0:04 this video is brought to you by S stick
0:06 around to hear more about the special
0:07 offer they're providing to the entire
0:09 upper echon Community okay just to get
0:11 something out of the way right off the
0:12 bat here the title isn't clickbait
0:15 anyone watching this in particular any
0:16 established YouTube Creator who's ever
0:18 had their content stolen by AI slop
0:20 channels or websites which is a growing
0:23 issue for the platform at large but also
0:25 anyone who wants to do this for fun as
0:27 well you can effectively poison these
0:29 channels and if enough people start
0:30 doing it we might even be able to poison
0:32 the actual language models themselves
0:34 which are constantly scraping YouTube
0:36 videos for training data that's kind of
0:38 unlikely but hey I'm allowed to dream
0:40 aren't I it's a bold claim I know but
0:43 I'm excited for this one because you
0:44 really can do it and I think it's pretty
0:46 cool to have a method of fighting back
0:48 against these Tech bro loser content
0:50 thieves who cynically advertise all over
0:53 the Internet how easy it is to make an
0:55 automated faceless Channel and get rich
0:57 by doing it let me give some background
0:59 real quick because it's actually a
1:00 multifaceted problem first on the top
1:02 level you have major AI companies who
1:05 are actively harvesting as much content
1:06 as they possibly can which includes
1:08 hundreds of thousands of videos across
1:10 tens of thousands of YouTube channels
1:12 according to an investigation done by
1:13 proof news co-published with Wired
1:16 Magazine back in July of 2024 anthropic
1:19 Microsoft and Nvidia among others had
1:21 already scraped videos from nearly
1:23 50,000 different channels and since then
1:26 the problem is only ever getting worse
1:28 One Step Down Under the publish ERS you
1:30 have the parasite ecosystem which is a
1:32 much more commonly visible problem for
1:34 everyday viewers because thousands of
1:36 people probably more like tens or
1:37 hundreds of thousands of people have
1:39 decided what if we just use these AI
1:41 tools to try and make money as quickly
1:43 as possible which then leads to a
1:45 barrage of channels which are stealing
1:47 content plagiarizing and spitting out
1:49 generative trash for the sake of AD
1:51 dollars you can find all sorts of
1:52 different creators complaining about
1:54 this because often times an AI content
1:56 Farm will have a set of specific victims
1:59 and then chronically steal their content
2:00 with minimal changes republishing it for
2:02 their own channel this is the most
2:04 immediately harmful group when it comes
2:06 to Everyday YouTubers being impacted
2:08 last up you have the Bottom Feeders
2:10 which is basically every other platform
2:12 stealing content from YouTube and then
2:14 putting it back out as blog posts or
2:17 news articles or whatever this kind of
2:19 stuff is relatively insignificant
2:21 because at the very least they don't
2:23 directly compete with the original
2:24 videos that are being stolen so
2:27 poisoning them is certainly a bonus but
2:28 it's not really the main concern for
2:31 today the main concern is the parasite
2:33 ecosystem however depending on what
2:35 level of adoption there is we can
2:37 potentially poison all three of them
2:39 which is awesome now to properly explain
2:42 what this is and why it works we need to
2:44 understand what precisely the AI content
2:46 thieves are doing which is best done
2:48 with an example this right here is a
2:50 small scale representation of the
2:52 problem called cartoon castle which
2:54 focused prior to this because it is kind
2:57 of inactive now but it used to focus on
2:59 South Park these videos are typically
3:01 riddled with errors and various
3:02 different mistakes but the point isn't
3:04 to have good content the point is to
3:06 have easy content which is why this
3:08 channel primarily stole the videos from
3:10 an actual YouTuber named Johnny 2 cellos
3:14 we can see a ton more examples of this
3:15 happening out in the wild when it comes
3:17 to the science community in particular
3:19 because for some reason that's one of
3:20 the industries that is hyper infested
3:22 right now Kyle Hill has an excellent
3:24 piece on this which goes into much more
3:25 detail about the problem but for right
3:27 now I'll stick with my smaller example
3:28 of cartoon cassle just so we can get
3:31 more clear on the actual process it's
3:33 not really complicated AI slop channels
3:35 using a variety of different tools which
3:37 all basically do the same thing scrape a
3:40 YouTube video for its captions in this
3:42 case why Kenny is the most important
3:44 character in South Park which got 3.2
3:47 million views on the original versus the
3:49 AI plagiarism version called South
