00:00:00
Alright, so it's time to eat, you bust out your phone, you use your glasses as a stand, because
00:00:04
everybody does that for some reason, and you try to find an interesting video to watch.
00:00:08
And, well, you do find something that seems promising, you click on it and...
00:00:13
At first, everything seems normal, but then you quickly start realizing that something is
00:00:20
very wrong with the video. Like, the voice of the narrator sounds a bit too robotic,
00:00:26
and he seems to struggle with the most basic words, and also, the stock footage that's being
00:00:32
used is super generic, and not just that, but it keeps repeating for some reason. And the actual
00:00:38
script itself, like, what the video is about, doesn't seem to really make sense. It's not
00:00:45
getting anywhere. It's like, the narrator is just yapping, starting and stopping, starting and
00:00:51
stopping, without reaching any meaningful conclusion. It's like, no human thought was
00:00:57
put behind this video. And that's because if this happened to you, you've most likely not watched
00:01:05
something that was made by a human at all, but something that was made by a...
00:01:10
So, it is officially 2025, and I can confirm, the sound was wrong. We women are not having
00:01:18
more sex with robots than humans, and that's my roundabout way of saying that AGI is not here yet.
00:01:23
All this AI slop that's appearing on every social media platform is not made by robots trying to
00:01:29
steal our jobs, it's made by humans trying to make money using AI to launder other people's work.
00:01:36
A new grift has appeared on the girlboss side and the alpha male side of the internet. Which
00:01:40
side you're on depends on if you liked blue or pink when you were a kid. And this grift is called
00:01:45
Faceless YouTube Channel, and that's a great name because if I made that sh*t I wouldn't want to
00:01:50
show my face either. The idea is to leverage AI tools like ChatCBT to essentially outsource all
00:01:56
of the necessary work to make a video to these tools, including voiceovers that they make with
00:02:01
uncanny AI text-to-speech, and since no real work is actually being put in, this is, unfortunately,
00:02:07
extremely effective. More and more often recently you see AI slop being based not on Wikipedia
00:02:12
articles, not on forum posts, but on other YouTube videos. This is done by exploiting YouTube's
00:02:18
automatic subtitles feature, scraping those subtitles and then giving them to ChatBBL,
00:02:22
asking you to either change some words around in case you straight up want to steal a video one to
00:02:26
one, or make a summary in case you want to make a short case of plagiarism. You can easily find
00:02:32
countless YouTube creatives complaining about how their videos were stolen like this, and
00:02:36
the phenomenon doesn't seem to be stopping at all. As Hbomberguy said in that video about that thing,
00:02:41
"On YouTube, if you have an original idea, if it's good, it won't be yours for long."
00:02:47
Automatic subtitles are, in my opinion, one of the best things YouTube ever did, and even though
00:02:53
that means AI grifters can now steal your videos easier, you shouldn't disable them. They are
00:02:58
incredibly useful and some people need them in order to watch your content. Unfortunately,
00:03:03
this incredible feature is being exploited by these AI grifters to steal videos, and we can't
00:03:10
do anything about it. Like, just take the link of this video, okay, and put it on a video summarizer,
00:03:16
like any summarizer of your choice. As you'll see, you're getting no summary for this video.
