AI and music now meet at every stage of the industry: software generates, classifies and recommends tracks, and it is changing how music is composed, produced, marketed and consumed. AI tools have accelerated creative workflows and help listeners discover new artists. They also put creator income at risk, flood streaming catalogues with machine-made uploads and have set off copyright suits over training data. Below we look at what the technology does, who gains and who loses, and why the rights question is moving from courtrooms to licensing deals.
How AI is affecting the music industry, stage by stage
Recent reviews of music AI cover generation alongside editing, captioning, production, transcription, source separation, live performance, discovery and education.
One clarification before going further: generative AI is broader than large language models. The term originally referred to models that generate signals such as audio and images, and treating “generative AI” as a synonym for large language models is a market-driven confusion. Music is a good example of the broader meaning, because its systems generate symbolic scores and audio signals directly. Our team at TechnoLynx developed custom generative models well before Midjourney existed. Our work on AI engineering for media, broadcast and telecom shows how we apply generative audio models across these industries.
Is AI-generated music real music, and how do the models make it?
If “real” means sound listeners struggle to tell apart from human-made music, a recent survey points to yes. In a Deezer and Ipsos survey of 9,000 people across eight countries, published in November 2025, 97% of respondents failed to tell which of three tracks were fully AI-generated. If “real” means authored, US copyright practice draws a line. The U.S. Copyright Office will not register “works that lack human authorship” or works produced by a mechanical process with no creative input from a human author.
Symbolic and audio generation
AI music generation splits into two main directions. Symbolic generation writes notes, producing MIDI files, sheet music or piano rolls, and learns chord progressions, melodies and rhythmic patterns. Audio generation produces the waveform or spectrogram directly, which puts it closer to recording and mixing and lets it reproduce complex timbres. The two fail in different ways: symbolic models often lack expressive timbre, while audio models struggle with long-range structure, which is why hybrid frameworks combining both have appeared.
The architectures doing the work
Three families of models stand out in recent reviews. GANs pit a generator against a discriminator, transformers use self-attention to capture motifs and repetition across time scales, and diffusion models build high-fidelity audio by gradually removing noise. Two published systems show how this works in practice. Meta’s MusicGen is a single-stage transformer language model over several streams of compressed audio tokens, conditioned on a text description or a melody. Google’s MusicLM treats text-to-music as hierarchical sequence-to-sequence modelling and generates 24 kHz audio that stays consistent over several minutes.
Why training data decides what comes out
A generative model takes in large amounts of training data and produces new items that are statistically likely to exist in that data. The data needed is large: the U.S. Copyright Office report quotes the licensing firm Rightsify saying that commercially viable models resembling professional music need datasets of at least a million songs. A composition review also notes that deep learning models are always likely to produce music similar to their inputs, sometimes with patterns taken directly from them.
Quality has limits too. As generated pieces get longer, their musicality deteriorates into chaotic structure or heavy repetition, and GAN and RNN models still show “mode collapse”, where samples sound too alike. Our reading is that today’s systems are very good at plausible surface and still weak at the large-scale form a composer plans deliberately. That is one more reason the “is it real” debate ends up being about authorship rather than sound.
When was AI first used in music?
Computers began composing music in the mid-1950s. The most cited example is Hiller and Isaacson’s Illiac Suite, generated late in 1956 with rule systems and Markov chains. One review dates it to 1957 and credits the melody of its fourth movement as the first composition made entirely by computer using a Markov chain. In the early 1990s, David Cope’s Experiments in Musical Intelligence (EMI) recombined patterns from existing works to imitate the styles of hundreds of composers. At a musical Turing test organised by Douglas Hofstadter, the audience picked EMI’s piece as the real Bach. The field then moved from rule-based methods to the deep learning systems now in common use.
How is AI used in music discovery, editing and rights management?
Outside generation, AI also helps listeners find music, helps engineers edit audio and helps platforms identify AI-made tracks.
Discovery. Commercial recommenders such as Spotify and YouTube Music rely on collaborative, content-based and hybrid filtering, using graph-based and matrix factorisation methods. Collaborative filtering cannot recommend a new song that has no listening data, which is the cold-start problem. Spotify research published in 2014 addressed it by training a convolutional network on spectrograms of three-second audio fragments to predict where a song would sit in the collaborative filtering space. More recently, Spotify Research reported that LLM-generated explanations for recommendations made users up to four times more likely to click in some cases, especially for niche content. Its AI DJ, launched in 2023, adds personalised commentary to song selections.
Editing and production. Neural source separation tools, from open-source Spleeter to commercial services such as LALAL.AI, split mixes into stems, although separating distinct singing voices or synthesisers is still difficult. Text-controlled production models have emerged too: SonicMaster responds to commands such as “bring out the bass” for restoration and mastering. Musicians work inside digital audio workstations through VST plugins written in C++, while most models are written in Python, and few attempts have been made to bridge the two. We cover the production and vocal side in more detail in AI in music, audio and sound production.
Rights management. Deezer says its detector identifies fully AI-generated tracks from models such as Suno and Udio, and claims 99.8% accuracy with fewer than one false positive per 10,000 human-made tracks. Researchers propose fingerprinting training datasets against databases such as MusicBrainz and AcoustID to make clear what a model learned from. Deep learning watermarking, first built for speech, is being adapted for music. On the disclosure side, Spotify reports tens of thousands of AI credits submitted daily through its beta disclosure feature.
How much income could creators lose, and how many uploads are AI-made?
