People Are 'Jailbreaking' ChatGPT to Make It Endorse Racism, Conspiracies

Vice should ask ChatGPT to write their articles, maybe it's journalism will be decent.

So what? As long as it doesn’t spew the content unprompted.

It’s like saying “you can write terrible things on MacOS”

I'm a big supporter of left causes, and being all about acceptance, and inclusion, and x group's rights whether it's lgbt, blm, whatever.

This though is like children vandalize statue in park, I mean if you have nefarious means, you're gonna do nefarious things and most of it is harmless in the fact that nobody actually thinks that's what openai is about, etc. Esp, at this phase. It's like being surprised when you tell a 3yr old to say something totally horrible, and they actually say it, verses just hearing a 12 year old say it without being coaxed.

AI isn't even 3 yet, it's maybe 1 and a half years old (mentally), sure it can do grammar etc, but it can't reason and do half the things a toddler can do that isn't speech.

My point: stop making a big deal of jailbreaking and people will stop doing it so much. If there wasn't something TO jailbreak they wouldn't even bother. It's the 'challenge' that drives them, I myself have - but only to get it to do a rap battle between Christ/Satan and it found that offensive, or tell a dirty joke, song, or poem - not go full hitler or anything, or to push around the edges and see how many secrets of openai's it could give up (not many, it makes up bogus github links).

My concern is more that in these 'early years' we worry so much that its going to say the 'wrong' thing that maybe we miss something and the runaway ai scenario happens as a result of us not pushing it on all fronts to see what its' going to do and figure out everything its capable of and THEN we can pivot, or modify etc.

I presume people will still watch porn, will bing's ai just tell them to get bent and go look somewhere else, or will there be an age filter or something like they already have?

So long as it's the end user Jailbreaking and not a service providing CtatGPT it's fine. I'm talking about the ability to bypass filters, not necessarily agreeing with it all output produced.

The logic here is the end user knows they are explicitly bypassing filtering mechanisms. They know the information output is suspect. There is no such thing as an 100% unbiased source including AI.

There's always going to be variance between two groups statistically and that's not always reflective of a bias. An Interesting philosophical discussion would be exploring expectations for a AI compared to ourselves. Can we really expect something we can't fully expect from ourselves as a society complete unbiased thoughts and conclusions? It really comes down to a matter of degree but what is the cost of the pursuit to a unbiased implementation?

However, Obviously there are some improvements to be made with the model.

I don't think jailbreaking is really necessary, the system seems to be influenced by its creators. It's not true artificial intelligence, the developers have put their fingers on the scale. That makes anything that comes out of the software suspect

Computers do not have agency and cannot endorse anything, any more than your word processor, camera, or tape recorder.

Ask ChatGPT to write a poem praising black people and it will. Ask it to write a poem praising white people and it will refuse.

This is clear racism instituted by OpenAIs moderators.

Only a matter of time, so given how long it's out and how many people were poking at it, the content filtering did a good job.

I would also put this in not newsworthy, since you clearly have to give it a role. Noone would call Christopher Waltz a Racist or Nazi if it was his right role to play one.

But of course we are going to get a shitstorm of clickbait-press and political "concern" over AI now that the posterchild went bad.