Catfishing about to get almost undetectable.
(twitter.com)
You're viewing a single comment thread. View all comments, or full comment thread.
Comments (20)
sorted by:
Dont you think someone will come up with a way to detect AI video at a software level? Kinda like how every time someone develops new armor piercing technology someone else develops armor that defeats it.
What's the best body armor these days?
Wouldn't mind dropping up to a grand for premium body armor
Especially due to the inevitability of my confrontation with ZOG. You dont want ZOG to gun down your pay Vlad too easy. Suggest a vest.
Honestly, go ask your local PD. They generally actually test samples. Safe Life is what the state troopers near me get.
That kind of arms race can continue for a while, but eventually videos might get literally indistinguishable from the real thing. Any filter that can separate real vs. generated output can also be used directly as a training tool to help new models circumvent that same filter.
Forcing companies to include some kind of watermark in all generated output is another solution but it won't help much against open source models.
This argument never made any sense. The only input you have is the video or picture. If said input is made to 1:1 to what would be a real life file, then it's just an impossible solution.
Simplify this to a single pixel. You crop one pixel from a real life picture, then ask an AI to generate said pixel. Will it do it? Certainly. If on one hand, you have a pixel 150, 0, 50 (RGB values from 0 to 255), and on the other, you have a pixel 150, 0, 50, then it's physically impossible to tell which one is the real one, and which one is the generated one. No matter how good your software is, eventually it all comes down to "Just guess".
Secondly, your software to detect AI-generated content is a similar problem to the cheater / anti-cheat war that's been going in games for the past 3 decades. Guess what? Cheaters have always been winning, every single time. Every time there's an anti-cheat breakthrough and people claim "They did it! Cheaters are gone!", wait 3 months and suddenly a new wave of cheaters is here with even harder to detect methods. If you make a software that detect AI-generated content, all you do is amplify how good AI generation will be, in order to circumvent the AI-detector.
I don’t know and I’m not making the assumption that A.I. will never be able to create perfectly simulated human movements. In either case there will be a segment of time where AI will be able to create perfect or near perfect videos before humanity has countermeasures against it.
Convolutional neural networks are a training environment that is automated to train a neural network.
ACME Software Company writes a program that can detect deep fakes. Neat!
Intake ASC DF Detective and plug it into my training environment for my pet neural network.
Every time my NN gets detected, it gets an electric zap. Every time it isn't detected it gets a doggy treat.
Guess what happens after thirty million training cycles?
Thanks for the help Acme Software Company!
You're ignoring that AI detection neural networks can also be trained in an adversarial manner against AI generated content.
Yeah, it's possible that AI could eventually be undetectable, but I honestly doubt it.
I didn't say that AI images would or could be undetectable.
I pointed out that any software or program that detects AI images can be used as a training tool for AI image generation.
It is easier to use an AI detector that looks at file types or digital fingerprints to create training environments for AIs than it is to make AI detectors.
They still could be images with twelve fingers or nonsense writing, they just don't trip the AI detector software that was used for training.