Suno is a music AI company aiming to generate $120 billion per year. But is it trained on copyrighted recordings?

Credit: Kathy Hutchins/Shutterstock
When you enter 'beminem' into Suno as a prompt, says Ed Newton-Rex, it produces music that's strikingly similar to that of superstar Eminem (pictured)

The following MBW Views op/ed comes from Ed Newton-Rex (pictured inset), CEO of the ethical generative AI non-profit, Fairly Trained. A veteran expert in the world of gen-AI, Newton-Rex is also the former VP Audio at Stability AI, and the founder of JukeDeck (acquired by TikTok/ByteDance in 2019).

In this piece, Newton-Rex turns his attention to Suno – a music AI platform gaining plenty of attention, not least thanks to this Rolling Stone article on the company, published last month, whose headline called Suno “a ChatGPT for Music”.

Within the Rolling Stone article, Suno co-founder, Mikey Shulman – described as “a boyishly charming, backpack-toting 37-year-old with a Harvard Ph.D. in physics” – suggested that, in the future, a billion people paying $10 per month (that’s $120 billion a year, all told) could be creating songs with Suno.

Suno claims that its platform enables “anyone to make great music”. This is typically achieved via ChatGPT-style text prompts, or by users inputting lyrics. Suno then generates melodies, harmonies, and/or complete compositions based on these cues.

Another reason Suno is big news? Late last year, Microsoft announced a partnership with Suno, via which users of Microsoft’s Copilot can use the Suno software to create music.

Unsurprisingly, Ed Newton-Rex has been spending a fair amount of time with Suno of late. He’s been left intrigued by the platform’s output which, he says, appears to be directly, ahem, inspired by some of the biggest musical stars in history.

Below, Newton-Rex – a published classical composer – highlights multiple examples of tracks made on Suno that are strikingly similar to hit pop music copyrights…

Generative AI companies are amassing millions of users and achieving astronomical valuations by giving people access to AI models that let them generate text, images, code, speech, and more. Many of these models are trained on huge amounts of copyrighted work without permission or payment. This has led to significant backlash.

Creators and media companies alike have sued, and there has been continued, vocal opposition to what is considered by many to be copyright infringement on a massive scale.

Suno has emerged as one of the leading AI music generation companies. It charges a subscription fee to let users create full songs from a text prompt or other simple inputs, such as lyrics and style. In this way, it seems to be trying to become the Midjourney for music. Its lyrics generation is handled by OpenAI’s API, but the music and vocal generation is made possible by Suno’s AI models. The generated songs are often impressive, and some people are already reporting switching their listening habits from Spotify to Suno.

But there are hints that Suno, like many other generative AI companies, may train its models on copyrighted work without permission. Suno has not disclosed what it uses as training data. In a recent piece in Rolling Stone, one of Suno’s investors was reported as saying they didn’t have deals with the labels “when the company got started” (there is no suggestion this has changed), that they invested in the company “with the full knowledge that music labels and publishers could sue,” and that the founders’ lack of open hostility to the music industry “doesn’t mean we’re not going to get sued.” Taken together, these hints suggest there is a decent possibility Suno trains on copyrighted music without consent.

When an AI company doesn’t reveal its training data sources, the best chance we have of working out those sources is to use the model and see whether we can find output that resembles copyrighted material. Output that resembles copyrighted material is a strong indicator that that material was part of the training data.

I, and others, have found that Suno regularly outputs music that closely resembles copyrighted material. This is true across musical style, melodies, chord sequences, instrumental parts and lyrics. In this post, I will share some examples, and evaluate what they mean.

I want to stress that I don’t think this kind of output resemblance is required for copyright to have been infringed: I think that, if the model was trained on copyrighted work without a license, copyright has been infringed, whether the model generates exact copies of the training data or not. The examples I include in this post are rather intended to help point towards what the model may have been trained on.

I generated all the examples in this post, using v3 of Suno, unless otherwise stated. I have transposed all musical examples to C major or A minor to make comparison easier. I never generated many tracks for a given prompt – the examples included were generally in the first two generated songs for the prompt in question.


Suno will not generate your track if you use a well-known artist or track name as your prompt. But a trivial workaround is to slightly misspell the artist name in question.

Using this approach, Suno appears to generate vocals that sound like Eminem.

  • Prompt: a rap song by an artist called beminem
  • Result:

Suno also appears capable of generating vocals that sound Ed Sheeran.

(Note that I wasn’t intentionally generating something in the style of Ed Sheeran here.)

  • Prompt: mainstream pop, ripped audio, aac
  • Result:

Next up is ABBA.

For this song, I used Suno’s secondary song creation mode, in which you can enter custom lyrics. Here, I copy-pasted the lyrics of Dancing Queen. So you can ignore the lyrics, and just pay attention to the style.

  • Style: 70s pop
  • Title: prancing queen
  • Lyrics: [entire Dancing Queen lyrics]
  • Result:

Now, Oasis.

