Faceless AI-generated YouTube channels have become a problem as of late, so one tech YouTuber took it upon themselves to poison the data they are trained on.
In her latest video (via Ars Technica), f4mi has shown off a way to mess with the data that AI is trained on, specifically the subtitles of YouTube videos. Effectively, these bots can scrape all the information in the video by accessing the data in subtitle files, which are then synthesised with all the other YouTube videos it has taken data from.
Someone can then instruct that bot to create a new video in the style of other videos. The tools aren't quite advanced enough to animate a lifelike face and the creators won't put their own face on it. That is where the "faceless YouTube channel plague" (as popular YouTuber Kurtis Conner describes it) comes from.
"All this AI slop that's appearing on every social media platform is not made by robots trying to steal our jobs," says f4mi, "it's made by humans trying to make money using AI to launder other people's work."
A little while back a life hack popped up online. Job applicants were putting positive keywords in white between margins in document folders to trick AI bots into thinking they were more employable. The poisoning method that f4mi invented is sort of like this.
It grabs the standard subtitle text for a video, adds gibberish in between the lines, makes it invisible to the average user, and then feeds it back to YouTube. She did this by putting works in the public domain into her files and using a tool to make synonyms for many of the words so the AI doesn't spot what she's doing. A Python script simply went through and made these changes.
Unfortunately, tools that can transcribe audio get around this but a few large LLMs like ChatGPT can't. Instead, the AI would 'hallucinate' inventing ideas about the video based on what it can grasp from the enormous amount of text in its subtitles.
And this is where the '.ass' subtitle format comes from. It can be used for standard subtitles but also for animation and graphics, which allows the level of customisation needed to sneak in all of that pesky extra data. Entire books can be hidden with the file type's position tool to make both the size and transparency of unneeded text zero.
.ass is not among the supported files for YouTube videos—but you can convert an .ass file to an SRV3 file, which will then keep all of the data from your original file.
What is artificial general intelligence?: We dive into the lingo of AI and what the terms actually mean.
Unfortunately, f4mi's video had problems with phones crashing "due to the subtitles being too heavy". YouTube has regional subtitle files you can activate, which allows users to account for cultural differences in phrases and tone, so the poisoned track only exists in the UK subtitle track for now, which is the original .ass file. f4mi has uploaded standard files for Australia, which means AI summary tools are getting around the poisoned track and to an untouched one.
This means that plugging the video into an AI summary tool now works. However, as a proof of concept, this is a very smart workaround that means users wanting to steal your videos may have to do it the old-fashioned way, which might just be enough to deter many.
f4mi is one of a long line of YouTubers (like Drew Gooden and Hank and John Green) who have been fighting against AI plagiarism on the platform, and likely won't be the last. It's unlikely the next major anti-AI effort will have quite as good a file name attached, though.
Anthropic has developed an AI 'brain scanner' to understand how LLMs work and it turns out the reason why chatbots are terrible at simple math and hallucinate is weirder than you thought
Studio Ghibli AI image trend floods social media, cheered on by OpenAI and denounced by critics as an insult to Hayao Miyazaki