00:03:22
A few months ago, I had an idea, and like the true scientist that I'm not, I wanted to put it to the
00:03:31
test. And the idea was trying to use the subtitle track in my video to poison any AI summarizers
00:03:38
that were trying to steal my content to make slop. The most basic way to do this would be simply
00:03:43
removing your subtitle track and creating a fake one that is only meant to be seen by AI bots, that
00:03:49
basically only contains yapping, like pure garbage. However, this was unacceptable for me. Despite it
00:03:57
working, I didn't want to do this because I really care about having proper, well-formatted
00:04:03
subtitles in my videos. They are very important to me. So after some experimenting and iterating,
00:04:08
I figured out a way to both have working subtitles for you guys, like in this video,
00:04:14
but also in the subtitle data, hide garbage, like pure nonsense that is only visible to AI trying to
00:04:23
steal my content. And in this video, I'm going to teach you how to do it. But first, this video
00:04:28
might make me some enemies, so just to be safe, give me one minute to pay my bills. Just last
00:04:32
year, several internet providers were victims of a massive security breach, where hackers are
00:04:36
suspected to have gained access to a network infrastructure that ISPs used to answer court
00:04:41
authorized wiretapping requests, meaning that potentially your private data and everything you
00:04:44
do on the internet might have been exposed to a third party. This is one of many leaks happening
00:04:49
seemingly every week. And hackers are not the only problem. Data brokers can legally harvest and sell
00:04:54
your personal data, letting anyone straight up buy it and use it for whatever they want. And that's
00:04:58
why I have partnered with today's sponsor, Aura. Aura monitors your personal data across both the
00:05:02
clear and dark web, alerting you immediately in case of a breach and providing you with up to
00:05:06
5 million dollars in insurance if that data is used to steal your identity. They also provide
00:05:10
you with free all-time alerts, even for breaches that do not directly involve you, an automatic
00:05:14
opt-out from data brokers so your data doesn't get sold to anyone, and a VPN for safe browsing,
00:05:18
potentially protecting you from attacks like the ISP one that I just mentioned. You can go to the
00:05:22
link in my description, aura.com/family, for a free two-weeks trial, meaning you can straight
00:05:26
up immediately check for free if your data was stolen or sold to anyone. Thanks to Aura for
00:05:30
supporting this crazy video right here. And now, back to destroying Skynet, I guess. Around one
00:05:35
year ago, one of my friends sent me a link to this website. I don't know who the f*** Toolify AI is,
00:05:41
but they just stole my video. They clearly just gave my video to a YouTube summarizer and published
00:05:46
a summary as their own article on their website. Now, instead of like tweeting about it and just
00:05:51
getting the content removed, I wanted to try doing something more interesting. And that is,
00:05:56
what if I can make it so that someone trying to do the same thing with one of my videos in the
00:06:01
future is going to waste their time and money because the subtitles are a lie. I mean, they're
00:06:06
real, but there's something in them. After a bit of experimentation, I figured out a way that works
00:06:10
with most LLMs. It doesn't work if someone is using Whisper AI, which is transcribing my video
00:06:16
based on audio and then they give that transcript to ChatGPT. But most people trying to steal stuff
00:06:22
wouldn't bother with that, they're just going to Google "video summarizer" and use that to steal
00:06:26
their stuff. And so, to start, we need to talk about...
00:06:29
Advanced Subtitle Alpha is a subtitle format released in 2002, technically being the fourth
00:06:36
version of the SSA subtitle format that was originally launched in '96. And it was originally
00:06:42
created by Cotus, a British programmer and anime fansubber, as the fourth format used for his own
00:06:47
fansubbing software, Advanced Subtitle Alpha. Now, if we compare the ASS format with the SRT format,
00:06:53
which is basically the standard when it comes to subtitles today, S clearly wins here. SRT was
00:06:59
originally launched in 2000 as part of SubRip, which was a software that would use OCR to scan
00:07:05
hard-coded subtitles from video and convert them to scripts. And because of this very narrow original
00:07:12
scope, SRT files have a very simple structure. When you open one with Notepad, you notice that
00:07:17
every subtitle is made of three parts. The sequence number, the timecode telling the player
00:07:22
when the subtitle should appear and disappear, and finally the actual text itself, the thing you see
00:07:26
on screen. This is very basic but also very clever, and it works, but it is no match for the ASS.
00:07:32
S also gives you fonts, positioning, effects like shadow, bold, italic, underline, karaoke, animations,
00:07:39
heck, you even get multi-line styling so you could have different styles in the same subtitle line.
00:07:45
S is how I managed to get color subtitles in my Format Wars video. However, that shouldn't have
00:07:51
worked. YouTube allows you to upload subtitles in different formats, but S is not one of them.
00:07:58
There are some other compatible subtitle formats that allow you to get different degrees of
00:08:02
customization, but none of them gives you access to every single one of these features at the same
00:08:07
time like S. However, after you upload your subtitles to YouTube, it doesn't matter what they
00:08:12
were because they internally get converted to YouTube's own proprietary format named SRV3, or
00:08:18
YouTube Time Text, YTT, which has already been reverse engineered, and you can find a few GitHub
00:08:24
projects that let you convert your S files to YTT, and YouTube is going to accept those just fine.
00:08:30
You can just upload them, and even though the styling doesn't show in the subtitle page,
00:08:34
after you save the subtitles and you go to the actual video player, they work.
00:08:39
So let's take one of my classic videos, okay, the Homebrew Channel Music one.
00:08:44
I've already made subtitles for that video ages ago, and they were in the SRT format.
