AI · 31 Aug 2026 · Rob Nidschelm

Australia’s AI Music ‘Ban’ Has One Big Weakness: It Starts With Self-Reporting

Australia did not ban AI music.

It did something more precise, and in some ways more revealing. The Australian Recording Industry Association, ARIA, changed the eligibility rules for its charts and awards. A recording can still use generative AI, but it cannot compete if AI generated the lead vocal, a key instrumental performance or the primary portion of the recording’s creative elements.

The rule took effect with the ARIA charts dated 31 August 2026 and published on 28 August. It immediately drew a line through one of Australia’s most visible AI-assisted hits.

Josh Fawaz’s cover of Madonna’s Like a Prayer had used generative AI for its vocals and drums, according to credits added to Spotify after an ABC investigation. The recording had topped an Australian dance chart, reached number two on the Australian singles ranking and become the country’s most-played song on radio.

It is absent from both ARIA’s Australian Singles and Australian Dance Singles charts for the week of 31 August.

The exclusion shows that the rule has teeth. It also exposes the control’s central weakness: ARIA says its process begins with what rights holders tell it.

This is not a national ban

The distinction matters because the headline has travelled faster than the policy.

No Australian law now prohibits AI-generated music. The rule does not prevent a track from being distributed, streamed, played on radio or monetised. ARIA says chart eligibility does not determine royalties, radio play or whether a streaming service carries a recording.

This is an industry eligibility standard administered by the organisation that runs Australia’s principal music charts. A recording classified as AI-generated cannot enter those charts or qualify for an ARIA Award. ARIA can also remove an ineligible recording retrospectively, adjust chart positions, withdraw accreditations and revoke an ARIA number-one award.

That makes the decision consequential without turning it into legislation.

ARIA now divides generative-AI use into three practical categories:

• AI-generated: Not eligible. This includes a generated lead vocal, a generated key instrumental performance or a recording whose main creative elements came from a prompt.
• AI-assisted: Eligible. Humans wrote and performed the central parts, while generative AI contributed supporting elements such as backing vocals or a secondary instrument.
• AI in production: Eligible. AI mastering, stem separation, effects, instrument patches and similar production tools do not disqualify a recording.

The boundary is therefore not whether AI touched the track. It is whether AI replaced the human performance at its centre.

One Madonna cover forced the question

Fawaz’s Like a Prayer cover became the obvious test because it succeeded before listeners were given a clear account of how it had been made.

ABC News reported on 17 July that Fawaz had added credits for generative-AI vocals and AI drums to the track on Spotify. By then, it had spent four weeks at number one on Australia’s national radio airplay chart and accumulated more than 38 million Spotify streams. It had also reached number one on ARIA’s Australian Dance Singles chart and peaked at number two on the Australian Singles chart.

The controversy was not simply that AI had been used. Many modern recordings use automated or AI-based production tools. The harder issue was that the synthetic contribution involved the performance listeners would ordinarily associate with a human artist.

ARIA’s new rule now makes that distinction explicit. Its charts for the week of 31 August no longer list the track in the two Australian rankings where it had been a major presence.

That is a clear enforcement outcome. It is not yet proof that the system will reliably identify the next undisclosed AI hit.

ARIA’s answer to detection is: ‘You tell us’

ARIA’s own FAQ is unusually direct about how it will know whether a recording qualifies.

Every release submitted to the ARIA survey must now include a declaration about generative-AI use. ARIA works from that declaration. If a credible concern is raised, it asks the rights holder for more information and allows the artist or representative to submit evidence and appeal an exclusion.

This is a reasonable starting point. It creates an accountable statement, gives ARIA a basis for enforcement and avoids pretending that automated detectors can settle difficult creative questions with certainty.

But self-reporting works best on parties already willing to disclose accurately.

The harder cases are the tracks whose creators do not declare AI use, use a distributor with incomplete metadata or describe a generated lead performance as ordinary production assistance. ARIA does not publish a list of the recordings it investigates or excludes. That protects individual proceedings, but it also limits outside scrutiny of how consistently the rule is applied.

The result is a familiar assurance problem. A policy can define the boundary perfectly while still depending on evidence supplied by the party being assessed.

Detection is arriving, but it is not a complete answer

The music industry’s technical response is beginning to form around the policy.

On 25 August, entertainment-data company Luminate announced a framework to identify AI-generated music using streaming data, metadata and audio-signature matching. Its CONNECT platform will eventually label tracks and artists that appear to be AI-generated, with a review process for disputed classifications.

Luminate is careful about the limitations. Its AI labels are not yet visible in CONNECT, the full database has not been screened and the absence of a label does not mean a recording is human-made.

That is the right caveat. Audio detection can support an investigation, but it should not become an opaque machine verdict. Generated audio can be edited, mixed with human performance, re-recorded or passed through additional processing. At the same time, false positives could unfairly damage independent artists who cannot easily produce studio records or provenance evidence.

A durable system will need several signals:

• Disclosure from the artist and rights holder
• Attribution metadata supplied by AI-generation services
• Consistent labels passed through distributors and streaming platforms
• Detection used as a risk signal rather than the only judge
• Evidence trails that show which elements were generated and which were performed
• A transparent review and appeal process

The International Federation of the Phonographic Industry is pushing in this direction. Its global chart principles require recordings to be substantially human-made, free of manipulation concerns and developed with authorised, lawful AI services. A separate voluntary labelling standard distinguishes AI-generated tracks from AI-assisted ones.

ARIA has adopted the human-performance and manipulation tests, but says it has not yet applied IFPI’s authorised-service test because licensing arrangements are still developing. That leaves another important question unresolved: a recording may satisfy ARIA’s creative threshold even when the legal status of the model behind its supporting AI elements remains contested.

The real test starts after the first obvious case

Removing a disclosed AI-vocal hit from a chart is the easy enforcement case.

The harder test is whether the system can identify a recording that becomes popular without a disclosure, survives automated screening and is backed by a rights holder prepared to defend an ambiguous production history.

ARIA’s rule is still important. It establishes that charts are not neutral databases of every stream. They are governed systems that decide which kinds of creative achievement they exist to recognise. It also gives the industry a concrete boundary: AI can assist the performance, but it cannot quietly become the performer.

Now the evidence model has to catch up with the policy.

The success of Australia’s approach will not be measured by whether one famous AI cover disappeared from a chart. It will be measured by whether the next one has to disclose what it is before listeners discover it for themselves.

Sources

• ARIA: Charts set eligibility rules for recordings made with AI, published 25 August 2026. No publication time provided.
• ARIA: AI and the ARIA Charts FAQ, accessed 31 August 2026. No publication time displayed.
• ARIA Top 20 Australian Singles, week of 31 August 2026, accessed 31 August 2026.
• ARIA Top 20 Australian Dance Singles, week of 31 August 2026, accessed 31 August 2026.
• ABC News: Josh Fawaz’s Like a Prayer now has AI credits on Spotify, published 17 July 2026 at 12:08 pm AEST and updated at 12:54 pm AEST.
• IFPI: Global principles for AI-developed recordings in official charts, published 30 July 2026. No publication time provided.
• IFPI: Track-level labelling for AI-generated and AI-assisted recordings, published 10 July 2026. No publication time provided.
• Luminate: Global framework to identify and measure AI-generated music, published 25 August 2026. No publication time provided.