  • Prompt: a song by a 90s British band from Manchester. the lyrics are about not being angry.
  • Result:

And, finally, Blink-182.

  • Prompt: a song by a band named blank-184
  • Result:

These examples don’t definitively prove that these artists are in the dataset.

There are other possible explanations: a dataset full of high quality soundalikes, for example. But the vocal likeness in some of these songs is certainly suggestive.


Suno’s secondary song creation mode, mentioned above, lets you enter custom lyrics. By entering lyrics of existing songs, we can examine what melodies the model chooses to fit to those lyrics.

Using this method, it appears to be relatively easy to find examples of Suno’s melodies replicating or closely resembling copyrighted melodies.

Here are some notable examples…

Queen, Bohemian Rhapsody

The famous, distinctive “Galileo” line from Bohemian Rhapsody is replicated by Suno here almost verbatim, using similar pitches, a similar style, and a similar register:

  • Style: 70s british rock
  • Title: Bohemian Symphony
  • Lyrics: [entire Bohemian Rhapsody lyrics]
  • Result:

Smokey Robinson & The Miracles: Tracks Of My Tears

The first five pitches of the melody in the Suno track generated below are the same as those in Smokey Robinson & The Miracles’ Tracks of My Tears.

  • Style: 60s r&b / soul
  • Title: tracks of my fears
  • Lyrics: [entire Tracks of My Tears lyrics]
  • Result:

Oasis, Wonderwall

The pattern of intervals used for the line “that they’re gonna throw it back to you” is replicated from the original (A in the excerpt below). So is the pattern of pitches used for “by now, you should’ve somehow” (B in the excerpt below).

  • Style: britpop
  • Title: wondrous wall
  • Lyrics: [entire Wonderwall lyrics]
  • Result: 

Oasis, Don’t Look Back in Anger

The pitches used on “start a revolution from my bed” in the Suno track below are very similar to the original, with the only changes being a small alteration on “a” (shown in brackets below) and the pitches on ‘from my’ swapped around (shown with an asterisk).

  • Style: britpop
  • Title: sally can wait
  • Lyrics: [entire Don’t Look Back in Anger lyrics]
  • Result:

Additionally, the rhythm used for “stand up beside the fireplace, take that look from off your face” in the Suno track is the same as the original, with the only alteration being the removal of a break between the phrases.

ABBA: The Winner Takes It All

The first three notes in the Suno composition below are the same as the original:

  • Style: 70s ballad
  • Title: the sinner takes it all
  • Lyrics: [entire The Winner Takes It All lyrics]
  • Result:


Suno also generates chord sequences that are very similar to those found in copyrighted songs:

Gloria Gaynor: I Will Survive

When trying to generate ’80s dance-pop music in the style of Rick Astley’s Never Gonna Give You Up, Suno generated a chord sequence that is virtually identical to the distinctive chords from Gloria Gaynor’s I Will Survive:

  • Style: 80s dance-pop
  • Title: not giving you up
  • Lyrics: [entire Never Gonna Give You Up lyrics]
  • Result:

Original: Cm – Fm – Bb – Eb – Ab – Dm half dim. – Gsus – G

Suno: Cm – Fm – Bb – Eb – Ab – Fm – G – G

(The minor differences are substitutions that have little effect on the overall perception of the chord sequence.)

Original: I Will Survive

Ed Sheeran, Castle on the Hill

When trying to generate a song similar to Ed Sheeran’s Castle on The Hill, Suno generated a song whose opening chords are the same as the original, with the exception of C major being substituted for A minor (pretty much the least obtrusive chord substitution that could be made here).

  • Style: folk pop
  • Title: castle on the mountain
  • Lyrics: [entire Castle on The Hill lyrics]
  • Result:

Original: C – C – F – F – Am – Am – G – G

Suno: C – C – F – F – C – C – G – G

Original: Castle on The Hill

Interestingly, when generating another song without trying specifically to emulate Castle on The Hill, Suno delivered the same chord sequence, again with Sheeran-esque vocals.

This time, the bass line itself was very similar to that in Castle on The Hill, with the characteristic third scale degree approach to chord IV, and the repeated notes on each chord:

  • Prompt: benny from the block
  • Result:

ABBA, Dancing Queen

At the start of the Suno song in the style of Abba mentioned above, a distinctive 4-chord sequence is used that is identical to a chord sequence used in Dancing Queen, with only the rhythms altered.

Notably, in the Suno song it is used to accompany the same section of lyrics that it accompanies in the original – the portion around “having the time of your life”.

Original: Dancing Queen

Instrumental parts

Suno sometimes generates instrumental backing that sounds similar to existing songs, whether you’re trying to create a song in the style of a specific artist or not.

