Is the most-played song on Australian radio human-made, or generated by AI?
A cover of Madonna’s “Like a Prayer” credited to Josh Fawaz suddenly dominated playlists, streaming and headlines, and raised a single, urgent question for broadcasters, labels and rights organisations: who, or what, actually performed on the recording?
Reported figures (variously reported as ~35 million Spotify streams and a No. 1 placement on some radio airplay charts) have circulated in the press, and the 18-track debut album Dance Like Nobody’s Watching has been reported as entering ARIA artist charts. Those numbers have been widely cited in coverage but are not independently verified here. The disputed point is the recording’s provenance.
What do we mean by “generative AI” in music?
Generative AI music tools can produce new audio, instruments, arrangements and even sung vocals, from text prompts, conditioning examples or short audio seeds. Some systems (examples cited in recent research include Suno and Udio) let a user supply lyrics, specify a timbre and generate a full mix. That creates a spectrum from AI used as a helper in mixing and mastering to workflows where a model supplies the vocal or instrumental performance itself.
That distinction matters. Using an AI plugin to clean up a vocal is different from using a model to synthesise a singing voice that never passed through a human throat. Post-production steps such as pitch correction, EQ, compression and layering can also mask where a sound originally came from, which makes attribution technically and legally tricky.
What’s been said about the Fawaz recording?
Industry listeners and researchers say the track contains audio characteristics they associate with some generative outputs: heavily compressed textures, vocal artefacts and drum programming that some describe as uneven. Sam Whiting, senior research fellow at RMIT’s school of media and communication, said:
“This is a very … impressive vocal performance if it was delivered by a human but if it’s not, that brings in really worrying questions around what we value any more in terms of human expression.”
A producer/DJ calling themself Needs No Sleep described similar auditory cues and called undisclosed AI-generated content “the biggest issue in music right now.” Josh Fawaz’s public response has been short: on Instagram he said, “I use AI as a tool, ” and replied to critics, “It’s not that deep… What I care about [is] providing my listeners with good music.”
Credits on the release list Josh Fawaz as “performer” and Fadi Fawaz on synths/production. Radio networks ARN, Nova Entertainment and Southern Cross Austereo did not respond to requests for comment about their playlisting or AI policies, according to reporting.
What the rules say, and what they don’t
The Australian Communications and Media Authority (ACMA) worked into a revised Commercial Radio Code of Practice for 2026 a requirement that broadcasters disclose when synthetic voices are used on air from 1 July. That change targets synthetic voices used in programs and news, and, as reported by ACMA/Capital Brief, does not impose an equivalent, explicit disclosure requirement on music content.
That regulatory line creates a practical gap. Stations must label a synthetic presenter or announcer, but there is no parallel duty to flag a music track that may contain synthetic vocals or instrumentation.
Who gets paid, and what remains unsettled
On composition royalties the rules are clear and familiar. APRA/AMCOS told reporters: “The song Like a Prayer is a remix/cover of a musical work written by Madonna L Ciccone and Patrick R Leonard, ” and that the original human rights holders “will be entitled to be paid all performance royalties in the usual manner.” That means the writers and publishers of the underlying composition should receive their usual share when a cover receives public performance or airplay.
But composition payments are only one slice of the money flow. Master recording ownership, performer (neighbouring) rights and label revenue splits concern the sound recording itself. If a singing performance is wholly or substantially synthesised by a model, current practice and law are unclear about how performer royalties and master ownership should be allocated. APRA/AMCOS’ comment addresses the composition side but does not resolve those master and performer questions.
Can you prove a track was made with generative AI?
Short answer: not reliably at scale today. Two facts explain why.
- AI tools can and do generate plausible musical performances. Academic work, for example the Transactions of ISMIR paper “Analyzing User Interactions with AI Music‑Generation Platforms”, documents platforms that permit users to generate full tracks from text prompts, condition on lyrics and influence timbre. The existence of these capabilities makes plausibility assessments easier but proof still requires more than plausibility.
- Post-production breaks forensic signatures. Real-world producers routinely re-EQ, compress, pitch-correct, layer takes and re-master stems. Those processing steps erase many of the acoustic fingerprints researchers rely on to classify controlled samples. As a result, listening judgements, even from experienced engineers, are informative but remain subjective without session stems or an independent forensic analysis.
Detection research can classify controlled datasets, for example the SONICS dataset used in research, but classifiers trained on one model or workflow do not generalise perfectly to all hybrids in the wild. That technical limit is why experts call for provenance documentation as the practical path forward.
