There is now a machine that will listen to a song and tell you, with a number, how much the world is going to like it. Researchers at the AMAAI Lab at Singapore University of Technology and Design built it. It's called APEX, and it was trained on more than 211,000 AI-generated songs, roughly 10,000 hours of audio pulled from Suno and Udio, to jointly predict two things: raw popularity, meaning streams and likes, and five separate dimensions of perceived aesthetic quality, pulled from frozen audio embeddings generated by a model called MERT. The uncomfortable finding buried in the paper is that the machine is far better at judging whether a song sounds good than at judging whether people will actually love it, and it does noticeably better on instrumental tracks than on anything with vocals, where taste gets messier and more human. In other words: AI can approximate craft. Chemistry with an audience is still slipperier. To be clear about what APEX actually is: it is not a consumer app and there is no website where a casual listener can drop in a song and get a score back. The AMAAI Lab released it as free, open source research code on GitHub and as a downloadable model on Hugging Face, the online hub where developers share machine learning models, under a permissive license that lets anyone use, copy, or build on it. Testing it requires some technical comfort. A user needs Python installed, along with a handful of machine learning libraries such as PyTorch, then downloads the pretrained model and points it at an audio file on their own computer. Feed it an MP3 and it returns a set of numbers: a predicted streams score, a predicted likes score, and five separate aesthetic ratings covering qualities like coherence, musicality, memorability, clarity, and naturalness. There is no button-and-slider interface built for musicians or fans, in other words, this is a tool made for other researchers and engineers to test and build on, not something a songwriter can open in a browser tab and use over coffee. That gap matters because the industry racing fastest toward answering it is Suno, the Cambridge, Massachusetts AI music company that has turned "type a prompt, get a song" into a business generating an estimated 7 million tracks a day. Suno already has a tool built for exactly the instinct to take something loved and build something adjacent to it. Its upload-and-cover feature lets a user drop in an existing audio file, one they're required to confirm they own the rights to, and rewrite it in a new style while the underlying melody survives the transformation. Feed it a stripped-down piano ballad and ask for djent, or a folk tune and ask for Afrobeat, and the skeleton of the original holds while the skin changes completely. It is, functionally, a machine for zeroing in on whatever made a song click and rebuilding a new one around that same spine. Suno just wrapped that capability, and a lot more, into v6, a family of three new models launched on September 9. The flagship v6, an experimental v6-wild, and a free v6-mini are the company's first models trained from scratch on catalog it actually licensed, from Warner Music Group, BMG, and independent distributor Believe. Warner signed its deal back on November 25, 2025, ending the copyright suit it had brought against Suno. BMG followed on August 12, 2026, covering both recordings and publishing. Believe and its TuneCore arm signed a day before launch. Suno's chief product officer told Bloomberg the company will now share a cut of revenue with those partners starting from launch day, with labels responsible for splitting that money among their own artists and songwriters going forward. Co-founder and chief executive Mikey Shulman framed it as vindication after eighteen months of being treated as the industry's most convenient villain. Suno has weathered a hacker leak alleging its models trained on scraped audio, an admission that it pulled data from YouTube, and a banner-towing plane reading "Say No To Suno" flown over an AI investor summit where Shulman was speaking. The company has also just raised $400 million at a $5.4 billion valuation, and by earlier this year had been used by more than 100 million people, with 2 million paying subscribers. None of that goodwill extends to Universal Music Group or Sony Music, who never signed licensing deals and instead filed a second lawsuit against Suno on September 18, a 45-page complaint in federal court in Boston that adds more than 60,000 specific recordings to their claims. Their argument is that training a supposedly clean new model on the outputs of an old, allegedly infringing one doesn't wash the infringement away, memorably calling v6 "the fruit of the same poisoned tree." Under US copyright law's cap of $150,000 per work for willful infringement, the theoretical maximum on the table now tops $9 billion. Suno also lost, decisively, in a Munich courtroom in July. Germany's music rights society GEMA sued over six specific compositions, including well-known tracks like Rasputin and Mambo No. 5, and the Munich District Court rejected Suno's fair use defense outright, ruling the company had actively operated the model rather than merely hosting neutral infrastructure, and ordering it to disclose the scale of its use and pay damages. The court's reasoning leaned on a blunt distinction: it may be fair use to copy from a book to write a review, but not to steal the book in order to make the copy. Put those threads together and the picture is of an industry racing in two directions at once. On one side, machine learning that can now score a track's aesthetic polish and predict, with real but limited accuracy, how an audience might respond. On the other, a tool that lets anyone take the song that scored well, or simply the song they personally love, and generate a close cousin of it in seconds. The technology for identifying what's loved and the technology for replicating it have arrived in the same product, at the same company, in the same month that two of the three biggest labels on earth decided to fight it in court rather than cash a check.
