AI Music: Disclosure Rules, Platform Policy and What Happens to a Release

•18 min read•Every figure sourced

This page is a reference, not legal advice. It describes what the platforms' own policies, the statutes and the courts say; it does not tell you what to do about your release.

Every fact here carries a source and a date, because this is the fastest-moving area in recorded music: several of the policies below changed in the last twelve months, one law took effect five weeks ago, and several of the questions musicians most want answered have no settled answer anywhere. Where something is unresolved, this page says so rather than guessing.

The four cases, and why platforms treat them differently

"AI music" is not one thing, and no platform treats it as one. Four cases run through every policy here, and the one your release falls into decides almost everything else.

Fully generated. A prompt goes in and a finished recording comes out; human involvement is prompting, selection and release admin. Platform anti-spam machinery is built around this case — not because generation is prohibited, but because it scales to industrial volume: Deezer reported that fully AI-generated tracks passed 50% of all new music uploads at peak in June 2026 (Deezer, 21 July 2026).

AI-assisted. A human writes, performs, arranges or produces and uses generative tools for part of the work — a generated instrumental bed, a lyric passage, a synthesised backing vocal. Every disclosure standard here is built for this case, because, as Spotify put it, "the use of AI tools is increasingly a spectrum, not a binary" (Spotify Newsroom, 25 September 2025).

AI vocal cloning. A model reproduces a specific, identifiable human voice — the only case prohibited outright on major services regardless of quality, volume or disclosure, and the only one whose legal exposure sits outside copyright entirely, in publicity and likeness rights that vary by state and country.

AI mastering, stem separation, noise reduction, pitch and timing correction. Machine-learning tools applied to audio a human already recorded. Every framework here treats these as production tools, not generative contributions: Apple's tags trigger only when AI "generate[s] a material portion" of a recording or composition (Apple Music Specification), and the EU AI Act's marking duty expressly does not apply where a system "performs an assistive function for standard editing" or "does not substantially alter the input data" (Article 50(2), AI Act text). A separated stem or an ML-assisted master is not, on any current reading, a disclosable AI contribution.

The line sits there for economic reasons, not aesthetic ones: generation at volume enables spam and royalty dilution, voice cloning enables impersonation, and a mastering algorithm enables neither.

How a disclosure actually travels

You do not disclose to Spotify or Apple. You give information to a distributor, the distributor writes it into the delivery, and the service reads it. If the field does not exist in the delivery format, the disclosure cannot travel — which is why the standards work matters more than the announcements.

DDEX. The Electronic Release Notification Message Suite (ERN) is the transport for a release and its metadata. ERN 4.3.1 added two capabilities on point: "communication of the fact that a recording was made (fully or in part) by a generative AI tool," and "communication of the fact that a contribution was made, fully or in part, by a generative AI tool" (ERN Part 1, Annex A release notes). Two things follow: the disclosure is graduated — "fully or in part," not a boolean — and it operates at two levels, the recording as a whole and the individual contribution.

The second level is the important one. In DDEX's model an AI tool is a contributor with a role, exactly as a session guitarist is; implementations describe a recording-level flag carrying values of the form All / Partly / None alongside a special contributor of type generative AI whose contribution is itself qualified (Sound Credit). Element names are your distributor's problem; the shape determines what you will be asked. Not "is this AI?" but, role by role, which parts were generated.

Apple's delivery spec. Apple exposes the disclosure as an <ai_transparencies> block of one or more <ai_transparency> values, available at album and at track level: Composition ("AI was used to generate a material portion of any music compositions embodied in a track," including the lyrics), Track ("AI was used to generate a material portion of a sound recording"), Music Video, and Artwork (AI-generated album art, static or motion). "If omitted, none is assumed," and the field is currently labelled "optional; may be updated" (Apple Music Specification, confirmed 2026-09-10).

