I'm creating this thread so that there's somewhere to direct questions about "ethical" considerations in using AI and LLM tech.
#Ethics, rights and activism.
fresh robin
I'm also not opposed to moderate discussions about how best to protect the rights of content creators (and readers) , such as the ability to create mature/controversial content with LLM's, and preventing LLM providers from imposing overly restrictive practices on their use.
We need a 'First Amendment' or 'protected speech' law for AI generation.
It should be competent readers and legislators that determine what might be "inappropriate" in a given situation, and not solely the content/usage policy of large providers.
Ethics, rights and activism.
gaunt vapor
Saving the link to this thread
gaunt vapor
We need a 'First Amendment' or 'protected speech' law for AI generation.
I'm curious, what type of rights/restrictions do you think should be enshrined in such a document?
fresh robin
The same that apply to other media.
I mean you don't when you buy a pen got told you can't use it to write naughty but legal stuff..
You don't have a clause in a typical EULA (well you used not to) syaing you can't use a word-processor to write a 'sex' novel.
You don't (or used not to get into trouble) for using a specifc manufactuers to take risque pictures of your girlfriend with her informed consent.
In the US, you have the situation (not that I agree with it) that you can have very strongly conservative evangelicals, producing religous material on public access cable etc.
gaunt vapor
I mean you don't when you buy a pen got told you can't use it to write naughty b...
I suppose nobody spells out the laws for you, yeah, but there's plenty of stuff that you'll get in trouble for writing. even in the US where we have pretty strong protections on freedom of speech there are still limits
or do you mean like, we should legislate AI text generation companies and mandate that they must allow production of certain classes of text, like sex novels?
fresh robin
I mean that they shouldn't necessarily be allowed to rule entire classes of work which are legal.
If there are good reasons (such as potential for public harm) however...
And of course if certain material IS illegal already then LLM's shouldn't be generating it.
gaunt vapor
And of course if certain material IS illegal already then LLM's shouldn't be gen...
well yeah, no argument here on that point.
I sort of agree with you in spirit but I don't know how it would actually work
I think a good example (from the US) would be Apple refusing to help the FBI unlock that mass shooter's phone
fresh robin
I would also say we need a Section 230 equivalent for the providerd of LLM's...
gaunt vapor
right now GPT-4 and ChatGPT are fine-tuned to avoid writing sexual works, so if a law went up enshrining the rights of people to use commercial LLMs to generate sexual works
fresh robin
(I assume you know what Section 230 is?)
gaunt vapor
it wouldn't just require unchecking the box that says "filter sex scenes"
it would require them to assemble a training data set and retrain the model so that it can produce such content
right now it's not an external filter (at least, not that I'm aware of) but encoded within the weights of the model to not generate sexual content
who pays for the compute costs of retraining existing models? and who pays for the data scientists to assemble the data set?
fresh robin
Well obviously you can't do much about existing models, which is why it's 'not restrict in policy' type of argument, as opposed to a 'must' allow and mandate to retrain.
It's a complex issue I know 😦
gaunt vapor
I disagree on the relation of section 230, since this regulation would be mandating that a company must produce a certain content versus simply saying that they aren't responsible for things other people produced
fresh robin
I'm not saying they should 'produce' a certain type of content...
I am saying they should not be allowed to stop people attempting to generate it with a given model. Important difference.
Provided that, being asked of the model is legal of course.
gaunt vapor
I suppose so, but where does the line of what they're willing to sell come into play? for example, I agree that it shouldn't be illegal for a book store to sell sex novels, but I don't think they should be legally mandated to sell sex novels.
fresh robin
yes , but do you get into toruble for asking a bookstore if they can find a certain type of book for you, that's otherwise legal ?
Certain artists come to mind..
My comment on Section 230 was I agree not the same issue, but a related one..
gaunt vapor
yes , but do you get into toruble for asking a bookstore if they can find a cert...
but you don't get in trouble for asking GPT to find you certain info either, right? it's not like it reports you or anything, it just doesn't give you the output you wanted
fresh robin
True.. but the way some LLM's word their content/usage policy, you could have something generated, that per it's usage policy it wasn't supposed to.. (like people using so-termed jailbreaks to get NSFW content).
My views on Section 230, is if that was extended to cover Ai generated outputs, LLM content providers could have a little more certainty and flexibility and thus might be prepared to be slightly more flexible. It also would rebalance responsibility back to users/readers of content.
That's probably an optomistic view though, given that you have had tech firms subject to 'mob-lobbying' and 'advertiser pullouts' cause some social-media platform didn't cancel this weeks Ranter of the week on some more controversial issue.
As I said elswhere, some of the banned content is to me more about liability issues than it is about actual harmful content being generated.
gaunt vapor
As I said elswhere, some of the banned content is to me more about liability i...
I think the issue of legal liability is an interesting one when applied to mandates about what LLMs must output
Sam Altman and co don't strike me as prudes. I'm making an assumption here, but I don't think it's that the people working at OpenAI are averse to anything sexual. I think they made their LLM not output sexual content as a business decision
if the US were to introduce legislation that mandates that they can't deny certain texts from being generated, doesn't that create an incentive for AI companies that aren't jerkoff chatbot companies to move to a different legal locale to avoid harming their bottom line?
gaunt vapor
they do, but I think that with any theoretical legislation we might introduce, we have to consider whether this law would be the one that makes them finally pack up and leave
fresh robin
Of course on the other hand, if firms had more clarity (like making users responsible for 'bad' content rather then the hoster), they might be prepared to be more flexible.
I mean web-hosts don't get sued because some news website hosted someone a little too conservative and said something they shouldn't have , or do they?
(Note : I am from the UK, where the 'publisher' of certain things CAN get sued as well 😦 )
Of course if LLM providers wanted to have 'safe-modes' as well as unrestricted ones, I don't see that as being an issue. Provided you can turn the safe mode on and off as a choice...
I mean that's how search engines now do it...
To be fair, that might be a sensible thing for some LLM's in the future...
having a 'safe' and 'Nsfw' mode as seperate interfaces ?
Provided they allow an unrestricted interface alongside the 'safe' one 🙂
@gaunt vapor : On something else , what are the Four Freedoms LLM users need ? ( https://www.gnu.org/philosophy/free-sw.html#four-freedoms) for context?
Currently most LLMS's aren't 'free' by those standards...