Inspecting Videos for Generative AI
 |
| A Picture Created by Google AI-I have no idea why it did this |
When I started the channel four years ago, AI was just beginning to find its footing. We officially signed up with YouTube and released our first video on March 29, 2022. Back then, we really didn’t know what we were doing. We depended on "YouTube Gurus" for suggestions on how to build a channel while trying to navigate YouTube’s official rules.
Our goal was to create a helpful, calming space. Initially, we used Pexels and Pixabay to find clips for our videos. The Gurus at the time insisted you could succeed by simply pasting free clips together and using the YouTube Audio Library or Pixabay for music and sound effects. It was marketed as the "free way" to become a content creator. Our first video, Calming and Relaxing Music with Rivers and Waterfalls, earned 227 views and 4 likes.
As it turns out, the Gurus were wrong. This wasn't a sustainable way to build a channel. While YouTube wasn’t as strict about "repetitive content" back then, they didn't push that type of media, and viewers generally didn't engage with it because the quality just wasn't there.
The Rise of Shorts and the AI "Trap"
YouTube Shorts expanded to the United States in March 2021, but as new creators, we didn’t really notice the "Shorts shelf" until August 29, 2022. Unsure of how to use the format, we treated Shorts as ads for our long-form content. Our first featured a piano track to advertise the Mood Room. This was a strategy suggested by a Guru called Dream Cloud, who claimed success came from melding multiple music videos together. It didn't work—the short only received 76 views.
Dream Cloud is also the reason we first got caught up in generative AI. Every Guru was pushing it, and around this time, we started adding cat shorts to our channel. Wanting to avoid "repetitive content" strikes and grow the audience, we experimented with AI. Back then, it was easy to spot: cats often had six legs, two faces, or two tails. Needless to say, it didn't do well. We eventually went back to Pexels and Pixabay, adding text to our clips and experimenting with AI voices, though those always sounded far too robotic.
The Turning Point
By December 15, 2023, when we made Sleep Serenity: Celestial Mushrooms and Ambient Harmony, AI had become an industry obsession. Pixabay was flooded with it. If I looked for a simple clip of a fish pond, almost every result was AI-generated.
A Timeline of the Takeover:
2019: Canva acquired Pexels and Pixabay. Initially, it felt like a win for creators.
September 21, 2023: Pixabay updated its license to allow AI-generated uploads. Almost overnight, "dead-eyed" AI clips began burying the work of real photographers.
2024: Platforms updated their terms to use high-quality human uploads to "train" their machines—essentially using our art to teach AI how to mimic us.
March 18, 2024: YouTube officially rolled out "Altered Content" labels to flag synthetic media.
2026: Today, tools like "Dream Lab" and "Magic Media" are the default. Systems are designed to generate machine-made images from scratch because it’s cheaper for the platforms than hosting human art.
We also saw the backlash against a famous artist who used AI to keep up with a tight comic-animation schedule. When he received the "Altered Content" label, the human pushback was intense; people called it "lazy" and claimed it wasn't "real art."
Seeing that, we promptly dumped AI-generated videos and paintings in favor of the human touch. Interestingly, our hand-painted watercolor elements and backgrounds perform just as well—if not better—than the AI ever did, and they look far superior.
How We Vet Content for Authenticity
Since we’ve committed to being a "human-made" channel, we now have to vet every clip. AI has improved; the six-legged cats are gone. Here is our process:
The Human Eye Test: We look for lighting inconsistencies. Does the wall or ceiling look like it is glowing? Do the shadows fall where they should? We check for "morphing" textures and look closely at fur—does it look like real cat hair or molded plastic?
Using the AI Against Itself: If something feels off, we use tools like Gemini to analyze the file. Often, these models can recognize their own "fingerprints" or metadata. You can also cross-reference clips with ChatGPT or Grok to see if they detect generative patterns.
Specialized Detection: We use external tools like Hive Moderation, Deepware, and C2PA viewers. Recently, I tested a clip where Hive flagged it as AI, but four other tools (including Gemini and Grok) cleared it. Because it passed the "eye test" and most tools, we decided it was real.
I personally miss the days when you could purchase a photo and just know it was real. We are fighting to keep that human element alive.
References & Resources:
YouTube Official Blog: Our approach to responsible AI innovation (March 2024).
Pixabay Content License: Updates regarding AI-generated imagery (Sept 2023).
C2PA (Coalition for Content Provenance and Authenticity): Technical standards for content origin.
Hive Moderation: AI-Generated Content Detection Tools.
If this post was useful, please consider following our blog using the box on the right. You can also visit our YouTube channel @serenityofthemind or check out our store via the green button.
What do you do to vet your own videos and pictures for AI? We would love to hear your methods in the comments!
No comments:
Post a Comment