AI-Written Music, Copyright and APRA AMCOS: A Practical Guide for Australian Songwriters

Somewhere in Australia this week, a songwriter opened an AI tool, typed a prompt, and produced something that sounded, to their ear, like a finished song. Whether that moment created anything they can actually own is a much harder question than it looks, and the answer has just shifted twice in the space of seven days. On 25 August 2026 the Australian Recording Industry Association banned recordings that are wholly or primarily AI-generated from the ARIA Charts and ARIA Awards.¹ Days later, an American publisher filed billion-dollar lawsuits against two of the biggest AI music platforms on the market,² a German court found one of those same platforms liable for reproducing existing songs from memory,³ and Apple Music told the industry it will start labelling tracks made with AI.⁴ None of this happened by accident, and none of it happened in isolation from a much older story: Australian copyright law has been arguing about what counts as genuine human creativity in music since long before anyone had heard of a large language model.

This guide is written for the person who actually has to navigate all of that: the working or aspiring Australian songwriter who has, at some point, let an AI tool touch a lyric, a melody, a mix, or an entire track, and now wants a straight answer about where they stand. We cover what Australian copyright law says about AI-assisted and AI-generated music, what moral rights add to that picture, what APRA and AMCOS are actually doing with your royalties right now, how the new ARIA chart rule and the Grammys’ own “meaningful human authorship” test compare, what happens if an AI clones your voice, and finally, a genuinely practical checklist you can act on this week, not once the law eventually catches up.

🎙️ Listen to the related Legal Matters episode: AI Music, Copyright and APRA AMCOS: Your Legal Questions Answered .

Why This Suddenly Matters: The Week’s Developments

Music copyright law has spent the better part of two decades being tested at the edges by sampling, mashups, streaming economics and remix culture. Generative AI has compressed years of that slow, grinding argument into a matter of months, and the past fortnight is the clearest evidence yet of how fast the ground is moving beneath songwriters’ feet.

First, ARIA’s Charts Code of Practice was rewritten on 24 August 2026 to draw a firm line between AI-generated and AI-assisted recordings.⁵ Under the revised code, a recording is only eligible for the charts if it is “substantially human made,” raises no stream or chart manipulation concerns, and otherwise complies with applicable laws, including copyright.⁶ A track is treated as AI-generated, and therefore ineligible, where generative AI produced the entirety of its creative elements or the primary portion of them, including where the lead vocal or a key instrumental performance was AI-generated, even if a human subsequently mixed or mastered the result.⁷ ARIA’s own published FAQ is blunt on this point: human mixing, mastering, equalisation and editing do not rescue an otherwise AI-generated track’s eligibility.⁸ The trigger, reported by the ABC and picked up internationally by the New York Times, was that a track that was almost entirely AI-generated had come close to topping the ARIA Charts, forcing the industry body’s hand before the next chart cycle.⁹

Second, the independent publisher Round Hill Music filed separate federal lawsuits against the AI platforms Suno and Anthropic on 17 August 2026, alleging unlicensed use of its catalogue and seeking statutory damages of up to USD 150,000 per work across what could ultimately be more than 500 compositions, a claim that could exceed USD 1 billion across both cases.¹⁰ Round Hill’s catalogue includes songs recorded by artists such as The Kinks and the Goo Goo Dolls, and the publisher has said it has no intention of settling either action.¹¹

Third, and arguably the development that matters most for anyone trying to assess real legal risk rather than reputational risk, the Munich Regional Court ruled on 31 July 2026 that the AI music generator Suno infringes copyright through “memorisation,” meaning the model can be prompted to reproduce protected songs from its training data closely enough to amount to unauthorised copying, in proceedings brought by the German collecting society GEMA.¹² This is one of the first substantive court findings anywhere in the world that an AI music generator is directly liable for what its outputs actually contain, as distinct from how it was trained, and it will be studied closely by APRA AMCOS, by Australian courts if a comparable dispute ever lands here, and by every songwriter wondering whether an AI tool might, without anyone intending it, spit out something that echoes their own earlier work.

