The Fight to Protect Music in the Age of AI
In a small studio in Georgia, surrounded by the hum of high-powered GPUs and the faint glow of solar panels, musician and technologist Benn Jordan is waging a quiet war. For over 25 years, Jordan has carved out a living as an independent artist, releasing music under various pseudonyms and building a loyal following. But in recent years, the rise of generative artificial intelligence (AI) has threatened the very foundation of his craft. Tech companies, armed with billions in venture capital, have been scraping music from platforms like Spotify and YouTube, using it to train AI models that generate derivative tracks—often without the consent of the original artists. For Jordan, this wasn’t just an ethical violation; it was a direct assault on his livelihood. “I still enjoy making music all the time,” he says in a YouTube video that has garnered over 600,000 views, “but I have entirely stopped releasing it.”
Jordan’s response to this crisis is a technological counterstrike: a pair of innovative tools called Harmony Cloak and Poisonify. These methods encode music files with adversarial noise—subtle, inaudible distortions that confuse AI models, rendering the music untrainable or even degrading the quality of the AI’s entire dataset. It’s a bold move, one that flips the script on the generative AI industry, which has long treated artists’ work as raw material to be mined without permission. As Jordan puts it, “Unethical generative AI companies have made artists feel incredibly powerless for quite some time now. But all of that is about to change.”
The Rise of AI Music and the Artist’s Dilemma
The story of AI-generated music begins in 2015 with the publication of U-Net, a convolutional neural network designed for pattern recognition in biomedical imagery. Its efficient architecture, which required minimal training data, inspired the diffusion models that power today’s generative AI, from image generators like DALL-E to music tools like Suno and Meta’s MusicGen. By 2016, Google’s Magenta project was scanning vast libraries of music to create tools that could, for instance, turn a voice into a saxophone or generate a basic melody from a few circles on a screen. These were fun experiments, but as Jordan notes, they were often “a really expensive and inefficient way to do what modular synths have been doing for decades.”
The real trouble began when companies like Open AI, ByteDance, and Suno started using copyrighted music to train their models, often without transparency or consent. A 2016 Sony project, “Daddy’s Car,” billed as an AI-generated song, turned out to be performed, recorded, and mastered by humans, with lyrics written by a human too. “How exactly does this qualify as artificial intelligence?” Jordan quips, quoting the song’s nonsensical lyrics: “From tax man until tomorrow, never know.” By 2019, the landscape had shifted. Voice cloning services and consumer-facing platforms like Suno and Udio emerged, offering AI-generated music for a subscription fee. These services flooded streaming platforms with millions of AI-generated tracks, siphoning royalties away from human musicians.
The scale of the problem is staggering. Suno, a generative AI company launched in late 2023, is now embroiled in a legal battle with the Recording Industry Association of America, facing potential damages of $1.5 trillion for training its models on copyrighted data. As Jordan points out, asking a company, “What data did you use to train your base model?” often results in silence or evasion, as answering could expose them to billions in intellectual property (IP) infringement liabilities. On X, users have echoed this sentiment, with one posting, “AI companies scraping music without permission is just digital theft dressed up as innovation. Good for artists like Benn Jordan fighting back.” Another user lamented, “The fact that Suno pays AI ‘creators’ but not the musicians whose work they trained on is wild.”
Harmony Cloak and Poisonify: The Artist’s Revenge
Jordan’s response to this crisis is both technical and philosophical. In 2023, he co-founded Voice Swap AI (now transitioning to Topset Labs), a company that trains vocal models on consensual data, paying artists royalties and offering them equity. The result? A platform with 150,000 users, no venture capital, and vocalists earning more from their AI models than from Spotify. But Jordan’s ambitions go beyond ethical AI. He wants to protect artists from exploitation at the source.
Enter Harmony Cloak and Poisonify. Harmony Cloak, developed by researchers at the University of Tennessee, Knoxville, encodes music files with adversarial noise that disrupts an AI’s ability to detect melody or rhythm. In demonstrations, Jordan shows how AI models generate coherent extensions of unencoded music but produce chaotic, unlistenable noise when fed files treated with Harmony Cloak. Poisonify, Jordan’s own creation, takes this a step further. By embedding inaudible distortions, it tricks instrument classifiers into misidentifying sounds—a cymbal might be mistaken for a harmonica, a synthesizer for a string quartet. This not only makes the music untrainable but can degrade the AI model’s overall performance, creating a “snowball effect” of false positives.
