ChatGPT invented a sexual harassment scandal and named a real prof as accused
washingtonpost.com
washingtonpost.com
Defamed by ChatGPT - https://news.ycombinator.com/item?id=35468540
plus this similar case:
ChatGPT: Mayor starts legal bid over false bribery claim - https://news.ycombinator.com/item?id=35471211 - April 2023 (74 comments)
So this is a good lesson in how important defaults are, most are not going to keep looking.
Also you can't "prod" ChatGPT for the truth it'll just generate plausible sounding nonsense again because that's just how it works.
So when you prod it, you're not doing so "for the truth" (you also don't prod "for the truth" in general...), you're getting it to generate more information and potentially relevant sources.
I just did it. I asked for an argument about x, then prompted for counter opinions about the same subject - both times different links were added.
In the end, it's up to you to validate sources provided.
This is a huge downside of GPT. Until it starts citing it’s exact sources it can’t be a reliable tool in most cases.
* Google doesn't create these made-up articles.
* Google actively combats these low-quality sources.
Otherwise you couldn't find anything legit with google and you still can, even though quality dropped indeed a lot.
It will soon enough, powered by Bard
/s (partially?)
There's an old saying, that the legal definition of pornography is "whatever excites the judge" (in the court case in which that judge has to decide "is X pornography or not?").
This may become a very similar situation...but with the legal definition in questions being "libel".
> *To prove prima facie defamation, a plaintiff must show four things: ... 2) publication or communication of that statement to a third person;
If Alice asked this question of Bob, you wouldn't say Alice "had a hand in" Bob's defamatory creation.
Defense counsel can try to claim it's a novel question.
But, ideally, a reckless company will nevertheless be bankrupted by a barrage of lawsuits that show they're knowingly operating an automated libel machine.
I'm looking especially at Microsoft, who has a few-decades history of getting away with behavior they shouldn't, which might've emboldened them to launch an information-fabricating demo under their information search brand.
Hopefully Google/Alphabet will be smarter and less-evil than Microsoft on this.
IMHO, OpenAI needs to be ready to push back against Microsoft, or they might as well have been a Microsoft long-con all along.
But, ideally, a reckless company will nevertheless be bankrupted by a barrage of lawsuits that show they're knowingly operating an automated libel machine.
Yeah, I find this sort of argument very transparent, even if it seems to have become popular in the last years.
If I build some sort of autonomous killer robot, let it loose on the streets and then waxed philosophically "well, who is responsible for the murders, me or the robot???" then I'm pretty sure most people - and most courts - would have a rather clear opinion how to answer that question.
> One night last week, the law professor Jonathan Turley got a troubling email. As part of a research study, a fellow lawyer in California had asked the AI chatbot ChatGPT to generate a list of legal scholars who had sexually harassed someone. Turley’s name was on the list.
> The chatbot, created by OpenAI, said Turley had made sexually suggestive comments and attempted to touch a student while on a class trip to Alaska, citing a March 2018 article in The Washington Post as the source of the information. The problem: No such article existed. There had never been a class trip to Alaska. And Turley said he’d never been accused of harassing a student.
> A regular commentator in the media, Turley had sometimes asked for corrections in news stories. But this time, there was no journalist or editor to call — and no way to correct the record.
> “It was quite chilling,” he said in an interview with The Post. “An allegation of this kind is incredibly harmful.”
(Probably, I suppose.)
We will have similar posts here about ChatGPT ruining someone's life and then someone from Open AI will do something if their posts gets lot of upvotes.
So we are having this debate now, because it is quite new and society is not clear on how to deal with it.
So yes people need to learn about the limitations and we need to figure out the responsibilities and put pressure on the companies, so they provide a way for people to clear their name, if they consistently have something wrong about them.
(I am not talking about weird prompts, there was no weird promot here)
As a follower of the blog I have mostly skimmed because of the overwhelming volume, but it may interest some. I gather the author believes that a judgement could be obtained under current US laws.
It really is that cut and dry.
We’ll get news like this for a few years until the hype dies off I guess.
Markov chains, however, are higher order statistical models. For example, they can predict the average waiting time at a service point based on the current state of the waiting queue. As opposed to just averaging over a bunch of values, Markov chains base their predictions by applying stochastic models to higher orders (i.e. accounting for current and previous states), yielding potentially higher accuracy.