00:08:48
When I convert them to the ASS format, what I get is this suit of options of like new stuff that I
00:08:55
can do. And there are two things that I'm very interested in, subtitle position and styles.
00:09:00
The way YouTube summarizers usually work is by scraping the subtitles from a YouTube video
00:09:06
and then giving those to an LLM like Chat NES, asking it to make a summary.
00:09:10
So the LLM then takes a look at the subtitle file in order and tries to explain what's going on
00:09:16
inside. So what if exploiting the ASS format? We add for every real subtitle line that a human is
00:09:24
supposed to be able to read, two chunks of text out of bounds using the positioning feature of
00:09:29
the ASS format, with their size and transparency set to zero so they are completely invisible.
00:09:34
And to avoid repetition, which is something that an LLM can easily figure out, like it can
00:09:39
understand we are trying to trick it, instead of putting random words, we actually copy paste
00:09:43
works from the public domain. And for extra measure, we replace most words there with synonyms.
00:09:50
Now doing this manually would be a pain in the ass. So I just made a Python script that does
00:09:54
this for me and it kind of works. It spits out an S file that I have to then open using AggieSub to
00:10:01
then modify the styles to make sure that the ones out of bounds are also invisible and zero pixels
00:10:06
big. I can then finally use this other tool called YouTube Sub Converter to convert the S
00:10:12
file into a YTT and upload it to YouTube. And as you can see, the subtitle page is a f***ing mess
00:10:20
now, but when I actually reach the video page and I enable the subtitles, everything seems fine.
00:10:27
All right, so before we go any further, I want to go here and remove... Okay, there is no automatic
00:10:35
captions track yet, so I should just be able to go to summarize.tech. I can paste my video and see
00:10:44
what happens now. "The tomfoolery test video presents a thorough examination of the evolution
00:10:51
of mechanical engineering and aviation, emphasizing the complexities of engine designs,
00:10:56
including steam engines and various propulsion systems." That's the garbage data. It's only
00:11:04
summarizing the garbage text. There is no... Okay, wait, here. We have a slight mention
00:11:11
of the Homebrew Channel and then it goes back to the garbage text. Yeah, this is working. This is
00:11:18
working great. "Crisp YouTube Summarizer." Doesn't Crisp, like, make a noise cancelling thing? Why
00:11:27
are they making YouTube summarizers now? "The discussion begins with the necessity of springs
00:11:32
behind the delivery box." This one doesn't even try to talk about the Homebrew Channel.
00:11:40
"Sumcube.ai." Let's try this one. "The effects of Discord on programmer humor were catastrophic."
00:11:56
Oh my god, it's just like... 136.
00:12:03
It is very important here that when the automatic captions get made, we delete them,
00:12:12
because we already have our own track, which is the one that's poisoned for AI. If we have the
00:12:19
automatic one, then the summarizers are going to default to that one, and therefore this trick
00:12:24
won't work. Gemini is actually pretty smart about this. If you ask it to make a summary of a video
00:12:29
and the video doesn't have automatic captions enabled, it doesn't even try. When I first came
00:12:34
up with this method one year ago, I was over the moon. Then I tried opening one of these videos on
00:12:38
my phone and yeah, transparency and position don't really work there. So any reasonable person would
00:12:44
just give up. Do I sound like a reasonable person though? Since the problem is that transparency
00:12:50
and positioning don't work properly on mobile and they show black squares, I decided to write a
00:12:56
Python script that scans the video and finds every instance of a full black frame. So for example,
00:13:03
when I fade to black now, there are like 30 subtitles on screen right now, but they are
00:13:08
black on black so you can see them. I ended up having a local LLM generate a story that is
00:13:13
similar to the real script, but with completely made-up facts, and also threw that in the
00:13:18
out-of-bounds subtitle text repository and that ended up working perfectly. I managed to confuse
00:13:23
GPT-40 every time it tried to recap my video. So I can confidently say that as of today, 22nd
00:13:30
January 2025, this is a pretty effective way of fighting the most common slop makers. I can't
00:13:37
really do a lot about Whisper right now. I have to figure out a way to trigger audio hallucinations
00:13:42
without making them obnoxious to humans. Also, bigger and newer models like ChatGPT-01 are able
00:13:49
to sometimes filter the noise and actually understand the real topic of the video. They
00:13:53
are able to see that I'm tricking them. However, there would be yet another step that could
00:13:58
potentially work by simply filling the memory of any LLM so much, it's simply wasting so many
00:14:04
resources it can do anything about it, and that would be dividing every single sentence in the
00:14:09
subtitle file per single letter specifying the position on screen and the timing so to a human
00:14:16
watching the video they would see the complete result of like this patchwork because the player
