Music platforms are starting to ban AI songs. Where the line is
Beatport, ARIA and the EU AI Act all drew lines around AI music in 2026. What each one bans, how it is checked, and what it means for your tracks.
Within three weeks in August 2026, a DJ marketplace stopped accepting AI tracks and Australia's chart body stopped counting them. Both drew the same distinction, between a recording a model produced and one made by people who used a model along the way, and neither published a percentage that tells you which side your track sits on.
What follows is four entries, each with what it says in its own words, how it gets checked, and a verdict, then what all of it adds up to if you open a generator tonight. One disclosure first: MultiChats has no music generation. The only audio the product generates is the text-to-speech that reads answers aloud in voice mode.
Beatport, 12 August 2026
Where it comes from: Beatport's own announcement, published on its editorial site Beatportal and dated 12 August 2026, plus the undated policy page Beatport's Stance on Artificial Intelligence. The widely shared 28 August date belongs to The Decoder's write-up rather than to Beatport.
What it says: "it is our policy not to accept music on our platform that has been wholly or substantially generated by AI". The same page narrows the target: "this policy targets generative AI and the use of AI to clone or replicate the voices or performances of other artists without explicit, legally cleared permission", while Beatport is "not discouraging the use of assistive AI tools for mixing, mastering, stem separation, or standard audio processing". Its Content Policy lists "AI-generated music" under "What we don't want". CEO Matt Gralen, in the announcement: "there is a difference between a tool that assists human creation and a system that replaces it entirely. Beatport is built on the former."
How it is checked: by an expanded partnership with the streaming-fraud detection firm Beatdapp. Fully or majority AI-generated tracks are withheld during ingestion and the rightsholder is notified directly. Majority human-made tracks are accepted and tagged, giving Beatport's curation team a flag when it picks what to feature. In the survey Beatport cites, 60% of DJs said they would not play AI-generated tracks in their sets and 8% said they were open to it.
Verdict: a commercial door policy, enforced by a fraud vendor's classifier at upload. "Wholly or substantially" is the whole test, and Beatport has not said what substantially means in minutes, stems or percentages.
What Australia's recording industry body did to its charts
Where it comes from: ARIA's announcement of 25 August 2026, with a companion FAQ. The revised Charts Code is dated 24 August 2026 and the change took effect "from the ARIA Chart dated Monday, 31 August 2026".
What it says: "Wholly AI-generated tracks will not be eligible for the ARIA Charts. Recordings that use generative AI in a supporting role remain eligible." Eligibility needs a recording to be "substantially human made", to raise "no stream or chart manipulation concerns", and to comply with copyright law and any required AI labelling standards. The FAQ is blunter: "A lead vocal generated by AI makes a track ineligible, the same as any key instrumental component of a song. A human lead vocal with AI backing vocals underneath is considered AI-assisted and therefore eligible." Human mixing, mastering, EQ and editing change nothing either way. ARIA CEO Annabelle Herd: "music generated wholesale by services built on artists' recordings is a different matter."
How it is checked: by declaration, against borrowed definitions. ARIA applies the definitions in the industry labelling standard announced on 10 July 2026, and tells releasing parties that "what matters for the ARIA Charts is that your declaration to ARIA on the use of generative AI is accurate". Where it finds a recording ineligible it can refuse it for survey, remove it from the charts retrospectively, adjust positions and revoke an ARIA #1 Award, and strip ARIA Award eligibility. The disputes process was rewritten so artists can contest an exclusion with evidence, and the ARIA Board decides.
Verdict: the more legible of the two rules, because the definitions, the sanctions and the appeal route are published. It also admits the difficulty out loud: on "substantially human made", the FAQ says human creative input must be "the primary driver of the work", assessed case by case, "and we expect that to get harder before it gets easier".
The Suno source code, and why it is an allegation
Where it comes from: 404 Media's reporting on material stolen from the AI music service Suno, with TechCrunch covering it on 21 July 2026. The breach itself and what to do if you had an account are a separate story.
What it says: the stolen dump included Suno's source code, which those reports say revealed how the company allegedly scraped millions of songs and lyrics from sites including Deezer, Genius and YouTube to train its models. All of that is an allegation drawn from leaked material read by journalists, and Suno has not confirmed it. Major labels were already suing the company over mass scraping when the material surfaced.
How it is checked: it is not, by us or by anyone with subpoena power yet. Treat it as reporting on documents nobody meant to publish, worth knowing and too thin to carry a conclusion.