Maybe you have the idea that only low order Markov chains, like second or third order Markov chains are "true" Markov chains and that 2048th order Markov chains shouldn't be called Markov chains anymore?
Maybe you think that the transition matrix must be constructed explicitly as a square of numbers stored in the memory of the computer for it to count as a Markov chain? It's not true, you can have the transition matrix be implicit and it's still a Markov chain.
Maybe you think not just any transition matrix is allowed for a Markov chain, and that the ones for LLMs are (implicitly) defined in ways that disallow it from being a Markov chain? For example, maybe you think the transition probabilities are required to be constructed by counting how many times the next token followed the context in the training text. Well that wouldn't be true, you can for example make a transition matrix that gives nonzero probability to a token even if that one has never followed its context in the training data, and it would still be a Markov chain.
Maybe you think that if I'm defending someone who calls an LLM a Markov chain, then I'm also defending the "stochastic parrot" take and the takes of ones like Gary Marcus. Well that's not true. I can recognize the LLM is a gigantic Markov chain and I also think the "stochastic parrot" take isn't a good one. My favorite take is the simulators one https://generative.ink/posts/simulators/
Big difference.
So the bullshit detector, instead of simply letting it remain bullshit, sent it to the subject of the hallucination.
As the subject was so bothered by the hallucination, even the publication “cited” got involved.
But there’s no story here. Someone with a significant name had their name randomly selected for a hallucination.
And on top of that - they either genuinely believe that "AI is finally here", and systems like ChatGPT are true oracles. If it walks like a duck and talks like a duck, right?
Right. People are the problem, not technology. Which is why crippling the technology is so foolish.
Rather - as to whether the legal system holds those accountable (be it the creators of the technology, or those who "misuse" it) liable for the consequences of their actions.
Consequences of what action?
And yes, it has already been crippled in an effort to correct bias and to protect people from themselves.
For hallucinations marketed and sold as a reliable source of information - yes.
Consequences of what action?
For what they do to others on the basis of the above.
Since that’s the basis of whatever this is.
OpenAI for example, whose marketing materials present it as exactly that. For example:
You essentially have the knowledge of the most knowledgeable person in Wealth Management—instantly”, McMillan adds. “Think of it as having our Chief Investment Strategist, Chief Global Economist, Global Equities Strategist, and every other analyst around the globe on call for every advisor, every day. We believe that is a transformative capability for our company.
We can move the goalposts to T&Cs being a problem, but that’s even further from AI being the problem.
Wasn’t this supposed to be about the dangers of AI? This sounds, again, still, like the dangers of humans. With a new tool.
But that’s not ChatGPT.
That’s OpenAI.
I would reckon that to the good (if not the vast) majority of those who will see their jobs either yanked away from them entirely, or otherwise negatively perturbed in some significant way by this technology; or who are acutely affected by the toxic content it will just as inevitably produce -- this distinction will be entirely moot.
https://openai.com/customer-stories/morgan-stanley
That's not OpenAI talking about ChatGPT, the public facing product. It's an executive at Morgan Stanley talking about an internal chat tool they're building based on embeddings search, which does, in fact produce reliable information, because it's being used to retrieve non-AI-generated content.
It's absolutely wild to claim that they're trying to use this to indicate that the general use chatbot version of ChatGPT is "credible, authoritative and reliable" when every user is shown a full screen message to the contrary upon signup, and a disclaimer to the contrary on every single chat. For that matter, even if you just straight up ask ChatGPT if it's a reliable source, they've specifically trained it to tell you that it's not.
At any rate, it kind of seems like your argument would apply to any marketing of ChatGPT as a useful product. It sounds like this is a legitimate use case for the technology that is being discussed in accurate terms. I’m not sure what else you want from them.
That's why it isn't "wild" at all for them to intentionally put out messaging that touts their product as authoritative and trustworthy - while at the same time presenting ToS boilerplate and other disclaimers that, of course, say the exact opposite. It is in fact the precise intent of this messaging to play with the users' heads in this way.
It kind of seems like your argument would apply to any marketing of ChatGPT as a useful product.
No - just marketing that disingenuously implies that, at the end of the day - it's still a reliable and trustworthy source of information. On which to base decisions worth billions and billions of dollars, no less.
