A new report warns that the proliferation of child sexual abuse images on the internet could become much worse if something is not done to put controls on artificial intelligence tools that generate deepfake photos.
I’m not sure that has to be true. Like you can ask an AI to give you a picture of a sailboat on the moon, while it has not ever seen a sailboat on the moon.
It could be trained on photos that are not pornografic containing kids and images that are pornografic containing adults.
you can ask an AI to give you a picture of a sailboat on the moon
Yes, correct. I’ll try to explain why that comparison isn’t entirely correct in this case and why my point stands: If you ask the model to draw an image of a sailboat on the moon it will take its context definition of “on the moon” and will likely end up selecting imagery of moon landscapes and will then put a sailboat in there. That sailboat will likely be assembled from frontal or sideviews of sailboats it has seen and will contain typical elements like a small bow pointing up and a keel line down the middle and some planks or a fibreglass-like structure to fill the are in between, depending on the style of things it has seen in the context of “lots of sailboat in this training picture”.
If the model has never seen the underside of a sailboat it will likely reduce to “boat” and start to put a freightship or containership-type of bow and keel there, it probably has seen imagery of those in drydocks - the output wouldn’t look convincing to you as a viewer. In order to create a convincing sailboat in your example, the model needs a good idea what a sailboat looks like under the waterline. Which means, it has seen enough of that. Whithout further elaborating, I am sure you can understand how this implies massive ethical problems with generating a model for content that contains exploitative and abusive elements.
You are not logged in. However you can subscribe from another Fediverse account, for example Lemmy or Mastodon. To do this, paste the following into the search field of your instance: !technology@beehaw.org
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
I’m not sure that has to be true. Like you can ask an AI to give you a picture of a sailboat on the moon, while it has not ever seen a sailboat on the moon.
It could be trained on photos that are not pornografic containing kids and images that are pornografic containing adults.
Yes, correct. I’ll try to explain why that comparison isn’t entirely correct in this case and why my point stands: If you ask the model to draw an image of a sailboat on the moon it will take its context definition of “on the moon” and will likely end up selecting imagery of moon landscapes and will then put a sailboat in there. That sailboat will likely be assembled from frontal or sideviews of sailboats it has seen and will contain typical elements like a small bow pointing up and a keel line down the middle and some planks or a fibreglass-like structure to fill the are in between, depending on the style of things it has seen in the context of “lots of sailboat in this training picture”.
If the model has never seen the underside of a sailboat it will likely reduce to “boat” and start to put a freightship or containership-type of bow and keel there, it probably has seen imagery of those in drydocks - the output wouldn’t look convincing to you as a viewer. In order to create a convincing sailboat in your example, the model needs a good idea what a sailboat looks like under the waterline. Which means, it has seen enough of that. Whithout further elaborating, I am sure you can understand how this implies massive ethical problems with generating a model for content that contains exploitative and abusive elements.