Once upon a time, poison pills were what spies carried with them in case they wanted to commit suicide rather than be captured. In the financial world, meanwhile, a poison pill is a tactic that companies can use to fend off hostile takeovers.
In the music/tech sphere, though, Poison Pill is a new startup based in the UK that is building technology to help music companies fight back against unlicensed AI training by “poisoning” their own music. It launched in beta last week, and is also a finalist in our 2025 Music Ally SI:X startups contest.
It’s the latest venture from Ben Bowler, whose previous startups include Chew·tv and Aux. While the exact details of how Poison Pill’s technology works are necessarily secret, he talked to Music Ally about the bigger picture, and who the service is aimed at.
“Most musicians are pissed about the current state of AI in music: well-funded companies are scraping music without permission, creating services that claim to replace them. Streaming services with AI-filled playlists. AI music is taking over previous money makers like sync,” he said.
“That’s why, as a first step, Poison Pill is open to independent artists who want to stick it to these AI companies. Our aim is to protect 20% of independent music. With this poisoned music out there, we can shift the power dynamic and bring AI firms to the negotiating table for fair licensing of training data for independent musicians.”
How exactly does the company ‘poison’ the music? It’s all about ‘adversarial noise algorithms’ – one of the techniques that researchers have designed to disrupt AI models.
“It exploits how AI models learn to recognise genre, instruments, etc, by identifying highly specific sound features: finding shortcuts based on their training to make predictions quickly,” is how Bowler explained it.
“We never know exactly what these features are, but imagine that the model detects indie music by looking for a certain resonance pattern made when a guitar string is plucked, and classical music by looking for a different resonance found in wind instruments,” he continued.
“We can use this nature of AI against itself by subtly adding low-level noise to an indie track that closely matches what the model expects classical music to contain. The level of this noise is so subtle that humans cannot hear it, but AI, which relies on these shortcuts, will quickly mistake your indie track for classical.”
According to Bowler, this is disruptive because if GenAI models are being trained to produce music based on text prompts, they need to accurately understand the genre and instrumentation of the tracks they are learning from.
“If suddenly many tracks are embedded with these attacks, future versions of these models will start generating random styles, seriously disrupting their usefulness,” he said.
The pitch sounds appealing, but can it really work? Or rather, won’t AI companies spot the technology being used and develop ways around it? Bowler acknowledged the potential for that, but said that it does not negate the value of using Poison Pill’s technology.
“The protections are difficult to detect and remove entirely, but yes, AI companies can reduce their impact. Importantly, the system doesn’t need to be perfect to effectively change the power balance between creators and scrapers,” he said.
“As more music incorporates ever-evolving Poison Pill protection, the cost for these companies to identify and address these protections increases progressively. These companies depend on the fact they can quickly and cheaply process large volumes of music,” he added.
“If they have to check and re-process every track before training new versions of their models, paying a fair price to license clean music becomes a far better option.”
Poison Pill’s service is now available in its beta, with Bowler saying that the startup is now focusing on continuing to improve its protection algorithm, applying new attacks and testing how well it works against different AI systems.
“Beyond that, we are exploring how other creators: photographers, artists, and filmmakers could also use this technology to safeguard their work,” he said.
The company is currently bootstrapped (self-funded), but Bowler said that Poison Pill has already attracted interest from several potential investors.
“Additionally, we are in discussions with potential partners, such as labels and larger rights holders, to provide API access to our technology for the protection of larger catalogues,” he added.
Related Stories