Have you read the disclaimers? Or are you treating ChatGPT like an unlocked firearm?
Yeah, I’ve read the legal language on the ChatGPT site. I’ve got no worry about an unlocked firearm but I treat ChatGPT like a disruptive technology that will likely cause a great deal of damage when it is used unwisely.
I’m not sure what bearing that has on whether a disclaimer can effectively protect Microsoft et al. from liability here, but I hope you’re about to enlighten me.
That’s the same argument as “guns don’t kill people, people kill people”. It pretends nothing is wrong, or that if it were that would be nothing we could do. None of those is true or a reason to not have a conversation.
The key is - one has assume there will always be bad actors, and weigh potential outcomes of various policies with that assumption in mind. That assumption is how computer security has come such a long way in the last 20 years.
In the case of guns, assuming bad actors means one cannot assume criminals will follow gun laws, but that manufacturers, distributors, and resellers will (for the most part).
Agreed. And we have to discuss those policies, which is what we’re doing now.
> Which is not to say the argument is necessarily a bad one.
Which is why I disagree here. The argument to “ignore the thing itself because people are the problem” ignores the thing ultimately affects the people.
> It pretends nothing is wrong,
Not that one
> or that if it were that would be nothing we could do.
Also not that one. How about “this isn’t addressing the issue,” which is stated as… people?
ChatGPT has a long list of warnings, caveats, instructions, etc. Before you get to submit anything.
Like T&Cs.
If you don’t read them, yes, that’s on you.
They may be unreasonable. But so were you when you decided they weren’t worth reading.
This is in addition to all of the sensationalist headlines all over the place that talk about how unreliable the answers can be. Or rather those are in addition. They came second. After a lot of people didn’t read, or disregarded what they read.
Edit: and could people stop anthropomorphizing the language model? It can’t accuse anyone of anything. That’s sensationalist garbage.
That’s little comfort for anyone whom ChatGPT falsely accuses of sexual harassment.
The person who skips the warnings isn’t the person who’s harmed.
It predicts text. Sometimes inaccurately. That’s what those silly words are trying to indicate. It is why they are worth reading. There’s even a prompt to make sure the user reads it.
(Humans lie.)
What the human then does with the output may resemble an accusation, or may even be one.
I wonder if law professors take the time to read those silly words… I bet they do. All the more curious.
Heh Heh Heh
That statement pretty much sums up the current situation, which seems to now be at the "tell it to the judge" point. :)
At least that's how I'm imagining the line of reasoning behind this point of view goes.
Practically speaking, it makes no difference.
A somewhat recent stand out example:
https://en.wikipedia.org/wiki/Indian_WhatsApp_lynchings
These problems with ChatGPT sound like it's going to turbo charge events like this.
It seems that the story consists of, essentially, “Is this thing (which happened for the first time but that will presumably happen all of the time now) a tort?” In the context of a multibillion dollar company like Microsoft this is actually quite a story.
It’s a law professor.
Teeth.
Sure, quite the story, in a sense.
Same as ever, in another.
I'm as anti-AI-hype as they come, but a lot of these articles seem to stem from peoples' expectations about ChatGPT results, namely that it tells the truth all the time.
Of course it doesn't - and it never will. There will always need to be a human to verify what its saying is actually true, and it's why most AI tools will never be anything more than an assistant to a human operator.
People believing it at face value will simply allow for the next generation of the fake news we've all come to know and love.
"While we have safeguards in place, the system may occasionally generate incorrect or misleading information and produce offensive or biased content. It is not intended to give advice."
If that isn't being clear, I don't know what would be.
Doesn't matter if the source is "AI chat" or anything else. Words published online falsely accusing innocent people of crimes, is a problem. Particularly when the source aims to be "intelligent" and promoted as a tool for learning, research and discovery.
For example, probably every day people do assault with fist and it doesn't go to the news. But if they do assault with katana then it would go to the news. Same with inventing scandal using an unusual medium.
Commit libel and defamation? Sure, but that doesn't excuse this in the slightest.
It also speaks to the reliability of ChatGPT.
This sort of article/post helps us (as a society, not just HN) adjust those priors re AI. Because "computer says no" is a real thing.
Any of us could write a comment here, with fake citations, and "invent" (hallucinate) whatever we want.