Post

You can make LLMs say almost anything. A news organization reporting "what grok said" is engaging in journalistic malpractice. It doesn't know. It just creates plausible sounding words in response to a prompt.

Reuters@reuters.com

Grok says safeguard lapses led to images of 'minors in minimal clothing' on X reut.rs/498SoUk

reut.rsGrok says safeguard lapses led to images of 'minors in minimal clothing' on XElon Musk's xAI artificial intelligence chatbot Grok said on Friday lapses in safeguards had resulted in "images depicting minors in minimal clothing" on social media platform X and that improvements were being made to prevent this.

Worth noting that AI image moderation is an effectively solved problem. Any of the frontier models can detect NSFW images and refuse to return them to the user. Grok is clearly not doing this

2/1/26, 5:39 pm1 repost5 likes

They often err on the side of caution sometimes blocking non-NSFW images

zendarva.bsky.social

@zendarva.bsky.social

I mean sure... But any one of the available open source LLMs takes about a day to modify to just generate whatever you want by abliterating *cough search term cough* it.

And? We're not talking about offline apps like ComfyUI, etc. We're talking about a chat bot that's accessible to literally anyone on Twitter that has zero safety rails

zendarva.bsky.social

@zendarva.bsky.social

You mean tools that are accessible to literally anyo9ne on twitter, with even fewer safety rails? Takes 30 seconds to set up a local LLM, and Grok will *tell you how to do it*.

I'm specifically talking about the image generation guard rails, image generation cannot be set up in "30 seconds" and most people don't have a powerful enough computer to do so. I'm confused what point you're trying to make here.