Fourth, Apple Music told partners in the week of 17 August 2026 that it will apply “Made With AI” labels to tracks materially generated on platforms such as Suno, effectively pushing the responsibility for tagging AI content back onto labels and distributors before a track even reaches listeners.¹³ Universal Music Group also disclosed a portfolio of more than twenty issued AI music patents, licensed to platforms including Udio, a clear signal that at least one major label is moving to formalise and monetise AI’s role in production rather than simply fight it in court.¹⁴

Read together, these four stories tell you exactly where the pressure is coming from. Regulators and industry bodies are trying to preserve a category of “real” human music that audiences and voters can trust, courts are beginning to test whether AI outputs directly infringe the songs they were trained on, and the money, both licensing revenue and litigation exposure, is moving at a pace Australian copyright law was never built to match. Underneath all of it sits the same question courts have wrestled with for as long as there has been recorded music: what actually counts as a human creative contribution, and how much of someone else’s original work can a new work borrow before it crosses the line into infringement.

Can You Actually Copyright an AI-Written Song in Australia?

Start with first principles. The Copyright Act 1968 (Cth) protects original literary, dramatic, musical and artistic works, but only where the originality comes from a human author exercising independent intellectual effort.¹⁵ Australian courts have applied this requirement strictly, and the doctrine long predates the current wave of generative AI tools; it was forged in a series of cases about compilations, databases and automated document generation that had nothing to do with music but everything to do with the underlying legal question.

In IceTV Pty Ltd v Nine Network Australia Pty Ltd [2009] HCA 14, the High Court considered whether IceTV had infringed Nine’s copyright in its television guide by using time and title information drawn from Nine’s own weekly schedule.¹⁶ The Court found no infringement, but in reaching that conclusion it reinforced that copyright protection depends on the exercise of “skill and judgment” directed at the particular form of expression, not simply on the labour of compiling information.¹⁷ That emphasis on human skill and judgment, rather than mechanical output or sheer effort, has become the doctrinal spine running through every later Australian decision dealing with computer-generated material, and it is the exact test that will eventually be applied to a dispute over an AI-written song.

The clearest application to purely machine-generated content came in the Telstra Corporation Ltd v Phone Directories Co Pty Ltd litigation, where the Full Federal Court found that telephone directories produced largely through automated processes lacked the human authorship necessary to attract copyright at all.¹⁸ The reasoning is blunt but important: if a court cannot point to a human who exercised independent intellectual effort over the material form of a work, copyright may simply not exist in it, no matter how commercially valuable or superficially impressive the finished output appears.

Apply that logic to a song generated end to end by an AI tool, where someone types a prompt such as “write me an upbeat pop love song about a Gold Coast summer” and accepts whatever the model returns without meaningful further creative input. On the current state of Australian law, that song is unlikely to attract copyright protection at all.¹⁹ That is a genuinely counterintuitive outcome for most people. The absence of copyright does not mean the song mysteriously belongs to nobody in some poetic sense, it means there may be no exclusive legal right for anyone to enforce, no royalty stream that has to be shared by law, and critically, no protection stopping someone else from taking the exact same output and using it themselves.

The escape hatch, and it is a real and usable one, is genuine human authorship layered on top of the AI’s raw output. If a person selects, rejects, edits and rearranges what an AI tool produces, chooses the final chord progression from several AI-generated options, substantially rewrites the lyrics, or restructures the arrangement around a hook the AI merely suggested, that person is very likely exercising exactly the kind of independent intellectual effort Australian courts require. The more specific, deliberate and identifiable your creative decisions, the stronger your claim to authorship and to copyright in the finished work. Industry guidance from Arts Law reaches the same conclusion from a practitioner’s angle, noting plainly that AI tools “do not currently have a legal status and cannot own copyright,” and that it is the human contributor, and only a human contributor who has genuinely contributed independent intellectual effort, who can own whatever copyright exists in the result.²⁰

This mirrors how the United States Copyright Office has approached an identical problem from a different legal tradition: no registration for content generated by AI on its own, but registration remains available for works where a human has added sufficient original expression on top of AI-assisted material, provided the AI-generated content is disclosed.²¹ Two very different legal systems, arriving independently at the same functional answer, is itself a strong signal about where this law is heading globally, even before any legislature has formally acted.

Moral Rights: The Part Almost Everyone Forgets

Copyright ownership is one issue. Moral rights are a completely separate, and frequently overlooked, layer of protection that Australian songwriters need to understand, particularly once AI enters the picture.