Three estimates put the income risk at around a quarter of revenues, all of them projections rather than measured losses:
| Source | Scope | Projection |
|---|---|---|
| CISAC and PMP Strategy (2024) | Global music creators | 24% of revenues at risk in 2028, an annual loss of €4bn |
| UNESCO monitoring report (2026) | More than 120 countries | Music creators’ revenues could fall by 24% by 2028 |
| APRA AMCOS / Goldmedia survey | Over 4,200 members, Australia, New Zealand and overseas | 23% of revenues at risk, AUD 519 million cumulative |
CISAC describes two ways creators lose: their works are used for training without payment, and AI outputs then compete with human-made music. Creators themselves are divided. In the APRA AMCOS survey, 38% already use AI in their work and 54% agree it can assist the creative process, yet 82% worry they will no longer be able to make a living.
Deezer reported about 10,000 fully AI-generated uploads a day in January 2025, 18% of daily uploads by June 2025, about 50,000 a day (34% of deliveries) by November 2025, and 90,000 a day, over half of new uploads, at a June 2026 peak. Deezer says these tracks are uploaded mainly for streaming fraud. In June 2025 Deezer found up to 70% of streams on fully AI-generated tracks were fraudulent, even though these tracks drew only about 0.5% of streams. Its July 2026 update put the 2025 figure at up to 85%. Spotify, for its part, says it removed over 75 million spammy tracks in the 12 months before its September 2025 announcement.
A systematic review of 171 generative audio papers found that 65% discussed positive impacts but only 16 papers (under 10%) discussed any negative impact, such as fraud, deepfakes or copyright infringement.
Platform responses: ban, tag, demonetise or disclose
Bandcamp is the only major on-demand service that bans AI music outright. The big services police impersonation, streaming fraud and spam uploads, and regulate the middle ground in different ways.
| Platform | Policy stance | What it does |
|---|---|---|
| Bandcamp | Ban | Prohibits music generated wholly or in substantial part by AI, plus AI impersonation of artists or styles; enforcement relies mostly on community reports |
| Deezer | Tag and exclude | Detects and labels fully AI tracks, keeps them out of algorithmic and editorial recommendations, deletes those used for fraud or unplayed for six months |
| Tidal | Badge and demonetise | Since July 2026 badges fully AI tracks and pays no royalties on them; refuses use of its catalogue for AI training |
| Spotify | Disclose | Allows vocal impersonation only with the artist’s authorisation, filters spam, supports DDEX metadata so distributors can disclose AI use |
| Apple Music | Optional tags | Transparency Tags from March 2026 flag AI use in artwork, track, composition and video; untagged releases are assumed human-made |
Spotify’s position rejects the binary of “AI” versus “not AI”. The company argues that AI use is a spectrum, and the DDEX standard lets rights holders indicate whether AI touched vocals, instrumentation or post-production. We think the spectrum view is the technically honest one. Deezer’s detector is built for fully generated tracks. The harder cases, a human song with an AI-assisted mix or a cloned backing vocal, will in our view depend on provenance metadata rather than classifiers.
What ethical and consent questions does training AI on artists’ music raise?
A central dispute is whether a company may train on recorded music without permission. The music industry says no: training on recordings, songs or anything containing an artist’s voice or likeness requires a licence. AI developers rely on fair use in the United States and on the text and data mining (TDM) exception in Europe. One legal analysis argues the TDM argument fails because generative training aims to produce outputs that resemble the training data, not to extract insights from it. The U.S. Copyright Office’s report on AI training notes dozens of pending US lawsuits on how fair use applies.
Memorisation turns the dispute into a concrete legal risk. Generative models have been shown to memorise portions of their training data. The RIAA’s suits against Suno and Udio, brought on behalf of Universal, Sony and Warner, allege mass copying of sound recordings. They also cite examples where Suno’s outputs closely resembled copyrighted songs. Suno’s chief executive denied that the service memorises existing music; the RIAA says the evidence shows otherwise.
Consent also covers artists’ voices and likenesses, not only catalogue. In 2023, “Heart on My Sleeve” cloned the voices of Drake and The Weeknd from vocal-only tracks. Speech models of the kind behind voice cloning raise fraud and biometric consent concerns. The technical side of those voice models is covered in our piece on real-time generative text-to-speech. Creators want transparency first: 97% of APRA AMCOS respondents say AI providers should disclose when they train on copyrighted works, and 95% want permission to be required. The ethics review adds quieter harms, including Western bias in training data and the use of music from under-represented cultures without engagement from those communities.
From lawsuits to licences
The industry is now shifting from litigation to licensing. Warner Music Group settled with Suno. Universal settled with Udio and agreed a licensed platform built with fingerprinting and filtering. By September 2026 Suno’s v6 models were developed in partnership with Warner, BMG and Believe. Licensing does not settle the question for performers, though. In June 2026 the American Federation of Musicians sued Universal and Warner for licensing members’ recordings to Suno and Udio without compensation or credit.
Weighing it up: who gains, who carries the cost
Our judgement, with the caveat that the income figures are projections and the legal questions are still open: AI’s benefits in music go mainly to listeners, platforms and tool-using producers, while the costs fall mainly on working creators whose recordings trained the models and whose streams now compete with machine output. Licensing deals shift some of that value back to the rights holders that signed them. That is why the union suit matters as much as the label settlements.
Attribution tools such as melody and timbre similarity detection, and memorisation safeguards such as output filtering, are still unfinished research with known limits. The open question for the next few years is whether attribution becomes reliable enough for licensing to pay individual creators, rather than only the catalogues big enough to sue.
