.)

Keep the identifier beside the exact title, version, duration, contributors, rights notes and source record it describes. When you search a registry, save the result and the date. A title or creator-name match can be ambiguous; compare the underlying work or recording and its version before relying on it.

3. Review public disclosures and terms, one model at a time

Identify the provider, model or service, and version you are asking about. Review any public dataset list, training-data statement, model card, transparency report, opt-out page, rights policy, terms of service or FAQ that applies to that specific model and date. Save a copy or dated link to the exact text and note its scope: a provider may describe a dataset category without listing every item, or explain a policy without confirming whether a particular work was included.

Read the terms that applied when the relevant material may have been collected or used, as well as the current terms. Do not infer that a public-facing policy guarantees disclosure, removal, payment or a response. If no work-level source information is published, record that as unknown rather than treating silence as confirmation either way.

4. Ask what the provider or rights society can actually confirm

Use the provider’s published contact or rights channel, if it offers one. Make the request specific: identify the work and recording, give the ISWC or ISRC where available, describe the rights you control, name the model and version, and link to the provider’s relevant policy. Ask what records the provider can check, whether it can confirm any source or licence information for that work, and which process applies to a correction, objection or other rights request. Keep the request, date, ticket number and complete reply. A response only establishes what that provider says it checked; a non-response or a negative search may reflect incomplete records and is not proof of absence.

If a performing-rights organisation or collective management organisation represents the relevant work or right, ask what its repertoire records, member guidance and available process can confirm. Ask specifically whether it has access to training-data records; do not assume that registration or membership gives it access to a model’s dataset. Society rules and mandates differ, so apply only the policy of the organisation you contact.

5. Preserve the evidence and label its limits

Keep a dated chronology with source URLs, saved policy versions, registry results, correspondence, agreements and the catalogue records you used. Preserve original files and their provenance. If a specific generated output or released recording appears similar to your work, keep the exact output, its URL and date, and document the passages or musical features you are comparing. A qualified audio examiner may be able to describe a similarity between two audio files, but similarity analysis does not reveal a training record and does not by itself prove that a work was used in training. Common musical material, independent creation, an intermediate source, or other explanations may need to be considered.

What ASCAP says—and what that does not decide

On 28 October 2025, ASCAP announced with BMI and SOCAN that their policies accept registrations of musical compositions that combine AI-generated material with human authorship; wholly AI-generated compositions are ineligible under those societies’ stated policies. That is a performing-rights-society registration policy, not a universal rule for every society and not a decision about copyright ownership in an individual work. The same announcement states ASCAP’s position, shared with BMI and SOCAN, that AI companies ingesting and training on copyrighted musical works without permission, compensation or credit is “not fair use, but theft.” This is an attributed advocacy position, not a court finding or a settled legal conclusion for every use. (ASCAP, AI registration policy announcement, 28 October 2025; the joint announcement published by BMI, 28 October 2025.)

Current U.S. routes depend on a specific claim

If the records suggest a concrete dispute, a U.S. copyright lawyer can assess ownership, the right involved, evidence, any permission or statutory limitation, whether an existing agreement provides a contractual route, and the correct forum before you make a legal demand. The U.S. Copyright Office explains that registration or refusal is generally required before a federal infringement suit concerning a U.S. work, and that registration timing affects eligibility for statutory damages and attorneys’ fees. It also describes the Copyright Claims Board as a voluntary forum for eligible copyright disputes seeking less than $30,000; a claimant must have filed an application to register the work. These are procedural routes, not a finding that training occurred or that any particular claim will succeed. (U.S. Copyright Office, Circular 1, Copyright Basics, and What Writers Should Know About Copyright, checked 16 November 2026; see also 17 U.S.C. §§ 411 and 412.)