Under Part IX of the Copyright Act 1968 (Cth), a human author of a copyright work has three distinct moral rights: the right to be attributed as the creator of the work, the right against having someone else falsely attributed as its creator, and the right against having the work subjected to derogatory treatment that harms the author’s honour or reputation.²² These rights exist independently of who owns the copyright itself, they cannot be sold or transferred, and critically, they attach only to a human author or performer.²³

This creates two practical traps for anyone working with AI tools. The first is straightforward: if you contributed genuine independent intellectual effort to an AI-assisted song, you are entitled to be credited as an author, and a collaborator who did not contribute a substantial creative effort is not entitled to claim that same credit, regardless of who ran the AI prompt or who owns the studio.²⁴ Split sheet disputes, discussed further below, are very often really moral rights disputes wearing a financial disguise, someone feeling written out of a credit line matters just as much, sometimes more, than a disagreement over percentage points.

The second trap is subtler and more troubling for the industry generally. Because moral rights only attach to human authors, an AI tool itself has no moral rights, and there is currently no legal obligation to disclose that AI was used to generate a work at all, unless a specific platform, chart body or awards body imposes that disclosure requirement contractually or as an eligibility condition.²⁵ In other words, Australian copyright law does not currently require you to tell your listeners, your collaborators, or even APRA AMCOS that a slice of your song came out of a machine. That gap is precisely why ARIA’s chart rules and the Grammys’ eligibility criteria, discussed below, have had to step in and impose disclosure and authorship standards that the underlying copyright statute does not itself demand.

APRA, AMCOS and Getting Paid When AI Is Involved

Copyright and royalties are related but genuinely separate questions. Even where a valid copyright exists, you are only paid for it if the correct collecting body has been mandated to licence it, and knows the song exists in the first place.

APRA AMCOS is actually two organisations operating jointly. APRA controls the performing right, which covers your song being performed publicly, broadcast, streamed or otherwise communicated to the public. AMCOS controls the reproduction right, which covers your song being copied, recorded or reproduced in any form.²⁶ In February 2025, APRA AMCOS wrote to writer and publisher members setting out its position on generative AI, confirming that APRA can licence the performing right on members’ behalf even where it is exercised through generative AI platforms, but flagging that AMCOS needed a separate, specific mandate before it could licence the reproduction of members’ songs for AI training purposes at all.²⁷ That mandate, described internally as covering “Generative AI Inputs,” was subsequently sought from AMCOS members, so that the society could licence AI developers to copy catalogue material into training datasets, with royalties intended to flow back to writers and publishers.²⁸

Crucially, that input licensing does not cover everything, and this is the detail songwriters most often miss. APRA AMCOS has been explicit that a mandate to licence a song being used as training input does not extend to a situation where a “substantial part” of that same song later turns up in what an AI system actually generates as output.²⁹ Being one of millions of songs an AI model was lawfully trained on is a completely different legal question from that same model later reproducing your song’s original expression when a stranger types a prompt. The “substantial part” concept driving that distinction has a long and colourful history in Australian courts, examined in detail below, and it is precisely the test that would apply if a dispute over AI-generated output were ever litigated here.

The overseas litigation landscape should be read as a live preview of exactly this fight arriving in Australia sooner or later. Round Hill’s claims against Suno and Anthropic allege the platforms copied lyrics and compositions from its catalogue to build commercial models without a licence.³⁰ The Munich court’s ruling against Suno went a step further, finding the AI’s own outputs, not merely its training inputs, reproduced existing GEMA-affiliated songs closely enough to infringe.³¹ If an equivalent case were ever brought in Australia, the reproduction of protected material in an AI output would almost certainly be assessed using the same substantial part test Australian courts have relied on for decades, meaning the qualitative creative essence taken matters far more than the raw quantity of material or how different the finished output sounds overall.

The History of “Substantial Part”: What Counts as Taking Someone’s Creative Work

Long before generative AI existed, Australian and international courts had to work out precisely how much of someone else’s song you can borrow before you have infringed their copyright. These older cases matter enormously now, because the same legal test, developed decades before anyone typed a prompt into a text box, will be the one applied to future generative AI disputes.

