The EU’s AI rules are having an impact.
  • mrmaplebar@fedia.io
    184·
    3 days ago

    Not enough. AI generated content should be required by law to contain both visible and invisible watermarks.

    • floofloof@lemmy.caEnglish
      12·
      2 days ago

      You can’t enforce that with text though, because people will just delete the words. They use a subtle statistical signal because people can’t just spot it and delete it. Unfortunately it raises a problem: either they have to make the signal a secret, in which case only Anthropic and those contracted to keep the secret could detect it, or it’s not, in which case people can use a detector and some other AI to discover how to eliminate the signal. I expect Anthropic would try to keep control over the detection secrets and provide a detection API for third-party software to use.

        • floofloof@lemmy.caEnglish
          3·
          2 days ago

          As I understand it, you’re right that you need a good length of output to be able to detect the watermark, because only then can you see the statistical effect with confidence. And in code there are usually various options for how you get something done, and they could watermark generated code by adding a distinctive pattern to its preferences for certain constructs over others. But again, it would have to be subtle, so you’d need a large enough sample of its output before you could see the effect.

          • Ek-Hou-Van-Braai@piefed.socialEnglish
            2·
            2 days ago

            It should be fairly easy to stop this detention even in code.

            You just need to set very strict “rules” the AI must follow when writing code.

            Variable and method names etc. Must follow a specific formula.

            Have a different AI write your docs etc.

            If you’re just vibe coding everything, the watermark would work. So maybe a good thing. But it’ll be whack a mole, and if you’re careful I’m sure you can stop the detection

    • TrickDacy@lemmy.worldEnglish
      11·
      2 days ago

      I’m so silly. I saw this and thought, oh cool a teeny tiny surprising win. Guess I should only be happy with the destruction of all LLMs.

    • schnurrito@discuss.tchncs.deEnglish
      4·
      2 days ago

      Nor to me - but it doesn’t make a lot of sense with images either.

      Text is literally just an array of numbers (Unicode character points); where do you put “invisible” watermarks there? Images are, likewise, a three-dimensional array of pixel color values; while it’s possible to very slightly vary those color values and call that “watermarking” (similar concept: steganography), how does that not get lost even unintentionally on the first lossy compression step?!

  • godsammitdam@lemmy.zipEnglish
    8·
    2 days ago

    So then they can lie and say real pictures have the watermark?

    So they can lie and say fake images don’t have the watermark?

    So the AI looking for the watermark can hallucinate?

    Yeah, I’m sure that’s super reliable.

  • SnailMagnitude@mander.xyzEnglish
    11·
    3 days ago

    This seems a riot, trying to sneak in invisible copyright really does seem like the copyright system is well and truly fucked.

    It arrived to try and help with the deluge of slop from Guttenberg nonsense but the whole idea, and copyleft, seems to be imploding at the moment.

  • MattR@feddit.orgEnglish
    1·
    2 days ago

    There are multiple major flaws with watermarks for texts:

    1. Their method only works for longer texts as it relies on statistical effects that aren’t clear enough to detect in a very short text.
    2. They are resting their approach on an incorrect assumption: their models prefer certain words and if they appear more frequently than in a random text, they assume it’s generate by their model. But who says texts are random? Out of billions of people there will be some who have a similar preference for some words and their texts will always be wrongfully accused of AI-generated just because they happen to have a similar word choice preference. The shorter the texts the likelier that issue becomes.
    3. There will soon be tools that replace a random amount of words with synonyms automatically and therefore remove any chances of detecting the watermark.
  • Promethiel@lemmy.worldEnglish
    1·
    3 days ago

    Google already does this with Gemini.

    Not only does the technical side work very intermittently at detection when you do carry over the same file, but with an “invisible” watermark, it can be bypassed entirely as easily as with a screenshot.