What Is Really Inside AI-Generated Music: Risks, Frequencies and Facts
AI-generated music sounds more and more real, but listeners do not know where it comes from, what it was built with or who is responsible for it. This page explains simply how it works, which risks are real and documented, and which fears have no evidence behind them. To truly warn people, we have to be precise.
How an AI song is made
Programs such as Suno or Udio create a complete song, with voice, instruments and lyrics, from a few words typed by the user. To learn how, they were "trained" on huge amounts of existing music. Outside these companies, nobody knows exactly which recordings were used. The major record labels took them to court for exactly this reason, accusing them of using their recordings without permission.
The result is music that imitates human music without being human. There is no composer choosing each note, and no singer giving meaning to each word. There is a statistical calculation producing what "usually" comes next.
How widespread it is
The numbers are growing fast. On the Deezer platform, fully AI-generated tracks were about 10% of new uploads in January 2025. By June 2026, on peak days, they passed 50%: around 90,000 tracks a day (figure checked October 2026). They currently get only 1–3% of plays, but according to Deezer up to 85% of the plays of these tracks in 2025 were fraudulent, that is, generated by bots to collect royalties.
And telling them apart is almost impossible. In an Ipsos survey for Deezer of 9,000 people in eight countries (November 2025), 97% failed to recognise the AI-made tracks. More than half said this made them uncomfortable.
Did you know? In summer 2025 the Velvet Sundown, a 1960s-style "band" with over a million monthly listeners on Spotify, turned out to be entirely AI-generated: music, lyrics, photos and band members.
Sound affects the brain: what science says
Sound is not neutral. It has been shown that certain frequencies can affect the brain and the emotions:
- 40 Hz and the hippocampus. In 2019 an MIT team published a mouse study in Cell: one hour a day of 40 Hz tones for a week reduced the plaques typical of Alzheimer's in the hippocampus, the brain's memory centre, and improved test results. This is deliberate stimulation for treatment, and so far it concerns animals.
- Infrasound. These are sounds below 20 Hz, too low for the ear to hear as notes. In 2003, at an experimental concert in London, some pieces were played with infrasound at about 17 Hz. The audience did not know which: 22% reported more unusual sensations, such as unease, sorrow, chills or fear, when the infrasound was present.
- Binaural beats. Two slightly different tones in each ear are supposed to "tune" brain waves. Studies disagree: in a review of 14 EEG studies, 5 supported the effect, 8 contradicted it and 1 was mixed.
Does AI music contain subliminal frequencies?
To date no reliable source has documented AI music generators inserting hidden frequencies, deliberately or by accident, to act on the brain. Experiments like the London one need special equipment and high volume: phones, earbuds and ordinary speakers can barely reproduce sounds below 20 Hz.
The fear of hidden messages in music is not new. In 1990, in the United States, the band Judas Priest was taken to court, accused of hiding phrases in a song that allegedly drove two young men to suicide. The judge dismissed the case: there was no evidence of subliminal messages.
The real risk is a different one, and nobody can deny it: people listening to a generated song have no control over how it was built. They do not know which recordings the program learned from, what the lyrics really say, or who decided what. Music made by a person has an author who answers for it. Generated music does not.
Lyrics without discernment and stolen voices
Generators have filters against some content, but no filter can judge whether a text is true, respectful or fit for prayer. A generated track can mix sacred words and irreverent images without anyone noticing. We look at this in Sacred music, blasphemy and AI.
AI can also clone the voice of a real singer. In 2023 the track "Heart on My Sleeve", with imitated voices of Drake and The Weeknd, was played millions of times before being removed at their record label's request. Neither of them had ever sung it.
How to protect yourself
- Choose artists with a name, a story and live concerts you can verify.
- Be wary of "bands" that release dozens of tracks a month with no real photos and no concert dates.
- Look for labels: some platforms, such as Deezer, flag fully generated tracks. From 2 August 2026 European law requires "deepfake" content created or manipulated with AI to be declared.
- For ceremonies, churches and important moments, always ask for music performed by people.
This guide is curated by Maestro Ilarion, a tenor and pianist who sings live in churches all over Sicily. If you are planning a ceremony and want real music, sung by a person, see church wedding musicians and church funeral musicians.