Kookaburra and Down Under: Australia’s own answer

The most instructive Australian authority is Larrikin Music Publishing Pty Ltd v EMI Songs Australia Pty Limited [2010] FCA 29.³² Men at Work’s 1979 and 1981 recordings of Down Under included a flute riff drawn from Kookaburra Sits in the Old Gum Tree, a four-bar children’s round written by Marion Sinclair in 1934, whose rights had passed to Larrikin Music. Justice Jacobson found that although the flute riff used only two bars of Kookaburra, and Down Under was an entirely different song in structure, genre, tempo and length, those two bars still amounted to a “substantial part” of Kookaburra, because the assessment of substantiality turns on the quality of what was taken, not the quantity.³³ The judgment states plainly that the “emphasis is on quality rather than quantity of what is taken,” and that the copied features must be a substantial part of the earlier copyright work, not of the later, allegedly infringing one.³⁴

EMI and the songwriters appealed. The Full Federal Court dismissed the appeal in EMI Songs Australia Pty Limited v Larrikin Music Publishing Pty Limited [2011] FCAFC 47, upholding the trial judge’s finding that Down Under reproduced a substantial part of Kookaburra.³⁵ The High Court refused special leave to appeal further, drawing the litigation to a close. Men at Work’s record label was ultimately ordered to pay Larrikin five per cent of royalties from Down Under, backdated to 2002.³⁶

The lesson for anyone assessing generative AI risk is direct and slightly unsettling. It does not matter that the finished song sounds completely different overall, uses a different tempo, genre or arrangement, or borrows only a handful of seconds of material. If what was taken represents the original, recognisable creative essence of an earlier work, that can be enough to infringe. An AI model that reproduces even a short but distinctive melodic or lyrical fragment from a song in its training data carries exactly this kind of risk, which is precisely the scenario APRA AMCOS deliberately carved out of its AI input licensing, and precisely the scenario the Munich court found against Suno.³⁷

My Sweet Lord: when copying happens without anyone meaning to

Kookaburra is not the only cautionary tale worth knowing, and one of the oldest and most famous examples in music copyright history is not even Australian. George Harrison’s 1970 hit My Sweet Lord was found by a United States court to have infringed The Chiffons’ earlier song “He’s So Fine,” despite Harrison’s own evidence that he had not consciously copied anything.³⁸ The court accepted that the similarity was almost certainly the product of what is often described as subconscious copying, Harrison had heard the earlier song many times and, without deliberate intent, reproduced its musical structure closely enough to infringe. This case matters for the AI conversation because it demolishes any assumption that infringement requires deliberate, knowing copying. If a human songwriter can infringe without meaning to, purely by having absorbed an earlier song into their own creative instincts, then an AI model, trained explicitly and systematically on enormous volumes of existing music, is arguably at even greater structural risk of reproducing protected material without its operator ever intending it, which is exactly the mechanism the Munich court’s “memorisation” finding against Suno described.

Blurred Lines and Stairway to Heaven: the international counterpoints

Two well known United States cases illustrate the limits on both sides of the substantial part test. In the “Blurred Lines” litigation, the estate of Marvin Gaye successfully argued that Pharrell Williams and Robin Thicke’s song infringed Gaye’s “Got to Give It Up.” The verdict was controversial among musicologists precisely because it was widely seen as resting heavily on the overall feel, groove and style of the two recordings rather than any specific, identifiable melodic or lyrical copy.³⁹ Copyright is not supposed to protect a genre, an era’s production style, or a general musical feel, only the particular original expression a songwriter actually created. The case is frequently cited as a cautionary example of the line between protectable expression and unprotectable style becoming dangerously blurred in practice, and it should worry anyone building an AI music case on “it sounds too much like” rather than on a specific, identifiable borrowed element.

Skidmore v Led Zeppelin went the other way. The Ninth Circuit ultimately upheld a jury finding that Led Zeppelin’s Stairway to Heaven did not infringe the earlier song “Taurus” by the band Spirit, partly because the descending chord progression at issue was a common building block of Western music, not something original enough to Spirit to be independently protectable.⁴⁰ Some musical elements, basic chord sequences, standard scales, common rhythmic patterns, are simply too fundamental to the shared musical language to belong to any one songwriter, no matter how memorably they were first used.

Read together, these three international cases and Australia’s own Kookaburra saga sketch a fairly complete map of the risk terrain. AI models are trained to reproduce statistical patterns across enormous numbers of songs, which makes them very good at capturing genre conventions and overall “feel,” the exact territory that caused controversy in the Blurred Lines case, while occasionally reproducing specific, original, protectable phrases from individual songs in their training data, the exact territory Kookaburra, My Sweet Lord and Stairway to Heaven were all fought over in different ways. A songwriter using AI tools needs to hold both risks in mind at once, and understand that neither one disappears simply because the finished track sounds original to a casual listener’s ear.