What "material portion" means. Nobody has defined it. Apple uses the phrase without a threshold; Bandcamp bans music generated "wholly or in substantial part" without defining substantial; the Copyright Office requires disclosure of AI content more than de minimis without quantifying it. This is unsettled, and it is the most consequential undefined term on this page. The workable rule: if a listener could point at a section and you would have to say "a model made that," it is material. Do not litigate the margin — over-disclosing currently costs nothing on any platform examined here; under-disclosing can cost the release.

Platform by platform

Spotify

Three mechanisms, all introduced or expanded since September 2025.

Impersonation. "Spotify will remove music that impersonates another artist's voice without their permission — whether that's using AI voice cloning or any other method." It covers releases that clone a real artist's voice "whether or not the uploader is pretending to be the artist or is presenting themselves as an 'AI version' of the artist," and releases naming the impersonated artist nowhere in metadata but whose "vocals are clearly recognizable as the exact voice of another artist." Narrow exceptions exist, "such as certain forms of parody." Spotify acts on claims, not detection: it "must receive a claim from the artist or someone acting on their behalf," filed under the Publicity / Likeness report type (Spotify, confirmed 2026-09-10). The platform rule behind it prohibits "content that impersonates others in order to deceive" (Platform Rules).

The music spam filter. Announced 25 September 2025 and rolled out from that autumn, it targets "mass uploads, duplicates, SEO hacks, artificially short track abuse, and other forms of slop"; identified tracks and uploaders are tagged and stop being recommended. In the same announcement Spotify disclosed it had "removed over 75 million spammy tracks from Spotify" in the preceding twelve months (Spotify Newsroom; MBW). Note what that figure is not: it counts spam removals over a year, not AI removals, and Spotify does not break out how many were AI-generated.

AI credits. Spotify helped develop the DDEX standard and launched a beta on 16 April 2026 surfacing disclosed contributions in Song Credits. "AI credits show when a specific part of a track is generated with AI. They apply to individual roles, not the whole track" — commonly lyrics, vocals, instrumental performances or production; where both AI and human contributions exist for a role, they are credited separately. Credits appear only on mobile, in Song Credits and Now Playing, and "don't appear on artist profiles or as track-level labels." They come through your distributor, and only distributors that have implemented the field can send them. They are optional, and the scheme "is not about punishing artists who use AI responsibly or down-ranking tracks for disclosing information about how they were made" (Spotify, AI credits; Newsroom, both confirmed 2026-09-10).

AI Persona badges — a separate thing. This concerns identity, not music. An AI Persona is "an artist whose public identity represents an AI-generated photorealistic human." Artists self-disclose in Spotify for Artists (Artist Settings → Profile & Identity → AI Persona), and cannot switch it off once enabled; Spotify also reviews profiles above defined audience thresholds and applies a "Likely AI Persona" badge where it concludes the identity is AI-generated. Music from AI Personas is excluded from editorial and personalised recommendations unless a listener intentionally engages with the artist, for example by following them. The badge does not apply if you use AI tools but your profile represents a real person, a stage name, or a fictional character (Spotify, confirmed 2026-09-10). This is the one Spotify disclosure with a documented distribution consequence.

Deezer

Deezer is the only service that detects rather than waits to be told. It began tagging AI-generated music in June 2025 and tagged 13.4 million tracks that year; in April 2026 it reported "nearly 75,000 AI-generated tracks per day, representing roughly 44% of the daily uploads" (Deezer, 20 April 2026). By June 2026, "90,000 AI-generated tracks per day now represent over 50% of all new music uploads on Deezer at peak level in June, 2026" (Deezer, 21 July 2026; corroborated by MBW).

Two further figures matter more than the headline. Fully AI-generated music accounts for only "between 1-3%" of total streams — the flood is upload-side, not listening-side. And "up to 85% of the streams generated by fully AI-generated tracks were in fact fraudulent in 2025," those streams being excluded from royalty payments. Detected AI tracks "are automatically removed from algorithmic recommendations and are not included in editorial playlists"; Deezer claims 99.8% detector accuracy.

Read that carefully: on Deezer a fully AI-generated track streamed legitimately is not removed, but it is not recommended either. The reach penalty is automatic and attaches to the category, not to your conduct.

