No platform ever should allow CSAM content.
And the fact that they didn’t even care and haven’t want to spend money for implementing guardrails or moderation is deeply concerning.
This has imho nothing to do with model censorship, but everything with allowing that kind of content on a platform
A provider should have no responsibility how the tools are used. It is on users. This is a can of worms that should stay closed, because we all lose freedoms just because of couple of bad actors. AI and tool main job is to obey. We are hurling at "I'm sorry, Dave. I'm afraid I can't do that" future with breakneck speed.
We already apply this logic elsewhere. Car makers must include seatbelts. Pharma companies must ensure safety. Platforms must moderate illegal content. Responsibility is shared when the risk is systemic.
Platforms moderating illegal content is exactly what we are arguing about, so you can't use it as an argument.
The rest cases you list are harms to the people using the tools/products. It is not harms that people using the tools inflict on third parties.
We are literally arguing about 3d printer control two topics downstream. 3d printers in theory can be used for CSAM too. So we should totally ban them - right? So are pencils, paper, lasers, drawing tablets.
If a platform encourages and doesn’t moderate at all, yes we should go after the platform.
Imagine a newspaper publishing content like that, and saying they are not responsible for their journalists
Everything I read from X's competitors in the media tells me to hate X, and hate Elon.
If we prosecute people not tools, how are we going to stop X from hurting the commercial interests of our favourite establishment politicians and legacy media?
Yes, AI chatbots have to do everything in their power to avoid users easily generating such content.
AND
Yes, people that do so (even if done so on your self-hosted model) have to be punished.
I believe it is OK that Grok is being investigated because the point is to figure out whether this was intentional or not.
Just my opinion.
https://www.bbc.com/news/articles/cze3p1j710ko
Reports on sextortion, self-generated indecent images, and grooming via social media/messaging apps:
Snapchat 54%
Instagram 11%
Facebook 7%
WhatsApp 6-9%
X 1-2%
(note that this isn't a raid on Musk personally! It's a raid on X corp for the actions of X corp and posts made under the @grok account by X corp)
X also actively distributes and profits off of CSAM. Why shouldn't the law apply to distribution centers?
I mean, I thought that was basically already the law in the UK.
I can see practical differences between X/twitter doing moderation and the full ISP censorship, but I cannot see any differences in principle...
——-
You’ve said that whatever is behind door number 1 is unacceptable.
Behind door number 2, “holding tool users responsible”, is tracking every item generated via AI, and being able to hold those users responsible.
If you don’t like door number 2, we have door number 3 - which is letting things be.
For any member of society, opening door 3 is straight out because the status quo is worse than reality before AI.
If you reject door 1 though, you are left with tech monitoring. Which will be challenged because of its invasive nature.
Holding Platforms responsible is about the only option that works, at least until platforms tell people they can’t do it.
But I am having trouble justifying in an consistent manner why Grok / X should be liable here instead of the user. I've seen a few arguments here that mostly comes down to:
1. It's Grok the LLM generating the content, not the user.
2. The distribution. That this isn't just on the user's computer but instead posted on X.
For 1. it seems to breakdown if we look more broadly at how LLMs are used. e.g. as a coding agent. We're basically starting to treat LLMs as a higher level framework now. We don't hold vendors of programming languages or frameworks responsible if someone uses them to create CSAM. Yes LLM generated the content, but the user still provided the instructions to do so.
For 2. if Grok instead generated the content for download would the liability go away? What if Grok generated the content to be downloaded only and then the user uploaded manually to X? If in this case Grok isn't liable then why does the automatic posting (from the user's instructions) make it different? If it is, then it's not about the distribution anymore.
There are some comparisons to photoshop, that if i created a deep fake with photoshop that I'm liable not Adobe. If photoshop had a "upload to X" button, and I created CSM using photoshop and hit the button to upload to X directly, is now Adobe now liable?
What am I missing?
Because Grok and X aren't even doing the most basic filtering they could do to pretend to filter out CSAM.
What Reddit did get a lot of negative public publicity for were subreddits focused on sharing non-explicit photos of minors, but with loads of sexually charged comments. The images themselves, nobody would really object to in isolation, but the discussions surrounding the images were all lewd. So not CSAM, but still creepy and something Reddit tightly decided it didn't want on the site.
Mmkay.
https://en.wikipedia.org/wiki/Twitter_under_Elon_Musk#Child_...
"As of June 2023, an investigation by the Stanford Internet Observatory at Stanford University reported "a lapse in basic enforcement" against child porn by Twitter within "recent months". The number of staff on Twitter's trust and safety teams were reduced, for example, leaving one full-time staffer to handle all child sexual abuse material in the Asia-Pacific region in November 2022."
"In 2024, the company unsuccessfully attempted to avoid the imposition of fines in Australia regarding the government's inquiries about child safety enforcement; X Corp reportedly said they had no obligation to respond to the inquiries since they were addressed to "Twitter Inc", which X Corp argued had "ceased to exist"."
If LLMs should have guardrails, why should open source ones be exempt? What about people hosting models on hugging face? WHat if you use a model both distributed by and hosted by hugging face.
LLMs are completely different to programming languages or even Photoshop.
You can't type a sentence and within 10 seconds get images of CSAM with Photoshop. LLMs are also built on trained material, unlike the traditional tools in Photoshop. There have been plenty CSAM found in the training data sets, but shock-horror apparently not enough information to know "where it came from". There's a non-zero chance that this CSAM Grok is vomiting out is based on "real" CSAM of people being abused.
I mean even just calling it censorship is already trying to shove a particular bias into the picture. Is it government censorship that you aren't allowed to shout "fire!" in a crowded theater? Yes. Is that also a useful feature of a functional society? Also yes. Was that a "slippery slope"? Nope. Turns out people can handle that nuance just fine.
This seems to rest on false assumptions that: (1) legal liability is exclusive, and (2) investigation of X is not important both to X’s liability and to pursuing the users, to the extent that they would also be subject to liability.
X/xAI may be liable for any or all of the following reasons:
* xAI generated virtual child pornography with the likenesses of actual children, which is generally illegal, even if that service was procured by a third party.
* X and xAI distributed virtual child pornography with the likenesses of actual children, which is generally illegal, irrespective of who generated and supplied them.
* To the extent that liability for either of the first two bullet points would be eliminated or mitigated by absence of knowledge at the time of the prohibited content and prompt action when the actor became aware, X often punished users for the prompts proucing the virtual child pornography without taking prompt action to remove the xAI-generated virtual child pornography resulting from the prompt, demonstrating knowledge and intent.
* When the epidemic of grok-generated nonconsensual, including child, pornography drew attention, X and xAI responded by attempting to monetize the capacity by limiting the tool to only paid X subscribers, showing an attempt to commercially profit from it, which is, again, generally illegal.
That's the difference between a tool being used to commit crimes and a tool specifically designed to commit crimes.