The Global Convergence: ARIA, the Grammys and a New International Standard

What makes this moment genuinely significant, rather than just another news cycle about AI, is that multiple independent industry bodies around the world have landed on almost identical rules within months of each other, without any single piece of legislation forcing their hand.

ARIA’s rule, effective from the chart dated 31 August 2026, asks whether a recording is “substantially human made,” treating it as ineligible where generative AI produced the entirety or the primary portion of its creative elements, even where a human later mixed or mastered the result.⁴¹ The Recording Academy, which awards the Grammys, adopted a strikingly similar standard back in 2023 and has since sharpened it considerably: “a work that contains no human authorship is not eligible in any category,” and where AI-generated elements are present, the human authorship component of the submitted work “must be meaningful,” a phrase the Academy’s own president has described bluntly as meaning AI-generated music, including full text-prompt-to-song generation on platforms like Suno or Udio, simply is not eligible on its own.⁴² The Latin Recording Academy followed with its own version in March 2026, specifying that the human contribution must be “significant and not merely minimal” and directly relevant to the category being entered, meaning a songwriting entry needs substantial human input in the music or lyrics, while a performance entry needs meaningful human engagement in the actual performance.⁴³

Notice what these three very different institutions, an Australian chart body, an American recording academy, and a Latin American recording academy, are all doing. Each has independently rejected the idea that a “the AI did the work but a human clicked publish” arrangement is good enough, and each has instead adopted some version of the same test Australian copyright law has applied since IceTV: genuine, identifiable, meaningful human creative contribution. The industry consensus is arriving faster than the legislation, and it is arriving from an unexpectedly consistent direction. Where the various rules diverge is in the detail rather than the principle. The Grammys explicitly permit AI-assisted pitch correction, mixing, mastering, sample organisation and drum pattern variation while barring an AI-generated melody that becomes the song’s primary hook, drawing a distinction between AI as a production tool and AI as the source of the song’s core creative content.⁴⁴ ARIA draws a similar but not identical line, focused specifically on the lead vocal and primary instrumental performances rather than production tools generally.⁴⁵ Anyone releasing music with any AI involvement, whether or not chart position or an award nomination is the goal, would do well to document their process against the strictest of these tests, since it costs nothing to be ready for scrutiny and can cost a great deal not to be.

Voice Cloning, Deepfakes and Australia’s Missing Right of Publicity

There is a further layer of risk that has nothing to do with copyright in a song’s melody or lyrics, and everything to do with the human voice itself. Generative AI tools can now convincingly clone a specific singer’s voice, and Australian law is, frankly, underprepared for the consequences.

Unlike the United States, where several states, most notably Tennessee’s so-called ELVIS Act, have created an explicit property right in a person’s voice and likeness,⁴⁶ Australia has no equivalent statute. The common law of Australia does not recognise an independent cause of action protecting how a person’s identity, including their name, image and voice, is used commercially.⁴⁷ If an AI tool generates an entirely new vocal performance that merely sounds like a well known Australian artist, without reproducing any pre-existing sound recording of that artist’s actual voice, Australian copyright law is unlikely to offer a way to stop it, because copyright infringement of a sound recording generally requires a facsimile copy of the original recording or a substantial part of it, not merely a stylistic resemblance.⁴⁸

That does not mean an artist whose voice has been cloned is entirely without a remedy, but the available tools are indirect and imperfect. The most likely avenues are the tort of passing off, and the misleading or deceptive conduct and false representation provisions of the Australian Consumer Law, found in Schedule 2 of the Competition and Consumer Act 2010 (Cth).⁴⁹ Both routes generally require the conduct to occur in trade or commerce, meaning purely non-commercial deepfakes or parody content sit in a genuine legal grey zone, and both require evidence that consumers were actually likely to be misled about the artist’s involvement, which is a real evidentiary hurdle when the entire point of a good voice clone is that it sounds convincingly authentic.⁵⁰ If, however, the AI-generated vocal was trained on or reproduces a pre-existing sound recording of the artist’s actual voice, the owner of the copyright in that original recording, often a record label rather than the artist personally, may have a more direct copyright infringement claim against whoever created or distributed the AI output.⁵¹

The practical upshot for songwriters and performing artists is twofold. If you are considering using an AI tool to generate a vocal performance in someone else’s style, understand that the absence of a clear Australian statute is not the same as the absence of legal risk, consumer law and passing off remedies exist precisely to catch conduct that copyright cannot reach, and courts have shown a willingness to use them creatively where a high-profile individual’s identity has been exploited without authorisation.⁵² If you are the artist whose voice might be cloned, the strongest protection currently available is often contractual, building clear terms into any recording, licensing or platform agreement that expressly restrict AI voice replication, rather than relying on a general legal right that Australia simply does not yet have.

