~4m10:49
TUNED INTO TECH

Poison Your Data. Fight Back Against AI.

May 8, 2026

Read: ~4m · You save: 7 min

Poison Your Data. Fight Back Against AI.

How ordinary people are poisoning music data to fight AI scrapers, from Homer Simpson vocals on Soulseek to inaudible audio watermarks that corrupt AI training.

The Open Internet and Its Original Promise

In 1996, John Perry Barlow published A Declaration of the Independence of Cyberspace. His argument was that the internet exists outside the physical world and therefore should not be governed by the same rules and regulations. He envisioned no borders and no paywalls, only ordinary people connecting and sharing freely.

The period that followed is one most people have forgotten. Communities formed on peer-to-peer file-sharing networks, where people shared art, music, knowledge, and information without corporations, paywalls, or algorithms.

Soulseek is one such network. Launched in 2000, it was built for music lovers by music lovers. It offered rare underground vinyl rips, underground electronic music, and lossless audio that was difficult to find elsewhere. As of 2026, it remains active, and its activity may be higher than ever.

Extraction of Shared Music for AI Training

Barlow did not anticipate that large technology companies would turn these communities into a resource. Companies valued in the billions and trillions of dollars have found these platforms but do not participate in them. They extract data from them.

Sites such as Soulseek hold thousands of hours of high-quality lossless FLAC audio, which AI models require for training. These companies scrape the sites and download everything they can access. The files were shared by individuals and musicians without compensation in mind. AI companies use them to build commercial models, and the original musicians receive no payment.

Data poisoning is presented as a response to this extraction.

Case Study: Replacing Vocals on Soulseek

Mr. Daniels, a 25-year-old from England who experiments with AI and deepfakes, took his personal music library of more than 2,000 records. He removed the original vocals from every track and replaced them with the voice of Homer Simpson. He then uploaded the altered files to Soulseek.

He did not change the metadata, including file names, artist tags, and album information. Because these elements matched the originals, the files appeared legitimate. Some tracks had long intros, so listeners could play them for 30 seconds before noticing the change.

The method targets automated collection. An AI scraper downloads files and reads metadata but does not listen to the audio. It assumes the files are authentic and sorts them accordingly. The resulting training data contains Homer Simpson's voice labeled as music by the original artists. The model treats this data as accurate. Mr. Daniels intends to introduce noise into a system built on clean, high-quality audio.

Mr. Daniels has stated that he feels some guilt because human users can also be affected. These users may download a file, recognize the substitution, and move on to the next file. An AI scraper cannot do this.

Forms of Data Poisoning

Data poisoning takes many forms. During World War II, Operation Mincemeat intentionally supplied the Nazis with false information, which produced a specific outcome.

In the early 2000s, major record labels responded to peer-to-peer file sharing by flooding these networks with corrupted, low-quality audio files. The labels aimed to frustrate users until they stopped downloading and purchased albums instead.

The current approach differs in its target. Earlier efforts aimed at individual users. Current efforts aim at AI companies.

According to the source, Spotify and several major labels have jointly filed lawsuits against AI companies seeking approximately $13 trillion in damages. The source also states that AI companies have downloaded 99.6% of commercial sound recordings available online.

Audio Watermarking as a Poisoning Method

The Homer Simpson method is effective and attracts attention, but it has clear limits. It is time-consuming, and listeners can detect it easily.

Companies including Harmony Clock, Poison Pill, and Synth ID are developing an alternative based on audio watermarking. This method embeds data into an audio track in a way that human listeners cannot detect.

The process works as follows:

  1. An original track is selected.
  2. Additional signals are embedded in patterns within the frequency spectrum.
  3. Human listeners do not perceive the changes.

Technical constraints:

  • Human hearing covers approximately 20 Hz to 20,000 Hz.
  • Standard audio formats represent frequencies up to around 22,000 Hz.
  • In theory, data could be encoded up to 300,000 Hz, but the resulting files would be extremely large.
  • Compression into formats such as MP3 destroys hidden data encoded above the standard range.

In practice, most audio watermarking embeds subtle signals within the audible range. It uses psychoacoustic masking, which hides data beneath louder sounds so that a human ear cannot detect it.

The watermark is not designed for human perception. It is designed for machine reading.

Effect on AI Models

AI models do not listen to music as humans do. They process the complete mathematical representation of an audio file, including every frequency, waveform, and detail that humans cannot perceive. A model trained on watermarked audio absorbs the embedded data without the developer's knowledge.

The consequence concerns the model's understanding of musical fundamentals. A model may classify a track as rock and roll while the watermark has instructed it to treat the material as classical piano. Later, when a user requests a rock and roll recording, the output may sound like soft piano or Mozart.

