MBW Views is a series of op-eds from eminent music industry people… with something to say. The following MBW op/ed comes from Deviate Digital founder Sammy Andrews.
I’ve been posting a lot of genuinely impressive AI cover versions this past year on socials, much to the dismay of many. It seems a large part of the industry still insists AI music is a gimmick, an uncontrollable threat, or something to dismiss until it has calmed down. It hasn’t and it won’t.
Fans are consuming and making this material in serious volumes, and the people losing out are the ones pretending the shift isn’t happening or failing to build the structures needed to protect human artistry and monetise the behaviour.
AI-generated and AI-modified tracks now exist at industrial scale. Deezer ingests around 50,000 AI-made tracks a day, roughly a third of all new uploads. Most are quarantined and excluded, but that’s a technical filter, not an audience signal.
Research shows most people cannot reliably distinguish AI-generated tracks from human ones. And Deezer is small compared with the largest DSPs. If it’s seeing 50,000 AI uploads daily, services many times its size are handling far more. This isn’t gradual adoption. It’s a mass arrival.
The signs have been obvious for a while. In 2024, an AI act reached close to a million monthly listeners on Spotify before anyone realised it wasn’t a human artist. More recently, BBC Introducing presenters praised AI-generated songs believing they were new acts.
“If trained professionals can’t reliably detect AI, everyday listeners won’t either.”
If trained professionals can’t reliably detect AI, everyday listeners won’t either. The behaviour is stable and repeatable. Millions search for and share AI covers and remixes daily. That consistency is the basis of every revenue line the industry has ever built. What’s missing is licensed infrastructure.
China (specifically Tencent) made some fairly early and bold moves in terms of embracing AI capabilities widely, but Western markets are moving slower. At time of writing two major labels have now settled with a leading AI generator, shifting the conversation from litigation to licensing and opening the door to a commercial AI environment in 2026 (and I half suspect with some equity thrown in).
Progress is happening, just painfully slowly. Meanwhile, fan behaviour keeps accelerating.
The biggest short-term commercial opportunity is AI cover versions and remixes. They exist everywhere: TikTok, YouTube, Reddit, Meta, Discord, X, private tools and emerging generative platforms. Fans aren’t waiting for permission. Quality varies for sure, but the behaviour is entrenched. This is a clear revenue line.
The basic rules for covers are simple: If you don’t use the original master, you’re making a cover; license the composition, and 100 per cent of the publishing revenue flows to the songwriters and publishers; AI-rendered performances that create a genuine re-record without using the original master fall under this category; treat them exactly as conventional covers.
The complication is the huge volume of AI ‘covers’ that borrow from the original recording: stems, trained vocal models based on the master, or recognisable sonic elements.
At that point it’s no longer a cover. It belongs in the same world as remixes and sampling, where permission from the master owner and performers is required and splits reflect reliance on the original.
If a transformation uses a master in any meaningful way, the original rights holders and performers should receive the majority of the master revenue. That’s not a new idea; it’s the logic that already governs remixes, interpolations and samples but seems completely missing from any AI doom conversation you’re likely to be having.
Some argue this can’t scale, but it already does. Remixing and sampling have operated with negotiated approvals and splits for decades. The friction comes from AI covers that imitate originals closely enough to appear ‘new’, while sidestepping permission and compensation. That’s why a consent layer for master-adjacent uses is essential.
When a transformation is trained on or derived from an original master, it should require approval and deliver most of the master income to the original performers and rights-holders.
A notice-and-takedown mechanism is a necessary backstop. If a track demonstrably uses or derives from a master without consent – including models trained on that master – and the artist objects, DSPs should be able to immediately block, demonetise or reroute revenue until cleared.
This isn’t about limiting fan creativity. It’s about ensuring creators, performers and rights-holders are paid. Let lawful covers flow with full publishing revenue. Require consent and majority master shares when original elements are used. Remove or escrow the rest until licensed.
None of this ignores the depth of feeling among creators. Some artists simply dislike others reinterpreting their work. Prince famously hated covers, yet Nothing Compares 2 U became one of his most successful compositions because a cover carried the song.
Artistic taste and commercial frameworks aren’t mutually exclusive; this can be navigated and we can and should honour artists’ concerns.
“Pretending AI covers will disappear only pushes the behaviour into grey zones where fraud thrives.”
A clear conceptual split will save endless circular arguments. AI works from large language or music models that generate ‘in the style of’ should be treated like conventional covers for composition licensing, with strong guardrails against misleading attribution and abusive voice likeness.
Mash-ups, remixes and other transformations derived from identifiable masters should continue to require permission and attract majority master shares.
Once AI covers are licensed properly at scale, the economic upside is immediate. Publishers and songwriters win first because every AI cover depends on the composition and many of these covers will also drive streams of the originals (evident in a million Reddit threads and Discord server chats already). One hit could spawn hundreds of licensed versions, all paying the original writers.
Labels and artists benefit when they license stems, masters or voice models into controlled environments. Catalogue value rises as classic recordings are re-rendered for contemporary tastes. DSPs and AI generating platforms benefit because legal transformations can become premium features that attract younger listeners and deepen engagement.
A major DSP or AI platform will almost certainly launch a legal transformation layer soon. Udio is touting it in its plans and Spotify’s AI DJ already mediates between listeners and catalogue – all of the recent licensing settlements point directly towards permission-based AI music environments.
Combine these elements and you get personal, on-demand, licensed reinterpretations of any song, with automated payments to songwriters, publishers, labels and performers.
There will be losers if this is mishandled: mass unlabelled and unregulated slop to wade through, artists cloned without consent, session players displaced, composers undercut, and smaller acts drowned in synthetic upload volume unless separated and supported properly.
Pretending AI covers will disappear only pushes the behaviour into grey zones where fraud thrives. The practical choice is to recognise what audiences are doing, license it where lawful, label and categorise it appropriately, require consent where needed, and ensure money flows to the people whose work is being used.
Right now, most of this earns nothing. The demand exists. The tools exist. The audience has already moved. What’s missing is clarity, consent and compensation. It’s the industry applying what it already knows. It’s time we stopped circling the issue and started collecting the money.
MBUK is available as part of a MBW+ subscription – details through here.
All physical subscribers will receive a complimentary digital edition with each issue.Music Business Worldwide