What ARIA’s New Chart Rule Actually Means for You

If you are an artist hoping to chart in Australia, the practical rule from 31 August 2026 is fairly clear, though it carries real edge cases worth understanding in detail.⁵³ Your recording remains eligible if humans wrote the song and performed the lead vocal and the primary instruments, with generative AI used only in a genuinely supporting role, such as AI-assisted mastering, minor backing elements, or production tools that do not generate the song’s core creative content.⁵⁴ Your recording becomes ineligible if generative AI produced the lead vocal, a key instrumental performance, or the primary creative elements of the song, even where a human has since mixed, mastered or edited the result.⁵⁵ ARIA can decline to survey a recording, remove it from the charts prospectively or retrospectively, adjust chart positions, or revoke an ARIA number one award, though artists and their representatives can contest an exclusion decision and provide evidence supporting eligibility.⁵⁶ ARIA is applying its definitions consistently with a global labelling standard announced jointly by the IFPI and the Recording Academy on 10 July 2026, so the same substantially-human-made distinction is now being adopted across multiple markets simultaneously, not just introduced in Australia in isolation.⁵⁷

This has an immediate practical consequence for how you should actually be working, not just how you should think about the law after the fact. If you want a recording to remain chart eligible, document your own creative process as you go: which vocal takes and instrumental performances were genuinely human, which AI tools touched the production and in exactly what capacity, and be ready to describe that process clearly if ARIA, a label, or a rights holder ever asks. Treat this the same way you would treat any other piece of evidence you might one day need, because you will not remember these details as clearly in twelve months as you do today.

A Practical Checklist for Australian Songwriters

Bring all of this together into concrete, actionable steps.

1. Understand that copyright is automatic, but only for genuine human creative contribution. There is no registration system for copyright itself in Australia. The moment an original song is fixed in material form, whether written down, recorded, or saved as a digital file, copyright exists automatically, provided a human exercised genuine skill and judgment in creating it.⁵⁸ If a song was produced entirely by AI with no meaningful human input, that automatic protection may not apply at all, and you may have nothing enforceable to rely on.

2. Make your human creative contribution identifiable, and write it down as you go. If you use AI tools anywhere in the writing or production process, keep a running record of the specific creative decisions you made: which lyrics you kept, rejected or rewrote, which of several AI-generated melodic options you chose, how you restructured an arrangement around a suggestion. This record strengthens your copyright claim, your moral rights position, and your ARIA chart or awards eligibility all at once.

3. Register every individual song with APRA AMCOS through the Writer Portal. Registration does not create your copyright and does not decide who owns it, but it is what allows APRA to collect performance royalties and AMCOS to collect reproduction royalties on your behalf. An unregistered song generates no traceable income for you even if it is being played or streamed widely, no matter how strong your legal position might otherwise be.

4. Sign a split sheet on the day a co-written song is finished, and address AI contribution explicitly in it. Joint authorship arises automatically where two or more people contribute original creative material, in proportions reflecting their actual contribution, but almost nobody records this at the time. A one-page agreement recording percentages, contributions, and where relevant, which parts came from an AI tool and which specific human decisions shaped them, prevents the vast majority of later disputes, including the increasingly common disputes over whether an AI tool’s contribution should affect anyone’s credit or share.

5. Understand the “substantial part” test before you release anything that echoes an existing song. Whether a court is looking at Kookaburra and Down Under, He’s So Fine and My Sweet Lord, Taurus and Stairway to Heaven, or a future dispute over an AI-generated output, the question is the same: was the material taken original to the earlier work, and did it represent a qualitatively important part of that work’s own creative essence. Short passages can infringe. Generic chord progressions and stylistic conventions generally cannot. Subconscious or unintentional similarity is not a defence.

6. Check your current APRA AMCOS communications on AI licensing before assuming you are covered. The Generative AI Input mandate, AI training licensing arrangements, and the treatment of AI outputs reproducing a substantial part of a member’s song are all live, evolving positions, not settled rules. What is true this month may already be updated as litigation like the Round Hill and GEMA cases develops overseas.