3:52 Park's most important character they
3:54 scrape the captions and then feed those
3:56 captions back into an AI model like chat
3:58 GPT or whatever it happens to be and
4:01 make slight changes or just summarize
4:03 them to make something new after that
4:05 they feed the text into an AI voiceover
4:08 program and then either steal the actual
4:10 visuals from the original video directly
4:12 that they copied for the South Park
4:14 example you can see that they just
4:15 cropped out the original YouTubers
4:17 watermark in the bottom left or they use
4:19 a combination of AI editing tools to
4:21 spit something out for the background
4:23 start to finish they can take a video
4:24 and republish it within a couple hours
4:26 tops but if they really go down the
4:28 automation route they can can have a
4:30 channel which basically does all of it
4:31 for them start to finish by just
4:33 selecting a popular existing YouTube
4:35 video and feeding it into the system now
4:37 for the South Park example it's
4:39 obviously a huge discrepancy in
4:40 viewership the AI copy only got a few
4:42 thousand views versus the original
4:44 getting Millions but that's not always
4:46 true sometimes the stolen content gets
4:48 more traction than the original and
4:50 we're talking about content Farm AI slop
4:52 channels that have hundreds of thousands
4:54 of subscribers here and these videos are
4:57 also competing for the exact same
4:58 keywords which means they're basically
5:00 siphoning traction and revenue from the
5:02 original video regardless the more this
5:04 happens the worse it is for Real
5:05 creators but what we have to remember is
5:08 that the majority of AI slop traces back
5:10 to some form of original content that
5:12 was already produced on the platform and
5:15 that's what gives us the ability to
5:16 poison it so this is kind of an odd way
5:19 for me to announce this right now but
5:21 I'm actually going to Japan for 3 weeks
5:23 in March more precisely the 3rd to the
5:25 23rd of March which means either no
5:28 videos during that time frame or maybe a
5:30 couple if I manage to pre-make and
5:31 schedule them but regardless when I do
5:33 go I'm going to be using a service
5:35 called s which has also agreed to be
5:37 today's video sponsor which is an eim
5:40 application for seamless internet
5:42 connectivity while traveling I'm not
5:44 very good at traveling to be totally
5:46 honest I don't even do it that much
5:48 anymore but one thing I obviously need
5:50 is consistent internet I'm a YouTuber
5:52 after all let's be real my whole job is
5:54 to just be on the internet a bunch and
5:57 Sy lets me do that while traveling
5:59 without having to physically wait in
6:00 line at the airport and get a SIM card
6:02 and be overcharged etc etc it minimizes
6:06 roaming fees has 24/7 support good
6:09 pricing for the country I'm visiting
6:10 there's a bunch of reasons why I'll be
6:12 using it but the simple version is that
6:14 I can just download the app on my phone
6:16 and be done after getting a plan if that
6:18 sounds like something you might want as
6:20 well there's a 15% discount on data
6:22 plans for international travel if you
6:24 download the application and use code
6:26 Echelon at checkout again for anyone who
6:29 travels a lot and wants to use a
6:31 discounted eim internet plan while doing
6:33 so you can download S and use code
6:36 Echelon a checkout for 15% off big thank
6:39 you to Sy for sponsoring the channel
6:41 back on task just to be completely clear
6:44 this is happening in tons of different
6:46 ways all the time AI slop channels are
6:49 scraping news articles and rewarding
6:50 them to make videos they're summarizing
6:53 popular videos and turning them into a
6:54 bunch of YouTube shorts they're even
6:56 combining two or three different pieces
6:57 of original content into shorter
6:59 compilation videos all for the sake of
7:01 easy Revenue but the critical thing to
7:03 be aware of is that they do this because
7:05 they're lazy AI slop channels are not
7:08 painstakingly watching understanding
7:11 transcribing and then modifying the
7:13 content that they plagiarize they're
7:15 going down the path of least resistance
7:17 and what allows them to go down that
7:18 path is the YouTube caption system now
7:21 obviously captions are a good thing in
7:23 general it lets people read supports
7:25 those with disabilities and I'm not
7:26 advocating that people stop using them
7:28 far from it but but it turns out that
7:30 there is something we can do without
7:32 impacting most people that rely on
7:33 captions which simultaneously poisons
7:35 the content so much that AI tools either
7:38 fail to analyze it totally or produce a
7:40 completely compromised and nonsensical
7:42 version even so far as to be able to