If a specific allegedly infringing copy is available from an online service, the U.S. Copyright Office’s Section 512 guidance describes a notice process for asking the service to remove identified material; a compliant notice includes statements about authority, good faith and accuracy under penalty of perjury. That process targets material located on the service. It does not compel disclosure of a model’s training records or prove that training occurred. (U.S. Copyright Office, Section 512 notice-and-takedown resources, checked 16 November 2026.)

The TRAIN Act is a proposal, not a current disclosure remedy. Congress.gov lists S. 2455 in the 119th Congress as “Introduced”; its latest recorded action is 24 July 2025, when it was read twice and referred to the Senate Judiciary Committee. Senator Welch’s 25 July 2025 release describes the bill as a proposal for an administrative subpoena process to help copyright owners determine which works were used to train AI models. As of 16 November 2026, do not present that proposed process as an enacted right or available procedure. (Congress.gov, S. 2455 — TRAIN Act; Senator Peter Welch, TRAIN Act announcement, 25 July 2025.)

A practical next step

Build the rights file first, then write down what is known, what is only a lead, and what remains unknown. Check the relevant model’s public materials and terms; ask the provider and any relevant society narrowly framed questions; preserve the dated replies. If you have evidence of a specific use and are considering a demand, notice or claim, take the records to qualified counsel in the relevant jurisdiction before acting.

General educational information only—not legal advice for a particular work or dispute. Nothing here determines that a work was used for training, that a particular use was authorised or unlawful, or that disclosure, removal, payment or enforcement will follow.

FAQ

Can an ISRC or ISWC show that a model trained on my music?

No. An ISRC identifies a sound recording and an ISWC identifies a musical work. Either may help describe a catalogue entry in an inquiry, but neither is a training-use record or evidence that a model ingested the work. (IFPI’s ISRC FAQ; CISAC’s international identifiers, checked 16 November 2026.)

Does a similar generated audio passage prove that my work was in training data?

No. Similarity or forensic analysis can describe features shared by two audio files, but it does not reveal a model’s training records or by itself establish that your work was used. Keep the exact material and date as a lead, and get qualified advice before treating it as proof.

Can the TRAIN Act require a provider to disclose training records today?

No. As checked on 16 November 2026, Congress.gov lists S. 2455 in the 119th Congress as introduced, with its latest recorded action on 24 July 2025 referring it to the Senate Judiciary Committee. It is proposed legislation, not an available disclosure procedure. (Congress.gov, S. 2455 — TRAIN Act.)

Want to stop playing whack-a-mole with stolen uploads? See subscription pricing →Monthly is $19.99 per month and annual is $66.65 per year. Both include a seven-day free trial. A payment card is required at signup; billing starts when the trial ends unless you cancel before then.

Related resources

Put this into practice — or read more from independent composers.

Explore LicenseTrack monitoring

Monthly is $19.99 per month and annual is $66.65 per year. Both include a seven-day free trial. A payment card is required at signup; billing starts when the trial ends unless you cancel before then.

For independent composers

Monitor where your tracks appear — and protect your licensing rights.

Find where your music appears, preserve match evidence for review, and move from a detection toward the right licensing conversation. Monthly is $19.99 per month and annual is $66.65 per year. Both include a seven-day free trial. A payment card is required at signup; billing starts when the trial ends unless you cancel before then.

See subscription pricing →

See it in action

Curious what LicenseTrack catches in the wild?

Review a clearly labelled sample report; it is not a customer result or proof of a live match.

View the sample detection report

Convert the read into a result

Review a possible use, then choose your next step.

See a sample detection report or review the plans for the live LicenseTrack product.

Two ways in — whichever fits your week.

Ready to register a track for scheduled scanning?

Monthly is $19.99 per month and annual is $66.65 per year. Both include a seven-day free trial. A payment card is required at signup; billing starts when the trial ends unless you cancel before then. Provider configuration determines which monitoring paths are available after onboarding.

See subscription pricing

Subscribe via RSS

How Independent Composers Can Verify AI Training Use of Their Work: Auditing, Disclosure, and Remedies