AI music is more common – and harder to catch – than ever

Last fall, Joe Miller noticed something unusual while experimenting with Suno, a generative artificial intelligence platform that creates music based on text inputs from users. When he entered the lyrics “Daytrip took it to ten,” Suno generated audio that approximated the real-world producer tag of duo Take a Daytrip.

Producer tags — short, punchy audio clips common at the beginning of pop and rap tracks — function like an artist’s signature, identifying the record producers involved in a song. But Take a Daytrip had nothing to do with Miller’s creation; Suno, the AI program, made the association and generated an uncanny replica.

Miller, who recently deleted his Suno account, didn’t want to fool anyone with his work, but would that be possible? While sophisticated listeners can tell the difference between music produced by humans and AI, text-to-music models like Suno have improved to the point that some people could be fooled, according to Luca Comanducci of the Image and Sound Processing Lab at the Polytechnic University of Milan. 

“If you’re casually just listening to the music in an elevator, you probably wouldn’t realize in some cases,” that a track was generated by AI, Comanducci said.

But there is hope for protecting music creators. Scientists are developing systems to identify audio produced using generative AI, similar to deepfake video detection tools like Reality Defender and Sentinel. Building tools to spot AI-generated video has been a higher priority target for researchers because of the immense harm deepfakes can do in spreading misinformation, but Comanducci said audio detection programs are starting to catch up in response to the boom in text-to-music services like Suno and Udio.

Comanducci and his collaborators have created an open-source dataset for researchers working on AI music detection system. IRCAM Amplify, the commercial wing of the Institute for Research and Coordination in Acoustics/Music in France, recently launched a detection system which is already in use by an unnamed digital streaming platform.

Meanwhile, French streaming service Deezer rolled out its own detection model in January. In a press release, the company said over 10% of tracks uploaded to the platform are flagged as AI-generated by their model.

The new AI audio detection tools are coming at a time of growing controversy over the rise of AI-generated music, with experts disagreeing about whether the recent proliferation of text-to-music tools represents a threat or opportunity for musicians. This conflict echoes the broader conversation around generative AI in music as industry figures, legislators and consumers debate whether to constrain or unleash these new technologies.

In lawsuits filed jointly by several major record labels last June against Suno and Udio, the plaintiffs alleged the companies were training the models with copyrighted works, creating unfair competition for their artists. A Recording Industry Association of America press release about the complaints linked to examples of AI-generated tracks they claim lift directly from music heavyweights such as Jerry Lee Lewis and Jason Derulo.  

Both AI companies have filed responses arguing training their algorithms on copyrighted material constitutes fair use, as their models have simply learned from this music and are using it to generate new works. 

When asked in an email about the data used to create its model, a representative of Udio declined to comment, citing the ongoing litigation, but shared Udio’s frequently asked questions page for information about how the company addresses potential copyright infringement on the platform. Suno did not respond to a request for comment.

In March 2024, Tennessee lawmakers passed the Ensuring Voice Likeness and Security Act, which protects an artist’s voice from AI imitation. The following month, over 200 artists ranging from Nicki Minaj to the estate of Frank Sinatra signed an open letter warning that a glut of music produced with generative AI could devalue human creativity. 

But generative AI, including text-to-music tools, also has powerful proponents in the music industry. Producer Timbaland partnered with Suno for a contest to remix his new single “Love Again” on the platform in 2024.

Many users of text-to-music platforms feel these tools democratize music production. Before using Udio, creator Nathan Shea was interested in music but lacked the confidence to express himself. He compares the platform to a band of musicians that can help him work through ideas, emphasizing his active role in the production process. 

“There is a misconception that AI creators throw a prompt in to create something and just take that as it is,” Shea wrote in a direct message on Reddit. “While there are definitely those that do that, there are many who really use it as a tool to create things that they couldn’t have before.” 

The music-generating platforms are proliferating thanks to growing public interest in generative AI, which has been driven in turn by improvements in the technology, especially increases in computer power and refinements in diffusion models and large language models, according to Darius Afchar, a researcher from Deezer.

Diffusion models are an efficient form of machine learning that add noise, or meaningless information, to “diffuse” a training dataset. The models generate new content by removing this random noise. Some of the most popular text-to-image tools, such as DALL-E and Midjourney, rely on diffusion models. 

Large language models, meanwhile, have enabled text-to-music tools by organizing musical elements represented in word format. The models predict what words are likely to be linked together, and text-to-music platforms use this information to produce content based on a user’s text input. 

Researchers are building on their knowledge of how AI music generators work to construct AI music detection programs. Comanducci’s dataset consists of non-AI music and corresponding text prompts from the Google MusicCaps project and the resulting music produced when those prompts are fed into the most common open-source text-to-music algorithms. His team used this data to train a preliminary classification model that they have tested on Suno-generated audio.

Afchar built his AI detection tool by training a classifier he says can detect AI-generated music with more than 99% accuracy, but he warns that it still can be tricked. 

For example, generated music could be disguised with added elements that fool a classification model, in the same way that people who upload copyrighted music to YouTube can avoid detection by pitch-shifting the tracks. Afchar argues a detection model also needs to be interpretable, meaning that its reasons for flagging a track as AI-created should be transparent so they can be communicated to users. 

Finally, even a perfect model would have to keep up with advances in AI music creation. “We cannot magically generalize to everything,” Afchar said. 

For example, IRCAM Amplify recently updated their product to detect music created using the latest version of Suno. IRCAM Amplify’s AI detection program was created in just a few months and was the first commercially available tool upon its release in May 2024, noted Romain Simiand, the company’s chief product officer. He said this rapid development was deliberately prioritized over the tool’s ability to generalize to versions of AI music. Simian praised the Deezer model as a “scalable” technology as it is being designed to anticipate future updates to AI platforms.

Streaming companies that deploy detection models must decide how to handle AI-generated content uploaded to their services. Deezer’s current strategy is to tag AI tracks and remove them from the platform’s algorithmic recommendations to users.

Even if AI-generated works are allowed on streaming services going forward, that doesn’t guarantee they will find an audience right away. Shea, the sometime Udio user, said that while he supports other creators who use the platform, AI music without a human touch “leaves a lot to be desired.”

Others are more enthusiastic. Sebastien Ferguson studied music production in college and uses Suno to augment his workflow, taking samples he produced with more conventional tools and uploading them to Suno to create longer compositions he then promotes on commercial streaming services. 

Suno is a “crazy tool if used right,” Ferguson wrote in a direct message on Discord.

“I feel comfortable … the music I have released was done as ethically as I could,” Ferguson added, explaining that his main concern was others’ preconceived notions about AI music. “I am proud of anything I share.” 

While many AI music enthusiasts upload their work to platforms like YouTube and Spotify, where they can earn streaming royalties, not every creator hopes to monetize their output.

Before he deleted his account, Miller, the creator who replicated the Take a Daytrip tag in Suno, said he was mostly just interested in exploring the new technology. “It’s weird to treat this AI music stuff like you’ve actually made it.” 

Editor’s Note: This article was updated 11:05 a.m. May 20. A previous version named the commercial wing of the Institute for Research and Coordination in Acoustics/Music as IRCAM, not IRCAM Amplify.