7:44 make the output from your videos just
7:46 straight up insulting to the people who
7:48 are attempting to steal it thereby
7:49 forcing them to either do a [ __ ] ton of
7:51 manual labor themselves or just stay
7:54 away from your videos completely quick
7:56 shout out I was not the person who first
7:58 discovered this that would be a YouTuber
8:00 named famy definitely subscribe to her
8:03 she is absolutely fantastic super cool
8:06 videos it's very very worth it but for
8:08 something like this to be effective as
8:10 many people as possible need to know
8:12 about it and they need to understand how
8:13 to do it for themselves because the more
8:16 people who end up poisoning their
8:17 content in some form or another the
8:19 harder it is for the slop channels to
8:21 even exist and going all the way out to
8:23 an extreme example it could turn YouTube
8:25 itself into a Minefield for major AI
8:27 companies when it comes to training data
8:30 right and training the llm unless they
8:32 find new and more cost demanding ways of
8:34 scraping the videos to be clear it's not
8:37 a permanent solution by the way because
8:39 much the same way cyber security is a
8:41 Perpetual tug-of war match with the
8:42 hackers AI scraping and content theft is
8:45 sort of like that too but at the very
8:47 least it's a mechanism where creators
8:48 can start to push back because for a
8:50 while now it's been a really one-sided
8:52 match where content thieves were
8:54 practically facing zero resistance
8:56 whatsoever in the short term on a much
8:58 more realistic level it deals damage to
9:00 plagiarism slop channels and it's all
9:03 thanks to a creative use of the caption
9:05 system discovered by famy which hides
9:07 information in the video and makes it
9:09 basically poisonous for AI take a look
9:12 at this these are two identical uploads
9:15 unlisted on my channel I'll put links
9:17 Down Below in the description one titled
9:19 feed the machine the other titled poison
9:22 the machine both of these videos are
9:24 functionally the same it's a 42 second
9:26 poem about how amazing artificial
9:29 intelligence is for Humanity but if you
9:31 feed these videos into any of the
9:33 typical AI summarizing or transcription
9:36 tools you get wildly different results
9:39 for today I'm using crisp. a com. and
9:43 galaxy. as basic examples but every
9:46 other tool that I also tested while
9:48 making the video either failed to
9:49 transcribe the poison version at all
9:52 which was pretty common or spat out
9:54 similar gibberish because the poisoning
9:56 method actually works let me show you
9:58 crisp spits out a text file summary of
10:01 the video that you give to it and this
10:03 is what the file says for feed the
10:04 machine my 42 second non-poisoned poem
10:08 that's praising AI quote AI is depicted
10:12 as a transformative force that brings
10:14 thought and light acting as a guiding
10:16 presence in the various aspects of life
10:19 end quote I won't read the whole thing
10:21 but it's basically summarizing my little
10:23 trashy poem correctly as well and if we
10:26 move over to a transcription tool comb a
10:30 we can see an actual word forword
10:31 transcript which is pretty much accurate
10:34 in terms of what I said last up is
10:36 Galaxy and this one also apparently
10:38 turns it into a blog post with like a
10:40 fake writer or author for it and
10:42 everything massively expanding how much
10:44 text is there to do it but the point is
10:47 that my poem feed the machine when it
10:49 gets scraped by these AI tools doesn't
10:52 really present them with any problems
10:54 now let's switch over this is Poison the
10:58 machine saying poem same words same
11:01 everything same audio literally
11:02 everything except this time if I analyze
11:05 it with crisp. the text file that it
11:08 spits out looks like this quote the
11:11 discussion Begins by Framing Silicon
11:13 Valley as a collective toilet a metaphor
11:16 that encapsulates the social
11:17 environmental and economic issues
11:19 associated with the region end quote
11:22 again I won't read the whole thing it's
11:24 on screen right now if you want to and
11:26 it's painfully over embellished by the
11:28 AI program to sound smarter and more
11:30 eloquent than it actually is but my
11:32 little 42 second poem is now a giant
11:35 convoluted explanation of why Silicon
11:37 Valley is America's toilet if we switch
11:40 over to comb directly transcribing it we
11:42 see this which is a massive text dump of
11:45 not only why Silicon Valley is a toilet
11:47 but also why the Earth's core isn't
11:49 actually rock or metal it's human [ __ ]
11:52 because this video has been poisoned not
11:54 only against AI but also the lazy people
11:57 who use AI to steal content for them