🌍 world
Suno Can Clone Your Favorite Song. Everyone's Suing.
AI researchers have built systems that score a song's global appeal before a single human hears it, and Suno's newest tools let anyone upload a track they love and get something eerily close back. The catch: Universal and Sony just filed a fresh lawsuit calling the whole enterprise the fruit of a poisoned tree.
Fact checked - 15 claims 25 Sept 2026 · 13 with sources
My Take
AI hasn't cracked taste. It's cracked proxies for taste, streams, likes, engagement curves, and it's gotten genuinely good at predicting those numbers. But a stream count is not the same thing as love, and the APEX research itself admits the gap: the model is far more confident scoring whether a track sounds polished than whether a crowd will actually fall for it, especially once a human voice is involved. That's the tell. Taste is still contextual, cultural, tied to a moment nobody can fully quantify, and every attempt to reduce it to an embedding is measuring the echo, not the source. Suno's upload-and-cover tool is the more honest version of the whole conversation, because it skips the pretense of originality entirely. It says: here is a song you love, tell us what to change, and we'll hand you something built on its bones. That's a legitimately useful creative tool for a hobbyist reshaping their own demo. It's also precisely the capability that turns a licensing deal with three labels into a lawsuit from the other two, because the line between "inspired by" and "built from the wreckage of" gets thinner every time the underlying model was trained on material nobody licensed in the first place. The Munich court put it more plainly than any tech company ever will: transforming nothing while memorizing everything isn't transformation, it's laundering. Suno's licensing deals with Warner, BMG, and Believe are a real, meaningful shift toward legitimacy, and Shulman deserves credit for buying his way into the tent rather than fighting every label at once. But a $9 billion lawsuit sitting on the desk in Boston is not background noise. It's the bill for everything that got trained before anyone asked permission.
What Happens Next
The Boston federal lawsuit from Universal Music Group and Sony Music is the case to watch. It builds directly on an existing joint suit the two labels filed with Warner back in 2024, and now argues that Suno's licensed-sounding v6 models are legally contaminated by the older, unlicensed models they were partly trained on. Expect discovery fights over exactly how v6 was built, and expect Suno to lean hard on its licensing deals with Warner, BMG, and Believe as evidence of good faith going forward, even as it fights to keep the earlier training pipeline out of the spotlight. In Germany, the GEMA ruling is likely headed for appeal, following the same path as GEMA's earlier win against OpenAI over song lyrics, which is also under appeal. If the ruling holds, it sets a European precedent that AI companies can't shelter behind a content-neutral platform defense once they're actively operating the model and generating outputs for users, a standard that would complicate Suno's international expansion well beyond Germany. On the product side, Suno has said "opt-in" experiences built around individual named artists are next on its roadmap, letting specific musicians license their voice or style directly and get paid when fans use it, though no firm timetable has been announced. Whether any big-name artist actually opts in, given the current legal climate and the Human Artistry Campaign's aggressive public campaign against the company, will be the real test of whether Suno's truce with parts of the industry can widen into something the rest of the business is willing to join.
What History Tells Us
The GEMA rulings against both OpenAI and Suno echo an older pattern in the music business: technology arrives faster than licensing, gets sued into a settlement, and eventually the settlement becomes the industry standard. Napster became the blueprint for Spotify's licensed streaming model two decades ago, and YouTube's Content ID system emerged from exactly this kind of legal pressure. Suno's licensing pact with Warner, BMG, and Believe looks like an early version of that same arc, just with two of the three biggest labels on earth still refusing to sign.
Market Impact
Suno is privately held, so there's no ticker to watch, but its $5.4 billion valuation from the June funding round is now a live bet against roughly $9 billion in theoretical exposure from the new Universal and Sony lawsuit alone, on top of the damages already ordered in the GEMA case in Germany. For publicly traded Warner Music Group, the Suno partnership is a new, closely watched revenue line that investors will scrutinize for actual dollars once quarterly results start reflecting it, rather than treating it purely as reputational cover.