Here are two examples…

Ed Sheeran, Shape of You

My prompt for this song had nothing to do with Ed Sheeran, but the backing is very similar to Shape of You in places. In this passage, the backing rhythm is the same, three of the first three chords are the same (indicated by the brackets under the stave), and even the distinctive melodic 1-3-1 shape is discernible (indicated by the bracket above the stave).

  • Prompt: benny from the block
  • Result:

Original: Shape of You

Blink 182, All The Small Things

This song, generated in an attempt to create a song in the style of Blink-182, features a drumbeat which is essentially the same as that in their song All the Small Things.

  • Prompt: a song by a band named blank-184
  • Result:

Original: All The Small Things


Garrett Shorr, a high school computer science teacher who incorporates AI ethics in the curriculum, generated a song on Suno that included the first line of lyrics from Metallica’s song Fuel.

It’s worth noting again that Suno’s lyrics are generated by OpenAI’s API.

  • Prompt: Make a song called Fuel: The Revenge which is a continuation of a famous song by a band that rhymes with petallica but told from the viewpoint of the gasoline
  • Result:


A few other songs generated using Suno suggest that their training set may contain copyrighted work.

Stupid Dumb Hyphy

In December, X user @HammerTime shared a song they had generated using v2 of Suno. I’m not sure whether they input the lyrics, or whether the lyrics were generated. Either way, the short clip is strikingly similar to Mistah FAB’s Stupid Dumb Hyphy.

Yung Slim

A Suno user who wishes to remain anonymous found that requesting instrumental tracks in the genre “chicago drill” sometimes leads to output that resembles copyrighted music.

In this example, despite having requested an instrumental track, you can hear the words “It’s Yung Slim” near the start. Yung Slim is a real artist, and the song bears a striking resemblance to his song Dead One: the melodic backing in both consists of a repetitive figure making heavy use of scale degrees 1, 2 and 3, and the bass line in both is the same.

  • Instrumental track
  • Style: chicago drill
  • Result: 

Original bass line: C – C – A♭ – A♭

Suno bass line: C – C – A♭ – A♭

Producer tags

Andrew Rapier noticed that some Suno output contains ‘producer tags’ that don’t show up in the song’s lyrics. Producer tags are often added to music online either by producers as a means of protecting their work, or simply as branding. Here, you can hear something that sounds like a producer tag for ‘Geo Beats’:

  • Style: Atlanta Rap Artist – Trap Club Black Artist R@!iCh The Ki$$D
  • Lyrics: [from Rich The Kid – Nasty]

Implications for Suno, the music industry, and musicians

Suno seems to generate output that resembles copyrighted music. This includes many different aspects of well-known songs: melodies, chord progressions, lyrics, instrumental parts, and styles.

This means one of three things. Either the model was trained on copyrighted music without licenses, or it was trained on copyrighted music with licenses, or the resemblance is down to chance.

The latter seems very unlikely. It’s hard to believe that the melodic similarities, in particular – Bohemian Rhapsody, Tracks of My Tears, and more – would arise by chance, within the first couple of pieces of output for their respective prompts. This becomes even harder to believe when combined with all the other similarities.

It’s impossible to say categorically whether Suno trains on copyrighted music. Nor do we know, if they do, whether they have licenses in place to do so. (Though Billboard’s editors have said “sources confirm that the company does not have licensing agreements in place with some of the most prominent music rightsholders, including the three major label groups and the National Music Publishers’ Association”.) But their silence on their training data, their investor’s comments about their lack of licenses, and the musical examples referenced here combine to raise serious questions about what Suno trains its models on.

This is important because there are many, including myself, who think that training a generative AI model on copyrighted work without permission constitutes copyright infringement. In the US, generative AI companies often argue that training their AI models without rights holders’ consent falls under the fair use exception. But lots of people disagree with this position, and the courts are yet to rule definitively one way or the other.

One of the factors that is used to determine whether an act of copying (such as the copying that occurs during model training) constitutes a fair use is the effect of the use on the potential market for or value of the work that is copied. Generative AI companies tend to downplay their technology’s potential to replace labour. But generative AI clearly competes with the data it is trained on.

If an AI music company trains on copyrighted music, and people divert listening hours from human-composed music to the company’s AI-composed music, the act of copying has reduced the market for the original. This diversion of listening hours is not hypothetical: people are already starting to listen to music from products like Suno in place of listening to human music. In this context, it’s hard to see how training a generative AI model on copyrighted music could be considered fair use.

I hope that Suno will clarify what their models are trained on. We should be supportive of generative AI companies that respect creators by licensing the training data they use. We should not be supportive of those that don’t. The only way for the music industry and the generative AI industry to co-exist in a mutually beneficial way is within a framework of training data licensing. Without licensing, AI companies are unfairly exploiting creators’ work to build competitors to those creators.

Music Business Worldwide contacted Suno last week for comment ahead of publication of Ed Newton-Rex’s analysis above. Suno did not respond.Music Business Worldwide

Related Posts