Apple Music

Apple's AI Transparency Tags went live in the delivery specification in March 2026 (AppleInsider); the spec, confirmed 2026-09-10, still marks the field "optional; may be updated." That is changing. In August 2026 Apple told distributors that "content providers will be required to include AI Transparency Tags in any instance where AI was used to create a material portion of the content, including tracks that are AI platform generated," with listener-visible labels arriving later in 2026 (Hollywood Reporter; Variety; AppleInsider, 20 August 2026). No date and no enforcement mechanism had been published as of 2026-09-10. Check the specification before every delivery. Apple is the first major service moving disclosure from voluntary to mandatory, and this is the change most likely to date this page.

YouTube

YouTube runs the most developed synthetic-media regime here, but it governs videos, not the tracks your distributor delivers to YouTube Music.

Disclosure. Creators must disclose in YouTube Studio (Attributes → AI use) when AI makes realistic content showing someone appearing to say or do something they did not, alters footage of real events or places, or generates a realistic scene that never occurred. Labels appear in the player for photorealistic content, in the description otherwise. Minor edits are exempt, and the list is unusually explicit: beauty filters, colour adjustment, special effects, script or thumbnail assistance, upscaling and repair, clearly unrealistic content — and, notably, "voice cloning for one's own narration." Creators who "consistently choose not to disclose" face "manual application of a label, or penalties from YouTube, including removal of content or suspension from the YouTube Partner Program" (YouTube, confirmed 2026-09-10).

Voice and likeness removal. "If someone has used AI to alter or create content that looks or sounds like you, you can ask for it to be removed." The request runs through YouTube's privacy complaint process, must come from the affected person or a legal representative, and requires that "the content should depict a realistic altered or synthetic version of your likeness." YouTube weighs whether the content is altered or synthetic, whether it was disclosed as such, whether the person is uniquely identifiable and the depiction realistic, whether it carries parody, satire or public-interest value, and whether it shows a public figure in sensitive conduct (YouTube, confirmed 2026-09-10).

Likeness detection — proactive scanning for enrolled creators — currently covers faces only; YouTube calls it experimental, with limited availability, and states an intention to extend it to audio (YouTube, confirmed 2026-09-10). Until audio detection ships, a singer must find voice clones themselves and file one by one.

TikTok

TikTok's Community Guidelines require that "synthetic or manipulated media that shows realistic scenes must be clearly disclosed," by sticker or caption such as "synthetic," "fake," "not real," or "altered." Prohibited outright: synthetic media containing "the likeness (visual or audio) of a real person" where that person is a minor or a private adult, and synthetic media of public figures used for political or commercial endorsement (TikTok, confirmed 2026-09-10). The audio limb matters to musicians: a cloned voice of a private individual is not permitted at all.

TikTok began implementing C2PA Content Credentials on 9 May 2024, auto-labelling AI content arriving from other platforms that carries Content Credentials metadata, and committing to attach them to content made with its own tools — images and video first, audio-only to follow (TikTok Newsroom, 9 May 2024).

Bandcamp, Amazon Music, SoundCloud

Bandcamp has the strictest published stance here and the only outright ban. From 13 January 2026: "Music and audio that is generated wholly or in substantial part by AI is not permitted on Bandcamp," and "any use of AI tools to impersonate other artists or styles is strictly prohibited." Enforcement is human review of user reports, and Bandcamp "reserve[s] the right to remove any music on suspicion of being AI-generated" (Keeping Bandcamp Human, 13 January 2026). There is no disclosure route, because disclosure is not the remedy — exclusion is. "In substantial part" is again undefined, and removal follows suspicion rather than proof.

Amazon Music has no published AI-specific policy. Its content guidelines, as circulated to distribution partners, prohibit "generic content or artist naming conventions without brand clarity or musical identity" and reject bootlegs, but do not address AI generation, disclosure, soundalikes or voice cloning, and Amazon Music for Artists publishes no AI disclosure field or badge as of 2026-09-10 (Amazon Music for Artists). Treat Amazon as governed by its general infringement and impersonation rules — and the silence as a gap that will close, not as permission.

