Comment on With ‘AI slop’ distorting our reality, the world is sleepwalking into disaster

reddig33@lemmy.world ⁨2⁩ ⁨days⁩ ago

• BartyDeCanter@lemmy.sdf.org ⁨2⁩ ⁨days⁩ ago

  It’s a cat and mouse game, at best. If you have a tool that can reliably detect AI slop, then that tool can be used as part of the training process to fool the detection tool.

  source
• Blaster_M@lemmy.world ⁨2⁩ ⁨days⁩ ago

  AI auto detection is a coin toss at best - and it often hurts real artists/writers when they’re false positive’d, which is far too often.

  People didn’t magically forget how to proofread or use Photoshop (or similar tools) when AI generation became popular - the good fakes are just better now. Small “tells” can be fixed in post by anyone that has previously done “photoshopping” work… the kind of people who have already been doing such things before AI generation. AI generation just makes it easier and faster if you know how to use inpainting or img to img.

  source
• CarbonIceDragon@pawb.social ⁨2⁩ ⁨days⁩ ago

  For how long though? The issue with detecting AI generated stuff, Id imagine, is that a picture contains a finite amount of information, especially a digital one. These things have been improving relatively quickly, and I cant think of any fundamental reason why one could not eventually create images where every pixel is as it would be if that image were real, or at least close enough that detection is not even theoretically possible if you dont have some actual proof that the event depicted couldnt have happened. We may not be there yet, but the closer we get to it, the more prone to error and therefore less useful any detection algorithm must be.

  source
• desktop_user@lemmy.blahaj.zone ⁨2⁩ ⁨days⁩ ago

  yes, however avoiding that has been part of the training process of many machine learning models for years as a convenient way of preventing overfitting to the training dara.

  source