In his YouTube video, Jordan demonstrates Poisonify’s impact on Suno’s song extension feature. An original track produces a passable extension, but the Poisonify-encoded version results in what he describes as “music from an airport spa that somebody downloaded off Napster in 1999.” Similar tests on Miniax Audio and Meta’s MusicGen yield even more dramatic results, with the latter crashing entirely. On X, a user commented, “Benn’s Poisonify demo is savage. It’s like he’s poisoning the AI’s water supply.” Another wrote, “Harmony Cloak and Poisonify are the artist’s equivalent of a DDoS attack. Brilliant.”
The technology isn’t perfect. Encoding a single album with Poisonify requires two weeks of GPU processing, consuming 242 kW of electricity—costing between $40 and $150 depending on location. Harmony Cloak, too, demands high-end GPUs, though the Knoxville team is working on a more efficient model. Still, Jordan envisions a future where these tools are offered as an API through distributors like Symphonic Distribution, allowing artists to “AI-proof” their music when uploading to streaming platforms.
The Ethics of AI and the Pareto Principle
Jordan’s work raises deeper questions about the generative AI industry’s sustainability. He invokes the Pareto Principle, or the 80/20 rule, to argue that AI models have hit a wall. “AI image generators got really good really fast, but are still nowhere near perfect,” he says. Despite massive investments, issues like rendering text or hands remain unsolved without specialized workarounds. Music AI faces similar limitations: new features often prioritize bells and whistles like inpainting over core quality improvements, and models frequently ignore user prompts in pursuit of clearer sound.
This stagnation, Jordan argues, stems from the industry’s reliance on unethically sourced data. “Concentrating on the input data and working with vocalists is way less expensive and much easier than the trial and error of retraining base models,” he says. Voice Swap’s success—built on consensual data and artist collaboration—proves this point. A post on X echoed this view: “Benn’s right—AI companies burning billions on retraining models are just chasing their tails. Pay artists, use clean data, and you’ll get better results.”
The ethical dimension is inescapable. Suno’s CEO, Michael Schulman, has claimed that “the majority of people don’t enjoy the majority of the time they spend making music,” a statement Jordan finds delusional. On YouTube, a commenter responded, “Suno’s CEO saying musicians don’t enjoy making music is like a chef saying people hate cooking. Clueless.” By contrast, Jordan’s approach empowers artists, offering them tools to protect their work and a model that compensates them fairly.
The Broader Implications
Jordan’s technology has implications beyond music. Adversarial noise can manipulate AI assistants, embedding inaudible commands to unlock doors or disable alarms. In one chilling demo, Jordan plays soft classical music that tricks an Amazon Echo into processing a hidden command. He also envisions using targeted pressure waves to protect live performances from being recorded by smart devices, ensuring that an unfinished song shared at a concert doesn’t end up on Instagram.
These applications highlight a growing tension: as AI becomes ubiquitous, so do the tools to subvert it. Researchers at the University of Tennessee have refined adversarial attacks to bypass high-security settings, mimicking environmental sounds to fool voice recognition systems. On X, a user noted, “This is bigger than music. If Jordan’s tech can mess with Alexa or Siri, it’s a game-changer for privacy.”
Yet challenges remain. Poisonify’s misclassification could disrupt Spotify’s recommendation algorithms, potentially recommending abrasive techno to barbershop quartet fans. Jordan finds this amusing, but acknowledges that some artists might not. Efficiency is another hurdle; the computational cost of encoding music is prohibitive for individual artists, though an API could democratize access.
A New Hope for Artists
Jordan’s journey reflects a broader shift in the creative industries. In early 2024, distributor Tunecore falsely accused him of fake streaming, removing his catalog from streaming platforms and costing him 100,000 monthly listeners. After public backlash, Tunecore restored most of his music, but the metadata errors persisted. This experience underscored the fragility of artists’ livelihoods in the digital age, pushing Jordan to negotiate with distributors like Symphonic, whose CEO, Jorge Bria, is open to integrating Poisonify and Harmony Cloak into their platform.
For now, Jordan is finishing a new album, encoding it with a mix of poison pill methods to obscure how they work. “I’m not going to tell anybody which tracks are encoded with what,” he says, “so it can’t be avoided down the line.” His YouTube viewers have rallied behind him, with comments like, “Benn is out here fighting for all artists. This is what innovation looks like,” and “Poisonify is the punk rock of AI defense.”
As the generative AI industry grapples with its ethical and technical limitations, artists like Jordan are reclaiming control. His work with Voice Swap and Topset Labs offers a blueprint: prioritize consent, compensate creators, and use technology to protect rather than exploit. In a world where AI can mimic an eagle’s cry to unlock your door, Jordan’s fight is a reminder that creativity and ingenuity remain humanity’s greatest assets. “I’m not an AI hater,” he says. “But developing a useful tool will pay out much higher than developing an investment scheme.” For the millions of artists watching their work being scraped and resold, that’s a melody worth hearing.
Comments