AI generated YouTube videos

Published on 12/04/2023, 1629 words, 6 minutes to read

I like having YouTube videos and music on while I work. As an ADHD-haver in the acceptance phase, I like having them on in the background so that I can feel like someone else is being productive near me. This makes me more productive which enables you to be able to enjoy the fruits of my productivity.

Today while I was writing and editing for work, I stumbled upon this video:

<Cadey>

EDIT(2023-12-13 19:09): The video has been privated. I have uploaded my copy of it for posterity.

It's called "The Lost Floppy Disk that Beat Hackers for 30 Years". The title seemed a bit off, but the algorithm has taken me to weirder sides of video essay YouTube before and I've survived so I figured I'd let it run its course. I like watching videos from smaller creators. It helps me get ideas for my own video projects.

I'm pretty sure that a lot of the video on this YouTube channel is AI generated. I don't have any hard evidence, but there's a lot of subtle things that are either the sign of a non-native English speaker or the output of a large language model.

<Cadey>

This stuff is really hard to judge and I really don't know if I'm correct.

I don't want to sound like I'm gatekeeping here, but if the author of these videos is writing them partially or completely with AI tools or using AI text to speech models, I think that it would be more honest to be open about it. It's very frustrating as a user to be listening to something, get interested in it, and then find out it's AI generated garbage.

I've shown this video to a few other people and they also think it's AI generated, but we don't have any hard evidence than the vibes are off. I was able to find an article from Vice that talked about a lot of the similar things in this video, but I'm genuinely not sure

<Mimi>

Meta note: Greetings, I'm Mimi the catgirl! Typically, my delightful chats serve to personify large language models. It's a bit of a mystery whether The Black Tech uses these for more than just giving voice to text in their videos. Should we be correct, all quotes from The Black Tech's videos will be put in my boxes. I generated these transcriptions through the magic of OpenAI's Whisper model. You can read the transcript of The Lost Floppy Disk that Beat Hackers for 30 Years here on tclip.

Content generated by AI using the model gpt-4-1106-preview.

Here's a few examples that make me think this is the case.

<Mimi>

Not everyone has had the chance to see [a floppy disk] in person, it's simply a small, colorful device made of plastic and is quite fragile. Even touching the internal disk with just a finger can damage it. They were usually sold with limited capacity, like two or three megabytes. Often piracy became more of a necessity than a whim. In the 80s and early 90s, programmers faced limited resources.

Content generated by AI using the model ChatGPT???.

<Aoi>

...I was born after USB drives took over floppy disks but even I know that a 3.5" floppy disk can only store 1.44 MB of data. What the heck format can store three megabytes? I guess that could make sense if you added compression into the mix, but this does not add up.

<Mara>

I've seen some references to XDF floppy disks, but I don't think that they gained much popularity over 1.44MB high density floppy disks.

<Numa>

Who knows, maybe someone out there in the audience is a proper greybeard and can call us out on our shit in the Hacker News comments.

There's also a sudden jarring leap from the existence of physical storage as a concept and piracy in general. Yes floppy disks were used for piracy, but so is the internet. It makes sense that the dominant information transfer mechanism of the time (physical media) would be the dominant way to copy pirated data around.

<Mimi>

Piracy is a well-known problem, acknowledged by video game companies, that have tried to address it in every possible way throughout history. For instance, in recent times we have witnessed the emergence of so-called anti-piracy screens, used by various companies, including Nintendo. These screens, as the name suggests, appear when attempting to run an illegal copy of a video game. They often prove annoying and in many cases make playing the game practically impossible.

Content generated by AI using the model ChatGPT???.

Here's the video feed for this portion of the work:

It's worth noting that the "anti-piracy screens" come from a special effects meme where people created fake anti-piracy screens in video games. I'm pretty sure the Mario Kart DS one comes from this video.

Nintendo did have some pretty advanced anti-piracy logic such as with the changes that it made to Earthbound. However after a few easy to bypass mechanisms, they just made the game say "You can't grasp the true form of Giygas' attack!" then proceeding to wipe your save data and resetting the game.

<Aoi>

Huh, what's with that voice? It has such a bizarre cadence to it. I'm not sure if the text is wrong or the voice is wrong, but something is just off.

Here's a recording of me speaking that bit of the script aloud:

Except this isn't actually me recording this, because I tried a few times and found it impossible due to the obvious grammatical errors that I kept subconsciously correcting. That's a model from Eleven Labs trained on me making a mocking basic bitch voice with stereotypical nonsense about pumpkin spice lattes.

Then there's the completely technically inaccurate nonsense from the video, such as:

<Mimi>

These disks have a special code that tells the computer how to load data from the disk. This code is called the bootloader. The bootloader explains how to read data and where to put it in the computer's memory. Usually this process is simple and efficient, but in Gumball's case it's incredibly complicated. It's as if it had infinite numbers and indecipherable codes.

Content generated by AI using the model ChatGPT???.

There are many problems here. Well technically yes the bootloader pedantically DOES enable the computer to understand disks, this is a vast oversimplification of what is going on. Earlier in the script it was attempting to talk about disk formats (it was a wild west of various floppy disk data formats before the industry settled on cloning whatever MS-DOS did) and how it's difficult to copy data out of the disk for this exact game, but overall the attempt at simplifying down details ends up getting into misinformation.

This is what really made me think it's AI generated. I don't mean to be gatekeepy here, but if you had done your research about this to enough of a level to communicate what a bootloader was, I'm pretty sure you'd put it as something like:

<Cadey>

Normally Apple-formatted floppy disks have a fairly standard encoding that is easy to trace through, but Gumball had an incredibly complicated and custom format. Reverse engineering it was a nightmare because it felt like an indecipherable code.

I watched some of the other videos from this creator and they also followed much of the same pattern: mildly coherent information at a glance that completely falls apart under close inspection. There's a lot of comments on this person's videos that seem to think they are some unknown gem in the YouTube space and that's really kinda sad to me.

It's especially frustrating as someone trying (and failing) to get into video production to watch this happen. This creator doesn't have very many subscribers and views; But if a process was fully automated to make this kind of mildly passable drivel be done at scale through the use of vector databases and large language models with large context windows, it could put content mills to shame with the sheer amount of bullshit it could pump out.

Maybe this is an AI-washing of this post from /r/hobbydrama. It seems to follow a lot of the same structure of this post but there's more information in it that is probably hallucinated by ChatGPT.

The horrifying part is that I know almost exactly how you would go about doing this. This is the future of mass video entertainment, isn't it.

<Cadey>

I wouldn't have analyzed this video this critically if I hadn't seen this excellent video about plagiarism by hbomberguy. If you have four hours to kill, it's great fun the entire time.

Maybe that draft of my Tears of the Kingdom review I've been working on is in fact good enough and I don't need to rerecord the audio.

Facts and circumstances may have changed since publication. Please contact me before jumping to conclusions if something seems wrong or unclear.

Tags: ai, chatgpt, youtube