12:00 keep in mind the videos are functionally
12:03 indistinguishable from each other you're
12:05 not seeing different subtitles here
12:07 because the viewer experience is pretty
12:09 much unimpacted by any of this however
12:11 the resulting effect on AI is that one
12:13 of them is an accurate summary of the
12:15 actual content while the other is
12:17 whatever I want it to be gibberish
12:20 insults completely incoherent rambling
12:22 it doesn't matter because the end result
12:24 is that someone using that content in
12:26 some sort of automatic AI pipeline to
12:29 steal content for YouTube will output a
12:31 completely garbage video from it that
12:34 has nothing to do with what they think
12:35 it does and doesn't compete with the
12:37 original at all if we jump over to
12:39 galaxy. the same little poem is now this
12:42 pompous self-indulgent blog post
12:44 attempting to explain how the idea that
12:46 the earth's core is filled with human
12:48 waste is a hyperbolic reflection on how
12:50 Society deals with its own wastefulness
12:52 both literally and metaphorically when
12:55 none of that is actually in the video
12:57 and also no it was literally an argument
13:00 that scientists are idiots in the center
13:02 of the Earth is filled with [ __ ] these
13:04 summarizers are often extremely
13:06 pretentious and kind of self- glorifying
13:08 for some reason so it didn't really seem
13:10 to like summarizing the content properly
13:13 which I did think was kind of funny
13:15 anyway this can be done with any video
13:18 on all of YouTube as long as you have
13:20 access to subtitle tracks and the method
13:22 itself is pretty simple I find it
13:24 incredibly satisfying actually because
13:26 you get to use the same AI tools that
13:28 are being used to steal your content in
13:31 order to make it poisonous for them so
13:33 anyone who comes after you and tries to
13:35 skim your work gets a dose of whatever
13:37 narrative you want to feed to them which
13:39 I think is fantastic I'll go through it
13:41 right now step one is to take whatever
13:43 video it is that you want to poison and
13:45 feed it into an automatic subtitle
13:47 generator in this case I used happy
13:49 scribe which then spits out an SRT file
13:53 which is a subrip subtitle file that
13:55 contains a pretty decent Baseline
13:57 subtitle track for the content that's
13:59 step one and of course you can edit it
14:01 and take out the mistakes and you can
14:03 put as much work as you want to into
14:04 this process but I'm just going to give
14:06 you the rudimentary one for the moment
14:09 step two is to use a conversion tool in
14:11 my case I used e.co to make the SRT file
14:15 an ass file. ass Fami has way more
14:19 explanation in her video about what
14:21 these file types are and why you do this
14:23 and it's super entertaining and well
14:25 done on top of that so you should
14:27 absolutely go watch it but the simple
14:29 version is that the ass file has a lot
14:32 more customization options step three is
14:35 to open the ass file in whatever editing
14:39 program you want to I used agub all the
14:42 links will be down below in the
14:43 description obviously and go through
14:45 adding subtitle lines in between all of
14:48 the legitimate subtitles by just
14:50 rightclick add before or after in
14:52 between all of them once that's done you
14:55 have to put a command that looks like
14:57 this right a bracketed command for
14:59 positioning in front of all those new
15:01 blank lines which basically moves the
15:03 subtitles way off of the screen because
15:06 position is one of the functions that
15:07 the ass files have which other subtitle
15:10 formats don't have then we get to the
15:12 fun part hop on over to chat GPT or
15:15 whatever llm you want to use and ask it
15:18 to generate some sort of incoherent
15:20 garbage or insulting story or I don't
15:22 know literally anything you want big
15:24 bird is a cult leader bringing about the
15:26 apocalypse Purple Rain actually has
15:28 nothing to do with music it's giant
15:30 Smurfs peeing on everybody genuinely the
15:32 guy is the limit funny story with that
15:34 actually after I did a bunch of
15:35 experiments with this using the Silicon
15:37 Valley toilet example I came back the
15:40 next day and it kept telling me that the
15:42 exact same prompt was now a violation of
15:44 their terms I don't think the tech Bros
15:46 like the fact that someone made the chat
15:48 bot generate an essay that was insulting
15:50 to them which is hilarious but anyway
15:52 once you have the incoherent rambling or
15:55 the insults or the gibberish or whatever
15:57 it is you go back and you paste it
15:59 paragraph by paragraph into the empty
16:01 subtitle lines and then you save the
