Skip Navigation
Music Production @sh.itjust.works
trashcan @sh.itjust.works
Video

The Art Of Poison-Pilling Music Files | Benn Jordan

Only halfway in and it's already very interesting and too much for me to summarize but I do love the idea of subtly releasing music to poison models.

The demonstration at 15:07 is a good preview but the entire video is worthwhile.

Or 22:25 for a more malicious use of this tech.

Edit: fixed timestamped links.

4 comments

  • Some things to consider with regards to software like this. If music can be heard, it can be scraped, period. Even if you put up barriers on major streaming platforms or embed “anti-AI” tags, all it takes is:

    • Someone recording the audio with a mic (analog loophole)
    • A downloader that bypasses the protection (e.g., YouTube-dl)
    • A source that doesn’t respect the protections (pirate sites, leaks, live audience recordings)

    If a human can access it, an AI can be trained on it, even secondhand, and unlike traditional use cases where clean, labeled data is critical, AI models can learn from messy or partial data. Even if you degrade the quality or watermark it, a model can still extract style, rhythm, melody and timbre, just like how humans can recognize a song through static.

    Also, you can’t control every upload, every sample, every remix, every bootleg. As soon as someone puts your protected content in a place without safeguards, it’s back in the 'training pool'.

    Even if AI models never directly train on your content, they can still learn your style by training on other artists influenced by you, or on users uploading “in the style of” recreations. Protection doesn’t stop style emulation which is what many people want from AI anyway.

    Finally just because AI avoids your data doesn’t mean it avoids imitating you. You may block scrapers, but unless copyright law adapts to handle stylistic theft, there’s no real recourse when AI replicates your sound or vibe.