00:14:20
can do it easily as part of the logic that displays the subtitles, but an LLM would have to like read
00:14:26
every single letter in order and that's where we pull another trick because since the player doesn't
00:14:32
need the actual subtitles in the text file to be in order, when you're playing the video the player
00:14:37
just loads the entire file in RAM and displays the subtitles according to their timestamp, but we can
00:14:42
scramble the order of the letters in the text file and the video player is gonna be fine because he
00:14:47
can just reference the timing, but an LLM has to waste resources reordering every single letter
00:14:52
for every sentence and then from that it has to piece together the words which depending on how
00:14:58
the scraping is done means doing it without having the position data so it has to like play
00:15:02
scrabble for every single word and eventually if it can do that correctly it can theoretically try
00:15:08
to summarize something but yeah right now it just gives up it doesn't even try. There is yet another
00:15:13
trick this one works also on whisper summarizers sometimes because it is not exploiting any
00:15:19
specific quirk of the tech it's exploiting the economy behind it and that is since running large
00:15:24
good models can be expensive for people running websites like this it is common for them to use
00:15:29
caching so when you give the AI summarizer a link to a video it's going to make a summary and then
00:15:35
store that summary so that any future person trying to summarize the same video is not going
00:15:40
to waste any API credits because they already have the result it's already done so my idea to exploit
00:15:45
this is we make a video that is twice as long as the real video and the second part is just us
00:15:50
yapping saying stuff like "Android hell is a real place" we upload this video to YouTube and using
00:15:55
the YouTube editor we cut out the real part only leaving the yapping and make sure the yapping is
00:16:00
the same length as the real part for reasons you'll figure out soon. Finally give your video to every
00:16:05
video summarizer you can find so that they make a summary of the yapping and they keep that in
00:16:09
cache for the future. When that's done you go back to the YouTube editor you revert the changes and
00:16:14
this time you cut out the fake part you only leave the real part intact and what you're gonna have
00:16:19
now is your copy of the real video on YouTube and for that link associated with it fake summaries
00:16:26
about your yapping. Having the same length for the yapping part and the real part makes it so that it
00:16:30
is more difficult for these websites to figure out that something changed with your video because the
00:16:35
length is the same and so they can't just use that to update their cache. So did I just fix this
00:16:40
problem once and for all? No, not at all. When I first started working on this thing around one
00:16:45
year ago I was aware that the second I made my discoveries public people working at these tools
00:16:52
which by the way they're not to blame don't like hate them the tools are great some people just
00:16:57
use them to steal but they're not meant for that. But yeah the developers behind these tools are
00:17:01
going to fix this issue so my goal with this video aside from telling you how cool subtitles are is
00:17:07
not trying to sell you a cure for this problem. Making this video might realistically have just
00:17:12
closed some doors for me when it comes to future sponsors that maybe work with AI. But this video
00:17:17
isn't about making money or getting recognition it's about trying to do something about what's
00:17:22
going on. Us creators are being attacked by both other creators just stealing our content so they're
00:17:28
not really creators they're just thieves who don't really care about anything except making money
00:17:32
but most importantly huge mega corporations are being caught basically weekly trying to train
00:17:39
their product something they're selling on our art without any authorization they're trying to build
00:17:45
machines that are meant to replace us and they are using our passion to fill this project and I hate
00:17:51
it I hate it I'm a nobody I'm a woman trying to make her living doing what she loves and I'm not
00:17:58
trying to pick a fight or even like pretend I have a fighting chance against huge mega corporations
00:18:04
that in one second make more money than I've ever made in my entire life so realistically if they
00:18:09
want her data if they need it they're going to get it but what I'm trying to say is that we have to
00:18:14
stop being doomers about it and we don't have to make it easy for them to steal our sh*t. If I
00:18:18
manage to come up with this and I'm just a glorified art graduate maybe some of you watching
00:18:22
right now are going to come up with something that is even more complex and can make it less
00:18:27
convenient for people or huge corporations to steal our work without giving us any money any credits
00:18:33
so yeah that's pretty much it and after this video I have to really hope AI never takes over or
00:18:39
I'm definitely going to Android hell.
00:18:41
Nothing to see here nothing to see here there are no hidden subtitles here no.
