AI-generated models are causing another pain for internet moderation and safety — AI-generated child pornography. Thousands of AI-generated child-sex images have been found on forums across the dark web, with some users even sharing detailed guides about how other paedophiles can create their own images.
The generated imagery is problematic for a number of reasons, other than the obvious problem of child abuse. These images, in their current quantity, are capable of overwhelming tracking systems to prevent such content from surfacing on the internet simply because they were made to catch known images of abuse, instead of newly generated ones.
Additionally, given the realism of these images, there’s also the question of whether a particular image is real or not. This opens up an even bigger debate regarding whether these images even violate child-protection laws or not, considering the children depicted in such images aren’t always real. According to The Washington Post, Justice Department officials fighting against child exploitation say that while the images themselves are illegal, they can’t cite a case in which a suspect has yet been charged for creating them.
AI-based image generation models such as Dall-E, Midjourney and Stable Diffusion train on the millions of images available across the internet, which include images of children using which the models can generate images of their own. While companies like OpenAI and Stability AI strongly condemn the use of their AI models to generate explicit images of children and have even gone as far as removing all explicit content from their training datasets to prevent the model from being able to generate such images.
However, the open-source nature of some of these tools, Stable Diffusion, in particular, means that anyone can just download the tool and run it locally on their computers however they please. Any company rules or oversight placed on the tool can hence be largely avoided.
The tool does have an in-built filter to prevent such images from being generated, but that can be removed with a few lines of code added to the program. Stable Diffusion’s competitors — Dall-E and Midjourney, both ban sexual content entirely and aren’t available open-source, meaning any images generated are through company-run channels and tracked to prevent misuse.
But that doesn’t stop dark-web paedophile forums from openly-discussing strategies on how to evade any protection mechanisms and filters set in place. These even include using non-English languages which the users believe will be harder to track and are less vulnerable to detection or censorship.
The Washington Post reports that on one forum with 3,000 members, over 80% of the users claimed to either have already used or intended to use an AI tool to generate child sexual abuse images according to Avi Jager, the head of child safety and human exploitation at ActiveFence — a company that works with social media and streaming platforms to catch malicious content.
What can be done about this?
While the problem sure is big, all hope is not lost yet. Two officials from the Justice Department’s Child Exploitation Obscenity Section claim that the images are illegal under two American laws:
- 18 U.S. Code § 2256: Bans any computer-generated image that is sexually explicit and shows someone who is “virtually indistinguishable” from a real child. Passed in 1978.
- 18 U.S. Code § 1466A: Bans any computer-generated image showing a child engaging in sexually explicit conduct if it’s obscene and lacks serious artistic value. This second law goes as far as to state that it doesn’t matter if the minor depicted actually exists or not. Passed in 2003.
AI researchers aren’t giving up either. There is ongoing research on how models can be trained to stop such images from appearing on the internet by creating another model that generates examples of fake child-exploitation images. The method comes with its own harms and a “massive psychological cost” though, according to Margaret Mitchell, an AI researcher previously working with Google’s Ethical AI team.
Other researchers are working on identification systems that could imprint code into images that’d link back to the creator. Researchers at the University of Maryland published a new technique in May 2023 that adds an invisible watermark to AI-generated images that can identify the image’s creator and will be difficult to remove. That said, these methods would require industry-wide participation and would still not be able to catch all offences, according to Mitchell.
In the News: Hackers threaten to release 80GB of stolen Reddit data; demand API u-turn
