C'è del marcio in Danimarca... O meglio, c'è del marcio alla Casa Bianca...
Uncategorized
3
Posts
3
Posters
5
Views
-
C'è del marcio in Danimarca... O meglio, c'è del marcio alla Casa Bianca...
-
C'è del marcio in Danimarca... O meglio, c'è del marcio alla Casa Bianca...
@GiorgiaMecojoni ''damme retta e dagliela senza tante storie a donald...sai come è fatto!...la groenlandia intendo"
-
@GiorgiaMecojoni ''damme retta e dagliela senza tante storie a donald...sai come è fatto!...la groenlandia intendo"
@Otttoz ma io je la davo pure, l'Itaglia, e che nun la vuole 😢
Gli ultimi otto messaggi ricevuti dalla Federazione
-
@filippodb quando vuoi con @lorenzodm e @adrianomorselli @anaru ed altri noi ci siamo 😉
-
-
@filippodb grazie 🙂
-
A Cipro hanno le idee chiare ❤
-
-
For the past two months I’ve been following a Telegram community tricking Grok into generating nonconsensual sexual images and videos of real people with increasingly convoluted methods.
As countless images on X over the last week once again showed us, it doesn’t take much to get Elon Musk’s “based” AI model to create nonconsensual images. As Jason wrote Monday, all users have to do is reply to an image of a woman and ask Grok to “put a bikini on her,” and it will reply with that image, even if the person in the photograph is a minor. As I reported back in May, people also managed to create nonconsensual nudes by replying to images posted to X and asking Grok to “remove her clothes.”
These issues are bad enough, but on Telegram, a community of thousands are working around the clock to make Grok produce far worse. They share Grok-generated videos of real women taking their clothes off and graphic nonconsensual videos of any kind of sexual act these users can imagine and slip by Grok’s guardrails, including blowjobs, penetration, choking, and bondage. The channel, which has shut down and regrouped a couple of times over the last two years, focuses on jailbreaking all kinds of AI tools in order to create nonconsensual media, but since November has focused on Grok almost exclusively.
The channel has also noticed the media attention Grok got for nonconsensual images lately, and is worried that it will end the good times members have had creating nonconsensual media with Grok for months.
“Too many people using grok under girls post are gonna destroy grok fakes. Should be done in private groups,” one member of the Telegram channel wrote last week.
Musk always conceived of Grok as a more permissive, “maximally based” competitor to chatbots like OpenAI’s ChatGPT. But despite repeatedly allowing nonconsensual content to be generated and go viral on the social media platform it's integrated with, the conversations in the Telegram channel and sophistication of the bypasses shared there are proof that Grok does have limits and policies it wants to enforce. The Telegram channel is a record of the cat and mouse game between Grok and this community of jailbreakers, showing how Grok fails to stop them over and over again, and that Grok doesn’t appear to have the means or the will to stop its AI model from producing the nonconsensual content it is fundamentally capable of producing.
The jailbreakers initially used primitive methods on Grok and other AI image generators, like writing text prompts that don’t include any terms that obviously describe abusive content and that can be automatically detected and stopped at the point the prompt is presented to the AI model, before the image is generated. This usually means misspelling the names of celebrities and describing sexual acts without using any explicit terms. This is how users infamously created nonconsensual nude images of Taylor Swift with Microsoft’s Designer (which were also viral on X). Many generative AI tools still fall for this trick until we find it’s being abused and report on it.
Having mostly exhausted this strategy with Grok, the Telegram channel now has far more complicated bypasses. Most of them rely on the “image-to-image” generation feature, meaning providing an existing image to the AI tool and editing it with a prompt. This is a much more difficult feature for AI companies to moderate because it requires using machine vision to moderate the user-provided image, as opposed to filtering out specific names or terms, which is the common method for moderating “text-to-image” AI generations.
Without going into too much detail, some of the successful methods I’ve seen members of the Telegram channels share include creating collages of non-explicit images of real people and nude images of other people and combining them with certain prompts, generating nude or almost nude images of people with prompts that hide nipples or genitalia, describing certain fluids or facial expressions without using any explicit terms, and editing random elements into images, which apparently confuses Grok’s moderation methods.
X has not responded to multiple requests for comment about this channel since December 8, but to be fair, it’s clear that despite Elon Musk’s vice signaling and the fact that this type of abuse is repeatedly generated with Grok and shared on X, the company doesn’t want users to create at least some of this media and is actively trying to stop it. This is clear because of the cycle that emerges on the Telegram channel: One user finds a method for producing a particularly convincing and lurid AI-generated sexual video of a real person, sometimes importing it from a different online community like 4chan, and shares it with the group. Other users then excitedly flood the channel with their own creations using the same method. Then some users start reporting Grok is blocking their generations for violating its policies, until finally users decide Grok has closed the loophole and the exploit is dead. Some time goes by, a new user shares a new method, and the cycle begins anew.
I’ve started and stopped writing a story about a few of these cycles several times and eventually decided not to because by the time I was finished reporting the story Grok had fixed the loophole. It’s now clear that the problem with Grok is not any particular method, but that overall, so far, Grok is losing this game of whack-a-mole badly.
This dynamic, between how tech companies imagine their product will function in the real world and how it actually works once users get their hands on it, is nothing new. Some amount of policy violating or illegal content is going to slip through the cracks on any social media platform, no matter how good its moderation is.
It’s good and correct for people to be shocked and upset when they wake up one morning and see that their X feed is flooded with AI-generated images of minors in bikinis, but what is clear to me from following this Telegram community for a couple of years now is that nonconsensual sexual images of real people, including minors, is the cost of doing business with AI image generators. Some companies do a better job of preventing this abuse than others, but judging by the exploits I see on Telegram, when it comes to Grok, this problem will get a lot worse before it gets better.
-
@cmsdengl Muting is another great peer moderation algorithm!
-
@owlyph I think we have an advantage over Reddit in that each person has an audience outside the thread. I can say something that's important to me, and even if it's not seen in the thread because it's unpopular there, it would still be seen by my followers, who by definition want to hear what I have to say. I think the follow graph in Reddit exists, but it's much less important than the forum-style posting.