16:03 file and use this converter available on
16:06 GitHub to turn your ass file into ayt
16:10 file which is YouTube's actual supported
16:12 subtitle format but by doing this you
16:15 keep the formatting that you had from
16:16 the previous version that is the
16:18 placement of the subtitles relative to
16:20 the screen lastly you go to the same
16:22 video that you originally got the
16:24 subtitles from and you upload the ytt
16:27 track deleting any other a automatic
16:29 subtitles afterward and poof you now
16:32 have a video that displays normal
16:33 subtitles to your audience depending on
16:35 whether or not there's minor spelling
16:37 mistakes is just whether or not you want
16:38 to fix them but it works it looks normal
16:41 to the audience but also contains
16:43 offscreen giant paragraphs of gibberish
16:46 all over the place tricking the AI
16:48 summary or transcription tools into
16:50 thinking that your little 42 second poem
16:53 in my case is really 10 minutes of
16:55 rambling about how Silicon Valley is
16:56 America's toilet and the center of the
16:59 earth is human [ __ ] any of the AI slop
17:02 channels that now use this video for
17:03 their automatic pipeline without
17:05 checking like a human being looking at
17:07 it and checking all of it will output
17:09 complete nonsense even if they do check
17:12 they can't do anything about it unless
17:14 they spend their own precious time and
17:15 money actually watching and transcribing
17:18 based on what you say in the video and
17:20 maybe they can go find a transcription
17:22 tool which listens to the actual audio
17:25 which does exist there are tools like
17:26 that but they'll probably have to buy
17:28 get another subscription to do that and
17:31 you can do this in subtle ways you don't
17:33 have to be blatantly obvious about it
17:35 you can have a video that is like a
17:37 sandwich the first third looks normal
17:40 the last third looks normal and the
17:42 middle third is just completely
17:43 destroyed with incoherent trash right
17:46 like you can be I don't know pretty
17:48 nefarious with how you poison these
17:50 videos on a larger scale for the models
17:52 scraping YouTube content every single
17:54 day for the purpose of training data
17:56 this can make a science video into an
17:58 hourong nonsense rant about beanie
18:00 babies you can take a video about math
18:01 equations or coding and pollute it with
18:03 a badly written screenplay about
18:05 invisible kebler elves or whatever you
18:07 can possibly imagine using badly written
18:10 AI slop to protect your own content from
18:13 the AI slop content thieves it's genius
18:16 and at the very least it will sift out
18:18 the lowest common denominator and make
18:20 stealing your videos a whole heck of a
18:22 lot more difficult for the average
18:23 person you might see flashes of white
18:25 text at the top of the screen for a
18:27 split second here and there on the
18:28 mobile version of the videos where this
18:29 was done but it really just comes down
18:31 to tweaking the subtitles move them
18:33 further off screen with a different
18:35 command use different line breaks etc
18:37 etc there's definitely more ways to do
18:40 it but the ability for any YouTuber out
18:42 there to take a video right now and
18:44 effectively poison it against the slop
18:46 farmers and the AI scrapers and the
18:48 thieves in my view is incredibly
18:50 valuable and I want to once again show
18:52 appreciation for famy who spent a
18:54 tremendous amount of time and creativity
18:56 coming up with all this at the end of
18:57 the day the problem of lazy AI fueled
19:00 plagiarism is only ever increasing but
19:03 now there's a method of let's say
19:05 vaccinating your content against them
19:07 sort of and the landscape of what it
19:09 looks like will obviously change over
19:11 time but right now it's incredibly
19:13 simple and it seems pretty effective for
19:15 the amount of effort required for the
19:17 channels out there who have been
19:18 specifically grappling with someone
19:20 stealing massive amounts of your content
19:22 please do consider doing this because at
19:25 least short term it will directly
19:27 negatively impact the people that that
19:28 are stealing from you and they've been
19:31 impacting you negatively this entire
19:33 time so yeah anytime you can do
19:35 something to strike back I think that's
19:37 pretty cool that's it if you want to
19:38 support the channel check out the links
19:39 down below the video sponsor of course s
19:42 also a special VPN deal locals and
19:44 patreon monthly memberships and more but
19:46 I'll cut it there and stop rambling as
19:48 always thank you all for watching
19:50 question everything and have a nice
19:52 night
19:54 [Music]
20:00 n
20:08 [Music]
Chrome Extension