And if a person is using a model based on harmful training data, they should be held responsible.

I will have to disagree with you for several reasons.

• You are still making assumptions about a system you know absolutely nothing about.
• By your logic anything born from something that caused suffering from others (this example is AI trained on CSAM) the users of that product should be held responsible for the crime committed to create that product.
  • Does that apply to every product/result created from human suffering or just the things you don't like?
  • Will you apply that logic to the prosperity of Western Nations built on the suffering of indigenous and enslaved people? Should everyone who benefit from western prosperity be held responsible for the crimes committed against those people?
  • What about medicine? Two examples are The Tuskegee Syphilis Study and the cancer cells of Henrietta Lacks. Medicine benefited greatly from these two examples but crimes were committed against the people involved. Should every patient from a cancer program that benefited from Ms. Lacks' cancer cells also be subject to pay compensation to her family? The doctors that used her cells without permission didn't.
  • Should we also talk about the advances in medicine found by Nazis who experimented on Jews and others during WW2? We used that data in our manned space program paving the way to all the benefits we get from space technology.

Are there any guarantees that harmful images weren’t used in these AI models?

Lol, highly doubt it. These AI assholes pretend that all the training data randomly fell into the model (off the back of a truck) and that they cannot possibly be held responsible for that or know anything about it because they were too busy innovating.

There's no guarantee that most regular porn sites don't contain csam or other exploitative imagery and video (sex trafficking victims). There's absolutely zero chance that there's any kind of guarantee.

There are probably safeguards in place to prevent the creation of CSAM, just like there are for other illegal and offensive things, but determined people work around them.

Sure, but isn't the the perpetrator the company that trained the model without their permission? If a doctor saves someone's life using knowledge based on nazi medical experiments, then surely the doctor isn't responsible for the crimes?

Let's do a thought experiment, and I'd look to to tell me at what point a victim was introduced:

1. I legally acquire pictures of a child, fully clothed and everything
2. I draw a picture based on those legal pictures, but the subject is nude or doing sexually explicit things
3. I keep the picture for my own personal use and don't distribute it

Or with AI:

1. I legally acquire pictures of children, fully clothed and everything
2. I legally acquire pictures of nude adults, some doing sexually explicit things
3. I train an AI on a mix of 1&2
4. I generate images of nude children, some of them doing sexually explicit things
5. I keep the pictures for my own personal use and don't distribute any of them
6. I distribute my model, using the right to distribute from the legal acquisition of those images

At what point did my actions victimize someone?

If I distributed those images and those images resemble a real person, then that real person is potentially a victim.

I will say someone who does this creepy and I don't want them anywhere near children (especially mine, and yes, I have kids), but I don't think it should be illegal, provided the source material is legal. But as soon as I distribute it, there absolutely could be a victim. Being creepy shouldn't be a crime.