Close Menu
    Technik

    Es stellte sich heraus, dass ein viraler Reddit-Beitrag, in dem Betrug durch eine Lebensmittelliefer-App behauptet wurde, KI-generiert war

    Share.
    View 12 Comments

    12 Kommentare

    3. CircumspectCapybara on

      For those wondering how Gemini can determine a piece of content (whether text, image, audio, video, even code) was generated by Google AI with any sort of accuracy, it’s a neat piece of tech DeepMind came up with called [SynthID](https://deepmind.google/models/synthid), a sort of „watermark“ that Gemini embeds in its outputs that is resilient even if you tweak or modify the output, e.g., by cutting parts out or rearranging, cropping, pitch shifting, time shifting, etc.

      If you really want to know about the technical details, check out [this presentation](https://www.youtube.com/watch?v=xuwHKpouIyE), or the [Nature article](https://www.nature.com/articles/s41586-024-08025-4) on it.

      Basically, Gemini uses a keyed hash function to alter the probability distribution of the output tokens. Imagine for each token in the output you had n equally high-scoring candidates, and normally you would pick one at random. But with SynthID, you have a secret key which you hash with the context to produce a cryptographically pseudorandom bitstream (which cannot be guessed by anyone without the secret key) which you use to pick tokens from among candidates. The crux is if you’re using a cryptographic hash function and the key is secret, the bitstream looks indistinguishable from random. It’s functionally equivalent to choosing randomly. And yet it’s anything but random to one who knows the key.

      You can then evaluate piece of content (or sections thereof) by looking at the tokens that make it up and seeing if it matches this probability distribution for way too much of it to just be random chance or an unlucky coincidence.

    7. Kakariko_crackhouse on

      That’s the spin I’d try to push if I was one of those delivery apps too

    9. apocalypsebuddy on

      People on Blind confirmed the schema mentioned and other facts in the post. But we’ll continue to get astroturfed with the “debunk”

    10. bdoomed on

      I’ve seen more about the post being fake than discussion about the post itself. Makes me feel like it maybe wasn’t fake, or partially true enough that someone activated a massive PR campaign to combat it. Smells like money and I don’t trust money.

    11. No-Public3806 on

      I’ve tested those AI detectors on non AI text and a lot of them say it’s AI when it’s not. I would push the fake AI story if I was Uber Eats. We don’t need an employee to tell us how terrible they are, just open the app, close it and open it again and you’ll see how they jack up the price

    Leave A Reply