This degrades the model's reliability and value. A user who receives such output may not return to the tool. Data poisoning in this form is described as weaponized noise. Its presence remains undetected until the damage occurs.

Effectiveness and Limitations

The technology is at an early stage. The tools are new, still in development, and do not yet run locally. This limits adoption.

Data poisoning does not stop the development of AI or cause companies such as OpenAI to fail. Its function is to create friction. It makes data less reliable and forces companies to spend money, time, and resources verifying that training data is clean and legitimate. These resources could otherwise be directed toward advertising, scaling, and fundraising.

Implications for Artists and Creators

Musicians and creators have watched their work consumed without compensation or consent. Regulation and lawsuits have not yet provided a comparable remedy. Data poisoning offers these creators protection and control over their work.

Data poisoning is not a complete solution, and it does not constitute a victory. It is a meaningful constraint on unauthorized use, and it represents a starting point.

The Lost Open Internet and Extraction

Barlow's 1996 declaration envisioned a free internet, but peer-to-peer communities like Soulseek are now being harvested by AI companies for training data.

  • John Perry Barlow published A Declaration of the Independence of Cyberspace in 1996, arguing the internet should not be governed by physical-world rules.
  • Soulseek launched in 2000 as a music-sharing site for music lovers, offering rare vinyl rips, underground electronic music, and lossless audio.
  • As of 2026, Soulseek is still active and possibly more active than ever.
  • Sites like Soulseek contain thousands of hours of high-quality lossless FLAC audio, which AI companies want for model training.
  • AI companies are silently scraping these sites and downloading files without compensating the original musicians.

Poisoning the Water

Data poisoning is introduced as a way to contaminate raw music data so AI scrapers cannot take it without introducing doubt.

  • The speaker uses a river metaphor: clean water that AI companies could drink, unless someone poisons it first.
  • The video's stated subject is data poisoning and how ordinary people use it to fight AI companies and protect artists and musicians.

Mr. Daniels and the Homer Simpson Poisoning

Mr. Daniels, a 25-year-old from England, replaced vocals in over 2,000 records with Homer Simpson's voice and uploaded them to Soulseek with original metadata intact, targeting AI scrapers that rely on metadata.

  • Mr. Daniels is 25 years old and from England, and enjoys tinkering with AI and deep fakes.
  • He stripped original vocals from his library of over 2,000 records and replaced them with Homer Simpson's voice.
  • He kept metadata, file names, artist tags, and album information unchanged so files looked legitimate.
  • AI scrapers download files and read metadata without listening, so they file the audio as legitimate.
  • Mr. Daniels admitted some guilt that genuine human users may be caught in the crossfire.

Historical Precedents and Lawsuits

Data poisoning has historical precedents, including Operation Mincemeat and record labels flooding file-sharing networks, and today it targets corporations rather than individuals.

  • During World War II, Operation Mincemeat intentionally fed the Nazis false information.
  • In the early 2000s, major record labels flooded peer-to-peer sites with corrupted, poor-quality audio files to frustrate users into buying albums.
  • The modern difference is that poisoning targets AI corporations rather than individuals.
  • Spotify and some major labels reportedly banded together to sue AI companies for an approximate $13 trillion, claiming they downloaded 99.6% of commercial sound recordings worldwide.

Audio Watermarking and How It Works

Audio watermarking embeds hidden signals inaudible to humans but readable by AI, corrupting what models learn.

  • Companies named as working on this approach include Harmony Clock, Poison Pill, and Synth ID.
  • Human hearing is roughly 20 to 20,000 Hz; standard audio formats represent frequencies up to around 22,000 Hz.
  • Encoding data up to 300,000 Hz is theoretically possible but would produce enormously large files.
  • Hidden data in ultrasonic bands would be destroyed by MP3 compression.
  • Most practical watermarking embeds subtle signals within the audible range using psychoacoustic masking.
  • AI models process the full mathematical representation of audio, including parts humans cannot perceive, so they absorb the poison.
  • Poisoned training can make a model produce incorrect output, such as generating piano or Mozart when asked for rock and roll.
  • The speaker describes the result as weaponized noise that makes AI unreliable and less valuable.

Does Data Poisoning Work?

Data poisoning is early-stage and not a decisive solution, but it adds friction that forces AI companies to spend resources verifying data.

  • The tools are young, still being developed, and do not yet run locally, which limits adoption.
  • The creator states nobody claims data poisoning will stop the AI wave or bankrupt companies like OpenAI.
  • Its value is creating doubt and a speed bump, forcing companies to spend money, time, and resources verifying data cleanliness.
  • The speaker frames data poisoning as offering protection and agency to artists where regulation and lawsuits have not.
  • The speaker credits Ben Jordan's YouTube video, which covers data poisoning in greater technical depth and interviews creators of Harmony Clock.