Australia Banned AI From Its Charts. Enforcement Is a Checkbox.

TL;DR

On 25 August the Australian Recording Industry Association became the first national chart body to name a live date for the record industry's new AI eligibility rules. From the chart dated Monday 31 August, published Friday 28 August, a wholly AI-generated recording cannot chart in Australia and cannot be recognised at the ARIA Music Awards. The rule has real teeth, including retroactive removal and revoked number one awards. It also has one soft spot: the thing that triggers all of it is a declaration the person submitting the track fills in about themselves.

---

What actually changed

The ARIA Charts Code of Practice now says a recording developed using generative AI is eligible only where all three of these hold:

• It is substantially human made.
• It raises no stream or chart manipulation concerns.
• It complies with applicable laws, including those relating to copyright and related rights.

To decide which side of the line a track falls on, ARIA imports the definitions from the labelling standard announced by the global music community on 10 July by IFPI alongside the RIAA, A2IM, WIN, IMPALA, The Grammys, SAG-AFTRA and the Human Artistry Campaign. Two labels, one boundary.

AI-Generated covers a recording where generative AI produced the whole thing or the primary part of its creative elements. ARIA's own chart changes FAQ is blunt about the test: "A lead vocal or a key instrumental performance generated by AI puts a recording here." That bucket is ineligible.

AI-Assisted covers the case where "humans wrote the song and performed the lead vocal and the primary instruments" and generative AI did something minor on top. That bucket charts normally.

So the entire regulatory weight lands on two artifacts: the lead vocal and the primary instrumental performance. Your stem separator, your mastering chain, your generated pad in bar 33, none of it is the thing being tested.

The song this rule was written for

Brisbane producer Josh Fawaz released a rework of Madonna's "Like a Prayer" that did the numbers most human artists never see. It peaked at number two on the ARIA Top 20 Australian Singles chart, hit number one on the ARIA Top 20 Australian Dance Singles chart, and sat at number one on Australia's national radio airplay chart for four weeks. It reached number four on Billboard's Hot Dance/Electronic Songs and number one on the UK Official Independent Singles Breakers chart.

In July, after sustained criticism from Australian musicians, Fawaz added two credits to the Spotify listing: "Generative AI vocals" and "AI drums."

Read that against the definition above. An AI lead vocal is precisely the thing that drops a recording into the ineligible bucket. The track that produced the rule would not survive the rule. It also finished its entire chart run before the rule existed, and it kept accumulating streams the whole time it was being argued about: past 38 million on Spotify by mid-July, and 48.5 million by the day ARIA published the change.

The teeth are retroactive

This is the part that separates a chart rule from a disclosure badge. Where ARIA decides a recording is ineligible, it says it may decline to accept it for survey, exclude or remove it from the charts prospectively or retrospectively, adjust chart positions, withdraw accreditations, and revoke or request the return of an ARIA number one award. Wholly AI tracks are also out of contention for the ARIA Music Awards.

A chart is normally treated as a historical record. ARIA has just reserved the right to edit history, which is a genuinely unusual power for a trade body to write down in public.

There is a process attached. ARIA says it will contact you or your representative, state the concern, and give you a chance to respond with evidence before deciding. If you disagree, the matter can be referred to the ARIA Chart and Marketing Committee and then the ARIA Board.

The soft spot, and why it is smarter than it looks

Here is the sentence that matters most in the whole FAQ: "Every release submitted to the ARIA survey now requires a declaration about the use of generative AI, and that declaration is what ARIA works from."

No classifier. No forensic audio analysis at ingest. A field the submitter fills in about their own record. It is the customs card at the airport: you tick "nothing to declare", and the bag only gets opened when something else makes the officer curious.

The obvious criticism is that self-attestation collapses exactly where the incentive to lie is strongest. That is true, and it is why Apple Music's self-reported tag is a weak instrument. But a chart is a different shape of problem, and the design accidentally fits it.

Deezer says fully AI-generated tracks passed 50% of its daily uploads in June, around 90,000 a day. You cannot manually adjudicate 90,000 declarations a day. You do not have to. A chart is top-N by construction, so the surface that actually needs policing is roughly a hundred records a week, every one of which is already under a spotlight, already commercially significant, and already surrounded by competitors with a strong motive to complain. Enforcement scales because the prize does not.