7. If you intend to chart or seek an award, keep your production process substantially human and be ready to prove it. Human-written lyrics, a human lead vocal, and human-performed primary instruments are the baseline for ARIA chart eligibility from 31 August 2026 and for Grammy and Latin Grammy eligibility more generally. AI-assisted mixing, mastering, pitch correction or minor backing elements will not disqualify a track on their own, but the closer your use of AI sits to a song’s primary hook, melody or lyric, the more carefully you need to document your own creative role.

8. Treat voice cloning and soundalike vocals as a distinct legal risk, separate from copyright in the song itself. Australia has no general right of publicity protecting a person’s voice, so a genuinely new AI-generated vocal that merely resembles a known artist may sit outside copyright law entirely. If you are commissioning or using AI vocal tools in someone else’s style, understand that consumer law and passing off remedies may still apply, and get advice before release rather than after a complaint.

9. Get advice before signing anything with an AI music platform. Many generative AI music tools’ terms of service include broad licences over the content you generate, and Australian songwriters should read those terms as carefully as any recording or publishing contract, particularly given how quickly platform-level licensing and litigation positions are shifting overseas.

10. Revisit your process periodically, not just once. This is not a set-and-forget compliance exercise. APRA AMCOS’s own AI position, ARIA’s chart code, and the awards bodies’ eligibility rules have all changed materially within the past eighteen months, and there is no reason to think that pace will slow down. Build a habit of checking your assumptions against current guidance before each new release, rather than relying on advice that felt current a year ago.

Where This Is Heading

Nothing about this area is settled, and anyone who tells you otherwise is guessing. The Copyright Act 1968 (Cth) has not been amended to deal with AI-generated works directly, moral rights legislation still assumes every author is a person who can feel pride or embarrassment about their credit, APRA AMCOS is actively renegotiating its own mandates with members, ARIA’s new chart rule is barely a week old at the time of writing, and courts in Germany and the United States are only just starting to hand down substantive rulings on what AI music platforms actually owe the songwriters whose catalogues trained them.⁵⁹ Recent academic commentary on Australian copyright and generative AI has gone so far as to describe the current legal position as an “authorless” gap in the statute book, a category of creative output that exists in the world and generates real commercial value, but that the law was never designed to recognise, own or protect.⁶⁰

Expect further movement through 2026 and 2027 as the Round Hill litigation against Suno and Anthropic proceeds, as other rights holders test similar claims following the Munich court’s finding against Suno, as APRA AMCOS finalises how AI licensing revenue, if any, actually reaches individual writers rather than disappearing into an undistributed pool, and as Australian regulators eventually turn their attention to the voice cloning and deepfake gap that consumer law and passing off can only partially fill. The convergence between ARIA’s chart rules and the Grammy and Latin Grammy eligibility standards suggests the industry itself may end up setting the practical rules of the road faster than any parliament, simply because artists, labels and audiences all want the same basic reassurance: that when they are told a human wrote a song, a human actually did.

Until the law catches up properly, and it will take longer than anyone currently expects, the safest position for any Australian songwriter is the same one courts have applied to human creativity for well over a century: make sure a person is making, and can point to, the identifiable creative decisions in a song, document that involvement as you go, protect your moral rights and your voice through clear written agreements, and register properly with APRA AMCOS so you are actually in a position to be paid once royalties start moving. The technology will keep changing. The underlying legal test, real, provable, human creative effort, almost certainly will not.

Download the practical guides

One Page Practical Guide

Use the AI Music Copyright and APRA AMCOS One Page Practical Guide before you release, pitch, broadcast, submit or monetise an AI-assisted track.

Full guide PDF

Download the full AI-Written Music, Copyright and APRA AMCOS guide for a print-ready version of this guide.

This guide provides general legal information only; it is not personal legal advice. Everyone’s situation is different, so please seek independent legal advice for your own circumstances. Contact Bell and Senior Lawyers on (07) 5532 8777 or visit bellsenior.com.au .

See Also

• Legal Matters Ep 29 - AI Music, Copyright and APRA AMCOS: Your Legal Questions Answered
• Legal Matters: AI Agent Liability, Smart Glasses Privacy and Joint Tenancy
• Legal Matters: AI Law
• Legal Matters: Technology Law Part 2 - Consumer Law and AI