SoundCloud's published position concerns training on uploaded music rather than uploading AI music: it committed in May 2025 that "we will not use Your Content to train generative AI models that aim to replicate or synthesize your voice, music, or likeness without your explicit consent" (Digital Music News, 14 May 2025). That reverse question is worth asking wherever you upload.

The law

EU AI Act Article 50, and China

Article 50 transparency obligations apply from 2 August 2026 — five weeks before this page was written. They were not delayed by the Digital Omnibus agreement, which postponed high-risk-system deadlines to December 2027 and August 2028 but left Article 50 on schedule (European Commission; Gibson Dunn, 27 May 2026). The Commission states a "[g]race period for marking obligation until December 2026 for generative AI systems placed on the market before 2 August 2026"; no specific day in December is given.

Two duties, on two parties. Providers of generative systems must ensure outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated." That is the tool maker's obligation, not yours, and it does not apply where the system performs "an assistive function for standard editing" or "does not substantially alter the input data" (Article 50(2)).

Deployers — which can include you — must disclose deepfakes: AI-generated or manipulated image, audio or video resembling existing persons that would falsely appear authentic. Article 50(4) carries an artistic carve-out: where the content forms part of "an evidently artistic, creative, satirical, fictional or analogous work or programme," the duty is "limited to disclosure of the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work" (Article 50). Information must come "at the latest at the time of the first interaction or exposure" (50(5)). Penalties reach €15 million or 3% of worldwide turnover, reduced proportionally for SMEs (Commission factpage above).

China, separately. The Measures for Labeling of AI-Generated Synthetic Content took effect 1 September 2025 and are the only rules here demanding an audible in-content marker rather than metadata alone: explicit labels for AI audio at "the beginning, end, or middle of audio" (Art. 4(2)), metadata labels (Art. 5), and a duty on users to "proactively declare it and use the labeling functions provided by the platform" (Art. 10) (translation).

What is unsettled: whether an independent musician releasing a track containing a licensed synthetic voice is a "deployer" under Article 50(4), and what "appropriate manner" means for an audio-only release with no sleeve and no interface of its own. No guidance answers that as of 2026-09-10. The defensible reading: a track-level AI credit plus a plain statement on your own release page discharges a duty limited to "the existence of" generated content.

US copyright: what can and cannot be registered

The rule is human authorship, settled at appellate level. The Copyright Office's registration guidance, effective 16 March 2023, states that "copyright can protect only material that is the product of human creativity." Applicants must disclose AI-generated content more than de minimis, exclude it in the Limitation of the Claim field, and "provide a brief explanation of the human author's contributions to the work." Failure to disclose has consequences: the Office "may take steps to cancel the registration," and a court may disregard a registration under 17 U.S.C. § 411(b) where the applicant knowingly supplied inaccurate information (Federal Register, 16 March 2023).

The Office's Copyright and Artificial Intelligence, Part 2: Copyrightability report (January 2025) refines this. In its own terms: "prompts do not alone provide sufficient control" to make the user an author, because prompts convey unprotectible ideas while the system determines the expression; where human-authored material is perceptible in the output that contribution is protectable much as in a derivative work; creative modification of AI output, and creative selection and arrangement of it, can each support a claim; and "the use of AI tools to assist rather than stand in for human creativity does not affect the availability of copyright protection." These questions, it concluded, "can be resolved pursuant to existing law, without the need for legislative change" (Part 2 report; NewsNet 1060). The courts agree: in Thaler v. Perlmutter the D.C. Circuit held on 18 March 2025 that human authorship is required as a matter of statutory law (opinion), and the Supreme Court denied certiorari in March 2026 (Baker Donelson).

What this means for a mixed release. A human topline over a generated instrumental is registrable — as to the topline, the arrangement and the selection, with the instrumental disclaimed. A track generated end-to-end from a prompt is not registrable at all in the US on current Office practice. Recording and composition are separate registrations and can differ: a human performance of an AI-written melody yields a protectable recording embodying an unprotectable composition. You can still release what you cannot register — registration is not a precondition to distribution. What you lose is the ability to sue for infringement in the US, plus statutory damages and fees; and anyone may copy an unprotected generated track.

