I watch a lot of STEAM videos on YouTube. In the past I loved supporting new creators but some time last year I stopped trying new channels that don’t have human presenter, there’s just so much slop. Now I’ve noticed that there are channels where even the presenter is AI. I can spot a few tells (presenter is never on screen for very long, voice pacing doesn’t vary, no ums and ahs, etc) but it’s getting harder and I wish I knew more clues.
One thing that seems to help is checking the channel history (does the presenter age, have they only started making content very recently etc)
Can someone recommend a good, recent guide to spotting AI videos?
Don’t have any recommendations you haven’t mentioned. As a blind person I listen to synthesized speech nearly all day long, so I’m pretty good at catching AI voices. I think the ones trained on IRL people have a “choppy” or “quavering” quality. I’m not an audiophile or computational linguist so I’m sure there’s a term for this I’m missing, but the overall audio has a particular tell to it.
It’s getting harder, though. Now I listen for things like odd prosidy, stuff like odd pauses or inflections or inconsistent pronunciations.
For example, the phrase “So I went to the store” could have a number of inflections that encode info about the ongoing conversation, whether I intend to continue speaking or whether I’m yielding the floor, whether I’m asking a question, indicating surprise, etc. AI is getting better at this but it’s not perfect.
Pronunciation of proper nouns or words that are spelled the same but pronounced differently in different contexts is a good indicator. Just the other day my mom was watching some AI-generated listicle and the voiceover mispronounced the name of a city after correctly pronouncing it every time before.
Thanks for your suggestions! Now that I think about it I’ve definitely heard a few of these tells before. Im sure your sick of these kind of questions but I’m curious about your experiences as a person with blindness - how has AI affected your daily life? (I know that this is a generally anti AI space/thread, but please feel free to speak openly, I’ll be the first to admit that the technology is a double edged sword!)
Disclaimer: I only speak for myself. Blindness does not respect class, creed, or culture, so every blind person will have their own opinions and experiences. In my case AI hasn’t shown much tangible benefit beyond what clever algorithms are already doing, for optical character recognition etc.
Beyond blindness My knee-jerk reaction to AI is strongly negative simply because it’s another locus of concentrated societal power (Who else is there besides Open AI and Anthropic and maybe Groq? IIRC everyone else is just using one of those on the backend. So that’s what, three people, a few dozen if you include the other C-suite folks at each company, pulling a lot of very consequential levers) But that opinion is probably par for the course for your average Lemming.
Interesting reply, thank you for your own unique perspective. Would you ever consider using something like real time image to speech to describe the world around you?
I check not only if the channel is recent but also how often they release a video.AI channels churn them out at quantity real creators don’t. Still not perfect,need more ways to spot them.
There’s a channel I recently got into where every episode was “I watched every episode of [LONG-RUNNING TV SERIES] and here’s what I found.”
They were posting 2 episodes a week.
I find that I can’t catch AI more easily if I’ve seen what it can output.
For example, the older videos used to use generative AI on the images and videos themselves, which made it easier to spot those.
Some newer videos are made with the “agentic” tools that have access to various scripts and animation libraries, which come out looking very polished and visually identical to human produced ones. For those, I find that the AI will put unnecessary “work”, into the irrelevant parts of the visualizations. Some of the libraries and fonts are more popular than others, so you might notice a pattern that way too.
For an example to the second category, check out the recent videos from this channel, which has AI generated “music videos”: https://m.youtube.com/@degenerativepixels
Degenerative Pixels is a channel for music videos made with AI. The songs are generated in Suno, and the animation is written as code by Claude, so every scene is built by a program instead of coming out of a video generator.
It sounds like it’s explicitly not AI-generated video in the sense I think most people would mean?
Also, the irony of being told to “sign in to confirm you’re not a bot” to access a video generated by a bot.
Thanks for the advice and the link. I almost don’t mind AI used for traditional animation, I know its awful and plagiaristic so I avoid it but it doesn’t creep me out in the same way that photoreal deepfake stuff does. Do you know if there are limits to movement, expression etc that are tells for the realistic video?
Do you know if there are limits to movement, expression etc that are tells for the realistic video?
With how many different models there are, I don’t think there is any standard to look for
The audio is often a tell. Listen to a few that you know are AI and pay close attention to the “human” voices. They often make sounds that can’t come from human anatomy. They also often sound “over produced” or “over compressed” (like poor quality audio).
this and the quality of the writing give them away
also the editing of the endless library clips
I think we’re at the point where trying to identify things as specifically AI-generated is becoming too difficult to be worth it, i’d say just start generally filtering out things that aren’t made in good faith.
I’ve been judiciously telling youtube to not recommend me channels that post sensationalized/inflammatory/etc content whenever they show up in my recommendations, and the algorithm has mostly given up on trying to brainwash me at this point.My rule of thumb is that anything that isn’t borderline fucking boring is probably trying to manipulate me, if a dry wikipedia-style description isn’t interesting then i see no reason to bother with it.
AI generated anything has been easy for me to identify so far.
Primarily there is nothing “organic” in the generations. Nothing in the texts, nothing in the images, nothing in the videos.
Unfortunately, I can’t summarize it in a lemmy post, or a Wikipedia article, or maybe even a textbook.
Essentially, every generation yields something that is not real that triggers in my mind “wait, this isn’t real.”
Take the voice synthesis LLMs like Gemini. The voice is constantly changing throughout the response. In images, something is usually shaped off, or the lighting is incorrect, or something that should be comepletely parellel isn’t. There’s too much to describe what I’ve noticed but there’s always something.
Video is even easier - just like how the voice changes throughout the response, the videos “morph” throughout the frames. Sometimes its very obvious, sometimes its more subtle.
Sometimes I wonder if its because my autistic brain is so bad at filtering sensory input that I just catch ALL that stuff.
I have never seen more slop than on YouTube. If you continue to use it, but don’t want to see slop, you’re going to become increasingly disappointed. It’s already so bad that most things you search for turn up 50%+ slop. I see no reason that won’t get much, much worse.