The other pages in this guide
- AI music or human music: why stay human
- The human cost: artists, jobs and rights
- Live music and the sounds of nature
- Sacred music, blasphemy and AI
Sources
- Deezer Newsroom, 21 July 2026: AI music exceeds 50% of daily uploads
- Deezer/Ipsos, November 2025: 97% cannot tell AI music from human music
- The Fader, 7 July 2025: the Velvet Sundown story
- MIT, 2019: 40 Hz stimulation and the hippocampus (study published in Cell)
- NBC News, 2003: infrasound linked to "spooky" effects at a London concert
- Popular Science: do binaural beats really help you focus?
- Rolling Stone: the Judas Priest subliminal message trial (1990)
- Whynow: "Heart on My Sleeve", the fake Drake and The Weeknd track removed in 2023
- Forbes, 18 December 2025: the Suno and Udio licensing deals
- EU AI Act, Article 50: transparency obligations (from 2 August 2026)
What this story teaches
There is no need to fear everything: what we need is to know. Sound really does affect us, and that is exactly why it matters to know who made it, how and why.
Frequently Asked Questions
What is AI-generated music?
It is music created by a program, such as Suno or Udio, from a few words typed by the user. The program produces voice, instruments and lyrics after being trained on huge amounts of existing music.
How can I tell if a song was made with AI?
Often you cannot: in a 2025 Ipsos survey 97% of people failed. Some clues help: artists with no real photos or concerts, dozens of tracks released in a short time, generic lyrics. Some platforms, such as Deezer, flag fully generated tracks.
How much AI music is on streaming platforms today?
On Deezer, fully generated tracks passed 50% of new uploads on peak days in June 2026, about 90,000 a day. They get only 1–3% of plays, however (figure checked October 2026).
Can sound affect the brain and the hippocampus?
Yes. An MIT study published in Cell in 2019 showed that 40 Hz tones, heard for an hour a day, reduced Alzheimer's plaques in the hippocampus of mice and improved memory. It is serious research, on animals and for treatment.
What is infrasound and what effect does it have?
It is sound below 20 Hz, which we do not hear as notes. In a 2003 experiment in London, 22% of the audience reported unease, sorrow or chills when infrasound was added to the music, without knowing when this happened.
Does AI music contain subliminal frequencies?
To date no reliable source has documented this. The real problem is that listeners have no control over how the track was built or what material the program learned from.
Are there hidden messages in music?
The fear is old, but evidence is lacking. In 1990 Judas Priest went on trial in the United States over alleged subliminal messages in a song: the judge dismissed the case for lack of evidence.
Do binaural beats work?
Results are mixed. In a review of 14 EEG studies, 5 supported an effect on brain waves, 8 contradicted it and 1 was mixed. There are no firm conclusions.
Can AI copy a singer's voice?
Yes. In 2023 the track "Heart on My Sleeve" imitated the voices of Drake and The Weeknd although they had never sung it. It was removed from platforms at their record label's request.
What music are AI generators trained on?
The companies do not fully disclose it. The major record labels accused them in court of using their recordings without permission; some cases ended in settlements in 2025, others are still ongoing.
Can AI music contain offensive or blasphemous lyrics?
Yes, it can. Generators have filters against some content, but no filter can judge whether a text is respectful or fit for prayer. The lyrics are produced without a person's discernment.
What are fraudulent streams?
They are plays made by automated programs to collect royalties. According to Deezer, in 2025 up to 85% of the plays of fully AI-generated tracks were of this kind.
Do platforms have to say if a track is made with AI?
From 2 August 2026 the EU AI Act requires deepfake content, created or manipulated to look real, to be disclosed, and asks for generated content to be identifiable. Some platforms, such as Deezer, already label generated tracks.
Who were the Velvet Sundown?
A 1960s-style "band" that passed a million monthly listeners on Spotify in 2025. It turned out that music, lyrics, photos and members were all AI-generated.
Is it dangerous to listen to AI music?
There is no evidence of direct physical harm. The risk is knowing little about what you hear: who made it, with what material, with what intention. For important and sacred moments it is wiser to choose music made by people.
How can I protect myself from AI music?
Choose artists with a name, a story and live concerts you can verify, check the labels on platforms and, for ceremonies, always ask for real musicians.