re: AI child pron predator caught in Florida

quote:

---

So, that means the AI servers would have to access massive amounts of child porn.

---

That seems like the way to go at this.

quote:

---

What happens when some 35 year old male gets a detailed sex doll that looks like some 14-year old?

---

Damn... shite is getting crazy. This is a great point.

quote:

---

seems like a pretty slippery slope.

---

There is a (legitimate) knee-jerk jerk that it should be illegal, but there is arguably no victim. It could lead to illegal acts, but so can a lot of legal things; Exhibit #1 is alcohol. This is edging closer to thought crimes.

The emotional part of me doesn't care if people who view AI child pornography end up in jail, but the logical side of me sees the bigger picture and the potential ramifications.

quote:

---

What if a guy draws his own child porn?

---

From the way the law is written, it doesnt look like you cant have it. Just that you cant distribute, show or sell it. Unless I read it wrong

quote:

---

2) Any person who, with knowledge that the material is a deepfake depicting a minor, knowingly advertises, distributes, exhibits, exchanges with, promotes, or sells any sexual material that depicts a minor engaging in sexual conduct shall be punished by imprisonment at hard labor for not less than ten nor more than thirty years, a fine of not more than fifty thousand dollars, or both. At least ten years of the sentence of imprisonment imposed shall be served without benefit of probation, parole, or suspension of sentence.

---

quote:

---

I’m pretty sure that law only applies to deepfakes (i.e. fake images of real person who is a minor).

---

There are billions of unique looking faces in the world.

Its arguable most AI images could be taken as a deep fake of someone real.

I can’t have a logical conversation with someone who does not believe possessing child pornography should be a punishable offense and tries to spin the discussion into some obscure thought crime that it was never about.

Possessing child pornography no matter how it was created should be punished to the higher degree of the law because people who would possess such material are a threat to society.

This shouldn’t be a problem to anyone who doesn’t not have an ulterior motive to protect and or decriminalize child pornograpy i.e normalize this sick deviant behavior.

quote:

---

I suppose this is better than actual child pRon. Less kids getting sexually abused and taken advantage of which is a net positive to society.

---

I think this is a dangerous slippery slope. Porn is an addiction to some, and one that progresses from vanilla Playboy (or something similar) to increasingly hardcore and depraved subjects. Adding in easily accessible CP, even if it’s AI generated will invariably lead some to turn to it as they seek their next thing to get their jollies. An if some of those have access to children, well I don’t have to tell you what will follow.

quote:

---

the LA law requires “knowledge that the material is a deepfake depicting a minor.”

---

OP doesn’t care. He’s ready to prosecute thought crime.

quote:

---

SSBBW porn

---

Do we want to know what this is?

quote:

---

It should be. AI generates images from a massive number of real images.

---

quote:

---

So, that means the AI servers would have to access massive amounts of child porn.

That sounds highly illegal to me.

---

So this aspect is actually pretty complicated. Long post ahead…

TL;DR:
1. A model capable of producing CSAM doesn’t necessarily need to be trained on it.
2. There are certainly some creeps out there who seem to be going out of their way to intentionally produce models capable of creating CSAM, but it’s impossible to prove without someone getting their training data.
3. The major foundational models produced by the big AI companies likely had an extremely small amount of CSAM in their training data simply because they trained the models on images automatically scraped from the internet.

Long version:

First off, diffusion models are capable of generating images based on concepts outside of their actual training data.. e.g. concepts they weren’t explicitly trained on at all. This is especially true for the most advanced models. So the fact that a diffusion model can generate CSAM does not mean it was explicitly trained on CSAM.

However, there’s been a perverse (if not somewhat predictable) trend among publicly-available models that have been fine-tuned by users. These are models that are designed to run locally on consumer-grade hardware rather than remotely in server racks.

The thing is.. the #1 reason people fine-tune the publicly available models is to make them better at generating.. well, porn. But a model that can generate porn can probably also generate CSAM, because again - diffusion models don’t have to be explicitly trained on every concept. Since there is no way to tell what exact training images were used by analyzing the model itself, it’s impossible to say whether a given model was trained on CSAM or not.

Then there’s a separate issue where many of the big models (both those publicly available for download and those that run remotely) were trained at some point on LAION datasets. LAION (a nonprofit focused on open source AI development) basically scraped the internet for image/caption pairs to train diffusion models. The actual dataset consists of URL links to the images plus the text of the captions found in HTML.

There was a report a couple of years ago that the LAION-5B dataset included about 1,000 images that were confirmed to be CSAM (out of 5 billion images total, so 0.00002% of the dataset). LAION responded by pulling the dataset.

It’s unlikely anybody actually committed a crime here because LAION’s dataset came from links that were automatically scraped from the internet as a whole and LAION didn’t even actually host the images. Additionally, the AI companies would not have known that 0.00002% of the 5 billion images contained CSAM because surely nobody manually vetted each image. However, it does mean that many of the foundational models were likely trained on some amount of CSAM simply because it existed on the internet. It’s questionable (at best) whether this data actually made much difference to model capabilities, but there’s really no unringing the bell at this point regardless.