Verdict: it moved the argument from what a generator outputs to where its training set came from. We cannot show it caused either August policy. What is visible is that both reach for licensing language, ARIA warning against "unlicensed AI output" and Beatport singling out voice cloning "without explicit, legally cleared permission".
Europe's marking rule, and what it does not reach
Where it comes from: Article 50 of the EU AI Act, applying from 2 August 2026, read alongside the Commission's Article 50 FAQ and its Code of Practice on Transparency of AI-generated Content, signed by about 190 organisations by the end of July 2026, with two task forces due to launch in September 2026.
What it says: providers of systems generating synthetic audio "must ensure that AI-generated or manipulated content are marked in a machine-readable format and detectable as artificially generated or manipulated". That mark is written for detectors rather than ears. The separate deepfake labelling duty is narrower than people assume, because a deepfake is defined as content "that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful". A cloned vocal of a named singer sits inside that definition; an original prompt-made instrumental sits outside it. For "evidently artistic, creative, satirical, fictional or analogous works" the duty shrinks again, to disclosure "in an appropriate manner that does not hamper the display or enjoyment of the work".
How it is checked: mainly by national market surveillance authorities, with a limited role for the Commission's AI Office. The date to hold on to is the FAQ's carve-out: for AI systems placed on the market before 2 August 2026, the marking and detection duty applies "only as from 2 December 2026". Most generators you have heard of sit in that category, so this layer is not fully in force yet.
Verdict: useful later, and to platforms more than to listeners. A machine-readable mark helps Beatport's or Deezer's detector, and says nothing about whether a shop or chart has to accept the track. Older output stays unmarked, since the Commission only encourages labelling content made before August.
Three texts, one boundary, no number
Read the three definitions together and they converge on a place without measuring it. Beatport bars what is "wholly or substantially generated by AI". ARIA admits what is "substantially human made". The 10 July 2026 industry standard, from IFPI, RIAA, A2IM, WIN, IMPALA, The Grammys, SAG-AFTRA and the Human Artistry Campaign, applies its "AI-Generated" label where "generative AI was used to generate the entirety or the primary portion of the creative elements of the recording", and "AI-Assisted" where the recording "was created substantially by humans and expresses human creativity; however, generative AI was used for some expressive elements. Humans performed the lead vocal and primary instruments."
Each of those hinges on a word a lawyer would call a question of fact, and the worked examples do more than the definitions. An AI lead vocal is out. AI backing vocals under a human lead are in. A voice-to-voice translation counts as assisted, because a human performance is what it works from. Mixing and mastering launder nothing in either direction.
Two gaps are worth naming. The industry labelling system covers sound recordings only, and says so: it does not yet cover generative AI in lyrics, composition, music videos or cover art, so a song with a prompt-written lyric and a human performance carries no label under that standard. The enforcement layers also differ in kind, a classifier at ingestion for Beatport against a declaration audited later for ARIA, so two tracks made the same way can land differently.
The scale explains the hurry. Deezer, tagging AI tracks since 2025, reported on 20 April 2026 that it was receiving nearly 75,000 fully AI-generated tracks a day, about 44% of daily deliveries, while AI music drew 1 to 3% of streams, 85% of which were flagged as fraudulent. In a study it commissioned, 97% of listeners could not hear the difference and 80% wanted fully AI-generated music labelled.
If you make music with AI tools today
Where it comes from: the four entries above, read together.
What it says: keep the human performance at the centre, the lead vocal and the primary instruments, and everything these texts call assistance stays open. Beatport's wording names mixing, mastering, stem separation and ordinary audio processing as acceptable. Cloning a real artist's voice without cleared permission is the one move every text rules out.
How it is checked: two ways, both pointing back at you. The chart rule turns on your declaration, so an inaccurate one carries more exposure than the music does. The marketplace runs a detector, so expect mistakes in both directions and keep the session files, stems and dated project history that would answer a challenge, which is what ARIA's dispute route asks for.
Verdict: if the AI-made parts of a track are load-bearing, plan for it to be refused somewhere. A version re-recorded with human performance counts as a new recording and can go to the ARIA survey on its own merits, without inheriting the earlier version's chart history.
What to watch next
The transparency code set September 2026 as the start date for its task forces, the EU marking duty reaches existing generators on 2 December 2026, and ARIA has asked radio and others deciding what gets played to write similar rules into their codes. Beatport says its policy will develop. Where substantial stops is still unsettled, and every one of these documents leaves it to be answered track by track.