19
On AI-generated articles (piefed.social)
submitted 1 week ago by to c/piefed_meta@piefed.social
26 comments fedilink hide all child comments
 

I thought of sharing some issues that preoccupy me here, because of how active piefed development is. Btw thank you all for your hard work!

I have noticed that more and more articles are created with LLMs without disclosing it. So, it seems to me that if someone wants to avoid posting this sort of content, one needs to at least:

  • check how many articles the author posts per day in the specific site,
  • then if the author really exist and
  • finally copy-paste part of the text in a couple of ai-detector sites.

Initially, I thought of making a post for a feature request like the one that detects AI generated images, but for text. But I can't because if I got this right, the ai-detectors may flag an article as ai generated when the author is not a native english speaker [1].

Of course the feature that allows us to label ai-generated content ourselves before posting it is very important. In a way my questions are related to something else: what to do before that. As briefly as I can, here they are:

  • does the process mentioned above seem adequate?
  • if yes, what else can someone do to check an article before posting, and are there any alternatives/variations to this process?
  • if no, what would you suggest?
you are viewing a single comment's thread
view the rest of the comments
[–] 2 points 1 day ago* (1 child)

Sorry for being late to the party, but I had to give my two cents on AI detectors. (specifically GPTZero & co., watermarking is different) I think it's a bad idea to recommend them. These companies are overpromising their accuracy and essentially selling false certainty. This will also not fade over time since the inaccuracy is baked into the tech.

Currently the performance may be good for basic use cases, but recent papers seem to indicate you can still dazzle them by burning more tokens or explicitly prompting against detectors. But from what I understand about GenAI training* this is an eternal cat & mouse game, since the AI vendors don't want their output to be detectable.

what I understand about GenAI trainingA basic technique of training a generative model is training a discriminative model alongside it. The gen model tries to synthesize output and the discriminative model tries to distinguish it from actual training data. The models are then trained alongside each other, building better and better discriminator models to sus out better and better generative ones.

If these detector companies had some kind of breakthrough (better techniques, better datasets, whatever) to sus out generated output, they would be bought up to improve GenAI training. More likely, they don't. Maybe they can dial into some typical usecases for specific models and perform highly there, but that will fail to generalize for the edge cases.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 2 points 21 hours ago*

    Yes they're not perfect. I don't think it'd be a good idea to take moderation action against an account for a single post being detected as LLM-generated. Or two posts. But 3... It's just not possible for 3 posts in a row to be false positives. Combined with other signals (e.g. age of account, etc) you can make a pretty solid decision.

  • permalink
  • fedilink
  • source
  • parent