This video explores technological methods for musicians to combat the unauthorized use of their work by generative AI companies that scrape copyrighted data. The speaker, a musician and tech entrepreneur, introduces two primary techniques: Harmony Cloak and Poisonify. These methods embed inaudible adversarial noise into audio files designed to disrupt AI model training. Harmony Cloak aims to break AI's ability to detect melody and rhythm, while Poisonify targets instrument classification. Demonstrations show how these encoded files cause AI music generation services to produce distorted or nonsensical outputs and can potentially degrade the training data quality of AI models. The video also touches upon the broader implications of adversarial noise for voice assistants and recording devices. While challenges like processing power and unintended effects on services like Spotify exist, the speaker is developing these techniques into an API service, seeking partnerships with music distributors to offer artists a technical defense against non-consensual AI training, independent of often insufficient legal protections. The business perspective highlights the efficiency of ethical data sourcing and artist collaboration compared to the diminishing returns of endlessly retraining models solely on scraped data.
Fighting Back Against AI Scraping: Technical Tools for Musicians
0:00 A musician and tech entrepreneur presents a solution to generative AI companies scraping music without consent and generating new music associated with the artist's name.
0:35 The core idea is an encoding technology that makes music files "untrainable" for generative AI and can decrease the quality and efficiency of their entire data sets.
0:50 This technology, based on "adversarial noise," can attack AI in other ways, such as making smart devices respond to inaudible commands (e.g., telling Siri/Echo/Google Home to unlock doors) or misclassifying audio (e.g., Spotify classifying sex noises as folk music).
1:46 The video provides a brief history of AI music generation, from UNET's diffusion inspiration (2015) and Google Magenta (2016) to the rise of services like OpenAI Museet, ByteDance Juke Deck, Meta, Stability, Suno, and UMG (starting around 2019).
4:02 Many AI companies have recklessly scraped vast amounts of copyrighted data from platforms like Spotify, YouTube, and Audible without consent, leading to potential massive liability.
4:51 This unauthorized scraping has led to opportunists flooding streaming services with AI-generated music, siphoning royalties away from human artists.
5:54 The speaker co-founded Voice Swap AI, an ethical voice cloning company that trains models only on consensual data, pays artists ongoing licensing fees and equity, and has found this approach leads to superior models and artists earning more than from traditional streaming.
6:51 The speaker developed an AI music detector, but distributors lacked the incentive to use it, leading to research into proactive defense using adversarial noise.
7:46 Adversarial noise is a known information security concept (dating back a decade) that exploits the difference between human and neural network perception of sound.
9:39 Collaborating with researchers at the University of Tennessee Knoxville, the speaker explores "Harmony Cloak," a project that encodes music files with adversarial noise to break AI's ability to find melody or rhythm.
10:20 Demonstrations show that while humans cannot hear the difference in Harmony Cloak encoded music, AI models fail to generate coherent music from it, producing chaotic results.
10:54 Analysis using Tensorboard confirms Harmony Cloak encoded data prevents AI models from improving during training.
11:57 The speaker developed "Poisonify," a method focusing on targeted adversarial noise attacks to break AI instrument classification, potentially causing a snowball effect of misclassification in AI models.
15:07 Tests on popular AI music generators like Suno, Miniax Audio, and Meta Music Gen show that feeding them Poisonify/Harmony Cloak encoded songs results in poor quality, nonsensical, or crashing outputs.
18:08 Challenges include potential unintended effects (e.g., miscategorization on Spotify's recommendation algorithm) and the current high computational cost of encoding (requiring expensive GPUs and significant time/electricity).
19:14 The goal is to make the encoding process more efficient and offer it as an API service via Topset Labs (future name for Voice Swap) that willing distributors can offer to artists as an option when uploading music.
20:18 The speaker is exploring a partnership with Symphonic Distribution, whose CEO is open-minded about incorporating such an AI-proofing service.
21:16 The speaker is releasing a new album and will randomly encode their entire discography using various methods (including some ineffective noise) to obfuscate the techniques and prevent AI companies from easily adapting.
21:44 Future work includes exploring broad protection from AI listening devices in physical spaces using targeted pressure waves.
22:53 The aim is to empower artists with technological tools to protect their work and privacy, reducing reliance on legal frameworks that have struggled to adapt to AI.
23:17 The video argues that the generative AI industry faces an existential problem due to the Pareto principle: initial rapid quality gains are easy (80/20 rule), but achieving the final percentage of quality is an expensive, complicated grind, making ethical, consensual data sourcing and artist collaboration a more sustainable business model than scraping.