Stanford researchers find Mastodon has a massive child abuse material problem
Stanford researchers find Mastodon has a massive child abuse material problem
Not a good look for Mastodon - what can be done to automate the removal of CSAM?
You're viewing a single thread.
The article points out that the strength of the Fediverse is also it’s downside. Federated moderation makes it challenging to consistently moderate CSAM.
We have seen it even here with the challenges of Lemmynsfw. In fact they have taken a stance that CSAM like images with of age models made to look underage is fine as long as there is some dodgy ‘age verification’
The idea is that abusive instances would get defederated, but I think we are going to find that inadequate to keep up without some sort of centralized reporting escalation and ai auto screening.
The problem with screening by AI is there's going to be false positives, and it's going to be extremely challenging and frustrating to fight them. Last month I got a letter for a speeding infraction that was automated: it was generated by a camera, the plate read in by OCR, the letter I received (from "Seat Pleasant, Maryland," lol) was supposedly signed off by a human police officer, but the image was so blurry that the plate was practically unreadable. Which is what happened: it got one of the letters wrong, and I got a speeding ticket from a town I've never been to, had never even heard of before I got that letter. And the letter was full of helpful ways to pay for and dispense with the ticket, but to challenge it I had to do it it writing, there was no email address anywhere in the letter. I had to go to their website and sift through dozens of pages to find one that had any chance of being able to do something about it, and I made a couple of false steps along the way. THEN, after calling them up and explaining the situation, they apologized and said they'd dismiss the charge--which they failed to do, I got another letter about it just TODAY saying a late fee had now been tacked on.
And this was mere OCR, which has been in use for multiple decades and is fairly stable now. This pleasant process is coming to anything involving AI as a judging mechanism.
THEN, after calling them up and explaining the situation, they apologized and said they'd dismiss the charge--which they failed to do
That sounds about right. When I was in college I got a speeding ticket halfway in between the college town and the city my parents lived in. Couldn't afford the fine due to being a poor college student, and called the court and asked if an extension was possible. They told me absolutely, how long do you need, and then I started saving up. Shortly before I had enough, I got a call from my Mom that she had received a letter saying there was a bench warrant for my arrest over the fine
How do you plan to train the AI to recognise CP?