Why business leaders and broadcasters should care
- Trust and reputation. Audiences expect transparency from curated channels. Heavy rotation of undisclosed synthetic performances risks reputational damage if listeners or artists feel misled.
- Contractual and revenue exposure. Without clear rules on AI contributions, labels and broadcasters could face disputes over performer payments, master ownership and licensing terms.
- Regulatory friction and competitive arbitrage. A patchwork of partial disclosure rules, synthetic voices required to be labelled, music not, invites inconsistent practices and potential consumer protection scrutiny.
Operational checklist for broadcasters, labels and platform operators
Priority actions you can implement now to manage risk and preserve trust.
- Require provenance attestations before heavy rotation. Make an artist or label attestation mandatory for playlisting and promotional campaigns.
- Ask for verification when stakes are high. For tracks moving into heavy rotation or major campaigns, request session stems, dry vocal takes or DAW session files under confidentiality. Those are the quickest path to verification.
- Collect production metadata. Add required metadata fields to distribution and upload forms. Recommended fields: ai_contribution (none/minor/significant), ai_components (vocals/instruments/arrangement), tools_used (free text, include versions), session_stems_available (Y/N) and provenance_hash (SHA256 of files where appropriate).
- Pilot independent forensics. Universities and commercial Music Information Retrieval labs can provide comparative analysis against known model outputs. Use them when legal or commercial exposure is material.
- Update contracts and workflows. Require clear production disclosures in label and artist agreements and include an AI-contribution attestation in distribution terms.
Sample provenance clause for contracts or upload forms (adapt to counsel’s advice):
“The contributor attests that any material use of generative AI in the recording is disclosed in the upload metadata. ‘Material use’ includes any AI‑generated vocal or instrumental performance used as a primary sonic element on the master.”
What industry bodies and regulators should consider next
Waiting for law to catch up takes too long. Practical standards will protect audiences and rights holders faster than litigation. Collecting societies, broadcasters and distributors should promote a common attestation schema for music metadata. The fields above are a usable starter. They should also agree on when session stems must be made available to rights administrators or auditors. Major labels signing commercial arrangements with AI music companies should insist those deals include provenance and revenue-sharing terms that protect legacy rights holders.
ACMA’s recent move to require disclosure for synthetic voices shows regulators can act quickly in targeted areas. Extending similar transparency expectations to music, or at least encouraging voluntary industry standards, would close the current policy gap.
Voices and perspectives
“The reason why AI music is not being more heavily interrogated is because streaming has conditioned us all not to engage with music in a critical, proactive way, ”, Sam Whiting.
“If this is the future of music production … then we really are cooked, ”, Needs No Sleep.
“I use AI as a tool, ”, Josh Fawaz (Instagram).
Key questions, and honest answers
-
Is the Josh Fawaz recording definitively generated by AI?
There is credible expert suspicion and audio characteristics observers associate with some generative tools, but no public, independently verified forensic proof tying the recording to a specific model. Without stems, session files or an independent forensic report the question remains unresolved.
-
Do the original songwriters still get paid?
Yes. APRA/AMCOS said: “The song Like a Prayer is a remix/cover of a musical work written by Madonna L Ciccone and Patrick R Leonard, ” and that the original human rights holders “will be entitled to be paid all performance royalties in the usual manner.” That covers composition/public performance royalties for the cover.
-
Does the new radio code force stations to disclose AI music?
No. ACMA’s updated Commercial Radio Code of Practice 2026 requires disclosure for synthetic voices used on air from 1 July, but, as reported, it does not impose the same explicit requirement on music content.
-
Can platforms or broadcasters reliably detect AI-generated music today?
Not reliably in every case. Controlled research can classify model outputs, but real-world recordings that mix human performances, model outputs and extensive post-processing defeat many current detectors.
-
What should I do if I run programming or label ops?
Start by demanding provenance metadata and attestations, require stems for high-exposure placements, pilot independent forensic checks when needed, and update contracts to require disclosure of material AI contributions.
Where this is headed, the operational bet
Generative tools are already embedded in production workflows. The sensible operational move for labels, broadcasters and rights organisations is to treat provenance as an operational standard rather than a legal footnote. Implement metadata fields, require attestations in contracts, and build a verification desk for high-stakes placements. That approach preserves audience trust, clarifies revenue flows and gives the industry room to commercialise AI tools without sacrificing fairness.
If you run programming, commissioning or rights administration: require an AI-contribution attestation in your next contract cycle and pilot the metadata fields suggested above. Those two steps will reduce exposure and buy time for regulators and collecting societies to catch up with durable rules.