And the payoff structure is inverted from a badge. Lying on the declaration gets you a chart position. Getting caught later loses you the position, the accreditation, and the award, after the press cycle has already run. That is a bet with a bad tail.

This is a global rollout, not an Australian quirk

ARIA is implementing something bigger. On 30 July in London, IFPI published global principles for the eligibility of AI-developed recordings in official charts. Six criteria, all of which must be satisfied:

• Any generative AI services used are properly authorised and lawful.
• The recordings are substantially human made.
• They do not raise manipulation concerns.
• They comply with applicable laws, including copyright, related rights and personality rights.
• Making them available does not breach the terms of the AI service used.
• The use of AI is appropriately signalled to consumers on downstream services.

Criterion one is the sharp one, and it is not really about the recording at all. It is a supply chain requirement on your tooling. IFPI is explicit that many AI tracks reaching streaming services "have been developed with tools that have been trained on artists' music without authorization." Under this rule, the training provenance of the model you generated with becomes an eligibility question about your song.

IFPI runs 11 official chart programs directly, covering more than 30 territories across Latin America, the Middle East, Africa and Southeast Asia. It says more than 20 additional national chart programs are implementing the principles, including France's SNEP charts, Germany's Offizielle Deutsche Charts and South Korea's Circle Chart. ARIA is the first of those to publish a live date.

ARIA chief executive Annabelle Herd framed it as a floor rather than a ceiling, and pointedly aimed the next request at broadcasters: "We call on all parties who have a role in deciding the music played and promoted to Australian audiences, particularly radio, to support human artistry and implement similar changes across their own codes." Given that the Fawaz track's biggest number was a radio number, that is not a subtle ask.

What a builder should take from this

Strip the music out and there is a reusable provenance pattern here, one that applies to any ranked or curated surface you operate.

Police the prize, not the pipe

Verifying every artifact at ingest is a losing arithmetic problem the moment generation is cheap. Verifying the small set that wins something is tractable forever. If you run a leaderboard, an app store chart, a trending feed or a grant program, the enforcement point is where the reward concentrates.

Declaration plus audit-on-signal plus retroactive reversal

Three cheap mechanisms that compose into something with real deterrence, without a classifier in the hot path. The reversal is what makes the declaration honest, so if you build the first two and skip the third you have built a form, not a rule.

Draw the line on testable artifacts

"Substantially human made" is unfalsifiable on its own. "Did a human perform the lead vocal and the primary instruments" is something two people can actually disagree about productively. Pick proxies you can argue over.

Your model's training data is now a compliance surface

IFPI's first criterion makes the licensing status of the generator an attribute of the output. Anyone shipping a creative tool into a regulated distribution channel should expect that question to arrive from the distribution side, not just from a court.

Key Takeaways

• From the ARIA chart dated 31 August 2026, wholly AI-generated recordings are ineligible for the Australian charts and the ARIA Music Awards. ARIA is the first national chart body to publish a live date for the IFPI principles.
• The test is concrete: an AI-generated lead vocal or key instrumental performance makes a recording ineligible. AI used for minor expressive elements on a human performance stays eligible.
• Enforcement runs off a mandatory generative AI declaration at submission, not a detector. ARIA can remove tracks retroactively, adjust positions, and revoke number one awards.
• IFPI's six principles include that the AI service itself be properly authorised and lawful, which makes model training provenance an eligibility question about the song.
• More than 20 national chart programs have signed up, including France, Germany and South Korea. Watch for the second body to publish a date.
• The transferable pattern for builders: enforce at the top of the ranking where the payoff concentrates, and make the penalty retroactive so a cheap declaration carries real weight.

---

Sources: ARIA Charts set eligibility rules for recordings made with AI, ARIA Chart Changes FAQs, IFPI global principles for AI chart eligibility, ARIA on the global AI labelling standard, ABC News, ABC News on the Fawaz AI credits, Al Jazeera, The Register, Deezer Newsroom