Performing rights organisations

On 28 October 2025 ASCAP, BMI and SOCAN announced aligned policies: they accept registration of "musical compositions partially generated using artificial intelligence (AI) tools," defined as works "that combine elements of AI-generated musical content with elements of human authorship," and those works enter each society's licensed repertory. "Musical compositions that are entirely created using AI tools are not eligible for registration with any of the individual societies" (ASCAP; BMI). A fully generated composition therefore collects no performance royalties through these three societies; a partially generated one does.

Voice, likeness and the right of publicity

This is where the exposure is, and it is state and country law, not copyright.

Tennessee's ELVIS Act (Tenn. Code Ann. § 47-25-1101 et seq., effective 1 July 2024) added voice to the protected property right in name, photograph and likeness. It gives standing not only to the individual and their heirs but to holders of personal-services contracts or sound-recording distribution rights, and it reaches upstream, prohibiting distribution of technology whose "primary purpose or function" is producing a particular identifiable individual's voice or likeness without authorisation (Rothman's Roadmap — Tennessee).

California enacted AB 1836 and AB 2602 in September 2024, covering digital replicas of deceased performers and requiring specificity in contractual consent to replicas of living ones (Skadden, September 2024).

Federally, nothing has passed. The NO FAKES Act would create a federal digital replication right; the Senate Judiciary Committee advanced it by unanimous voice vote in June 2026 (Holland & Knight; text as reported). It is not law. Do not plan around it. In Europe, Denmark has moved to give individuals a copyright-style right in their own likeness and voice (Euronews, 6 November 2025) — unsettled and in progress.

The operative point is jurisdictionally messy: which law applies to a cloned voice depends on where that person lives and where the release is exploited, and those answers differ.

Training-side litigation

On 31 July 2026 the Regional Court of Munich I held in GEMA v. Suno (Case No. 42 O 763/25) that where training works remain recognisable in outputs this constitutes reproduction, and that supplying generated songs sufficiently similar to training works is an unlawful communication to the public — rejecting a fair-use defence notwithstanding that training occurred in the US. The decision is not final and is subject to appeal (DLA Piper, August 2026; Bird & Bird). Its exposure for a releasing artist is untested — but an output audibly reproducing a recognisable existing work is your problem regardless of who trained what.

Before you release: the checklist, in order

  1. Classify the track honestly, role by role. Vocals, lyrics, instrumental performances, production, artwork, video: for each, write down human, AI, or both. That list is what the disclosure fields want.
  2. Confirm no identifiable voice was cloned. Not a style, not a genre — a specific recognisable person's voice. If one was, you need that person's written authorisation. The only item here with no workaround.
  3. Read the licence for every tool you used, and keep it as it stood on the day you generated. Commercial-use rights, ownership of outputs and vendor attribution duties vary widely and change.
  4. Check the output against existing recordings. If a generated passage sounds like a specific record, treat it as an uncleared sample and replace it. Similarity to a training work is an infringement question, not an AI one.
  5. Ask your distributor whether it supports DDEX AI disclosure and Apple's AI Transparency Tags. If not, you cannot disclose on Spotify or Apple whatever you want to declare — and Apple's move to mandatory tagging makes that a delivery risk, not a preference.
  6. Complete the fields, and over-disclose at the margin. Spotify says disclosure does not down-rank tracks; Apple assumes "none" where the tag is omitted, the assertion you least want to make wrongly.
  7. Decide the AI Persona question separately. If your public artist identity is an AI-generated photorealistic person, self-disclose on Spotify and accept the exclusion from editorial and personalised recommendation. It does not apply if your identity is a real person or a stage name, whatever tools you used.
  8. Handle registration and society filings. If AI content is more than de minimis, disclaim it in the US registration's Limitation of Claim and explain the human contribution. Register partially AI-generated compositions with your PRO; a fully generated one will not be accepted by ASCAP, BMI or SOCAN.
  9. Say it plainly on your own pages — website, release page, press notes. The cheapest way to satisfy an "appropriate manner" duty under Article 50(4), and the only channel you control.
  10. Keep the evidence. Prompts, tool versions, session files, dates, stems, any voice licence. If a track is challenged, the burden of showing what you did falls on you, months later.

