It's not just the guardrails, but the ham-fisted implementation.

Grok is supposed to be "uncensored", but there are very specific words you just can't use when asking it to generate images. It'll just flat out refuse or give an error message during generation.

But, again, if you go in a roundabout way and avoid the specific terms you can still get what you want. So why bother?

Is it about not wanting bad PR or avoiding litigation?

The implementation is what gets to me too. Fair enough that a company doesn't want their LLM used in a certain way. That's their choice, even if it's just to avoid getting sued.

How they then go about implementing those guardrails is pretty telling about their understand and control over what they've build and their line of thinking. Clearly, at no point before releasing their LLMs onto the world did anyone stop and ask: Hey, how do we deal with these things generating unwanted content?

Resorting to blocking certain terms in the prompts is like searching for keywords in spam emails. "Hey Jim, I got another spam email from that Chinese tire place" - "No worry boss, I've configured the mail server to just delete any email containing the words China or tire".

Some journalist should go to a few of these AI companies and start asking questions about the long term effectiveness and viability of just blocking keywords in prompts.