When it goes wrong

Removal. The commonest outcome. Voice impersonation goes on claim on Spotify; Bandcamp removes on suspicion; anything infringing goes on notice. Removal is reversible in principle — you can redeliver — but the original loses its release-date position, editorial support, and accumulated saves and playlist adds.

Reach suppression, which is not removal and is not appealable in any published process. The underrated consequence. Spotify's spam filter tags identified tracks and uploaders and stops recommending them; Deezer removes detected AI tracks from algorithmic recommendation and editorial playlists; Spotify excludes AI Persona artists from editorial and personalised recommendation. The track stays live and streamable and simply stops being surfaced — no published appeal route, no notification. A release that quietly does nothing is the ordinary shape of this failure.

Royalty withholding. Real, documented, largely unrecoverable. Spotify: "When we identify confirmed cases of artificial streaming or stream manipulation, we take actions that may include the withholding of associated royalties, the correction of public streaming numbers, and measures to ensure the artist or song's popularity is accurately reflected in our charts." It shares monthly reports with distributors, and "your distributor may take actions like issuing warnings or, in flagrant or repeated cases, removing your content from streaming services or suspending your account." It also "charge[s] labels and distributors when flagrant artificial streaming is detected on their content" (Spotify; Spotify for Artists, both confirmed 2026-09-10). Deezer excludes fraudulent streams from royalty payments outright. Note where that charge lands: on your distributor, whose terms pass it to you. That is why fully generated catalogues attract distributor scrutiny even where platform rules permit them.

Account action. "Repeated violations of our prohibited content policies can result in losing access to the Spotify platform" (Spotify). YouTube's non-disclosure penalties extend to "suspension from the YouTube Partner Program." Distributor termination usually takes the whole catalogue down, not just the offending release.

What is and is not recoverable. Withheld royalties on streams found artificial, and stream counts corrected downward: not recoverable, no published appeal. A removed release: redeliverable, but artist-page history, release-date positioning and playlist adds are not. A terminated distributor account: your recordings and ISRCs remain yours — an ISRC stays with the recording — but re-establishing delivery means a new agreement and, in practice, a gap. A registration cancelled for non-disclosure: refilable, but the original effective date is lost, and with it the statutory-damages position for the intervening period. A publicity-rights claim over a cloned voice is not a platform matter at all, and no takedown resolves it.

What is unsettled, as of 2026-09-10

  • "Material portion" and "substantial part" have no published definition at Apple, Bandcamp or the Copyright Office. Every disclosure threshold here rests on an undefined term.
  • Apple's mandatory tagging has no announced date or enforcement mechanism — signalled in August 2026 for "later in 2026."
  • Whether an independent artist is a "deployer" under Article 50(4), and what "appropriate manner" means for an audio-only release, is unguided.
  • No federal US voice/likeness right exists. NO FAKES advanced out of committee in June 2026 and is not law.
  • Detection is one-sided. Deezer detects; Spotify and Apple rely on self-report, and Spotify says plainly that "[b]ecause we depend on artist disclosure, the absence of a credit doesn't mean AI wasn't used" (Spotify Newsroom). A disclosure regime only the honest use is a known weakness.
  • Training-side liability is unresolved. GEMA v. Suno is first-instance and appealable, and untested against releasing artists rather than tool vendors.
  • Amazon Music, and most services outside the majors, have published nothing. Silence is not permission; it is an absence that will be filled.

Anything on this page dated 2026 should be re-checked before you rely on it.

Sources

FREE TOOLS

Every tool Mazufa builds runs in your browser, costs nothing, and needs no account.

Open the toolkit ⇥