Is chat GPT a danger to humanity?

[ZRX61]

In the US a ChatGPT program recently created false allegations of sexual harassment against at least three people which included creating a newspaper report falsely credited to one of the big dailies & quotes from that report.

[Horse]

In the US a ChatGPT program recently created false allegations of sexual harassment against at least three people which included creating a newspaper report falsely credited to one of the big dailies & quotes from that report.

What was/were the brief and any qualifiers or refinements used?

[ZRX61]

In the US a ChatGPT program recently created false allegations of sexual harassment against at least three people which included creating a newspaper report falsely credited to one of the big dailies & quotes from that report.

What was/were the brief and any qualifiers or refinements used?

Apparently some researcher at UCLA asked *it* to find examples of sexual harassment. It came up with 5, 3 of which were complete bullshit & never happened. The researcher contacted at least one of the people named.

[MrLongbeard]

This is being discussed on another forum I frequent, it's basically makin' shit up

Seriously, tho', it's bobbins.

Take a look at this:

[ATTACH=full]85365[/ATTACH]

Exactly one of those six matches the prompt. You'd have better odds throwing a dart at the OED.

How about helping with an academic project?

[ATTACH=full]85369[/ATTACH]

Wow, that sounds like it would be an ideal reference for my work - it's exactly what I asked for! Let's have a look at the DOI...

[ATTACH=full]85370[/ATTACH]

Well, that's not a good start. The journal's real, and David Weaver's even written for it, but...

[ATTACH=full]85371[/ATTACH]

Yeah, there's no such article. Never was. ChatGPT just made it up, because it's exactly the format you'd expect for a response to my prompt. It's perfect, it has everything, it even breaks it down with an explanation of how the reference is formatted and "why it thought" it'd be a great reference.

It's a 10/10 ChatGPT response, perfect, no notes, exactly what it was built to do. The fact that it's also complete nonsense and a total fabrication doesn't enter into it. And until it does, I'd not recommend using any of the GPT models for owt important.

Hell, even the DOI is "right" - 10.1177 is indeed the prefix for the journal, and it's then followed by a bunch of numbers. Granted, it's important that they're they're specific numbers, but that doesn't matter to ChatGPT. The format is "10 dot 1177 slash bunch-o'-numbers," and that's exactly what I got. Great.

EDIT:
And, just to finish the job off, pp. 2027-2043 of Volume 19 Issue 12, which was indeed published in 2017 so well done ChatGPT on that one, would get you the latter three-quarters of Hatakka's When logics of party politics and activism collide: The populist Finns Party's identity under negotiation and the first few pages of Berkelaar's How implicit theories help differentiate approaches to online impression management: A preliminary typology.

EDIT EDIT:
[ATTACH=full]85372[/ATTACH]

(Snipped the explanatory bit for brevity).

Guess what? This article, which "provides a sociological perspective on enthusiast computer overclocking," doesn't exist either!

[ATTACH=full]85373[/ATTACH]

This time the DOI's real, but points to an article published in 2013 in 26(5) of the same journal, titled Self-certified proxy convertible authenticated encryption: formal definitions and a provably secure scheme. Nowt to do with overclocking at all, and by completely different authors.

[ATTACH=full]85374[/ATTACH]

Yeah, we're back to correct-format but made-up DOIs again. There's no such article, which is a shame: a paper that "provides a philosophical perspective on the relationship between overclocking and emotion" sounds fascinating.

I could keep going. It's never going to give me a valid paper. It's not designed to. It's just doing what it's supposed to: giving me correctly-formatted responses to my prompts.

EDIT EDIT EDIT:
Keep calling it out for long enough and it sort of gives up, telling you to go to Google Scholar instead. The twist?

[ATTACH=full]85375[/ATTACH]

Yeah, that paper doesn't exist either. It didn't search Google Scholar. It just knows that a response to being called out about making stuff up should include a reference to where you found something, so it makes that up too.

Then:

[ATTACH=full]85376[/ATTACH]

Yeah, none of those exist either. Bonus points for the correctly-formatted link to a TechSpot article that doesn't exist.

It never stops. Right after that, it gave me another three "references," all of which are fictitious. I could keep going, but there's actual work to be done. Y'get my point, though.

[Horse]

In the US a ChatGPT program recently created false allegations of sexual harassment against at least three people which included creating a newspaper report falsely credited to one of the big dailies & quotes from that report.

What was/were the brief and any qualifiers or refinements used?

ChatGPT just made it up, because it's exactly the format

It's just doing what it's supposed to: giving me correctly-formatted responses to my prompts.

[ZRX61]

We've got some muppet on a FB ZRX group telling us how great it is

"So I've been farting around with snapchats "my AI" feature. Today I spent quite some time talking about the possibilities on why my zrx1224r build failed. And I think we came up with a solid answer. This AI shit is absolutely wild, using it with what you know... keep in mind I did have to correct it when it said the zrx1200r had a 6 speed transmission. Then it said I was right it has 5. It's free too which is awesome."

"Imaging getting advice on a problem you're having with your motorcycle (or anything else you can think of) and being able to ask a super smart artificial intelligence what they think the problem is."

[Horse]

Apparently the people who write essays for uni students are suffering a fall in business.

[Yambo]

I don't know much about AI (probably like most of you lot) although a script I can add to my astro images processing program was developed with/using AI.

What I think I do know, is that AI is going to be lacking in the emotional department. You can't teach emotions, you can't teach (for a machine to recognise) gut feelings and you can't teach a machine fear, trepidation, concern etc.

To quote our new king ("whatever love is . . ." ). AI ain't gonna crack that one any time soon.

[Count Steer]

What I think I do know, is that AI is going to be lacking in the emotional department. You can't teach emotions, you can't teach (for a machine to recognise) gut feelings and you can't teach a machine fear, trepidation, concern etc.

I don't suppose they'll dream either. Not even of electric sheep.

[DefTrap]

Yeah some people are still snooty about smart phones and television, they were going to take over the world too.

Man is a super selfish bastard survivalist. We kill, imprison and enslave each other at the drop of a hat. Use AI to kill each other and / or otherwise gain advantage? Absolutely. I strongly doubt that it'll be allowed to run amok though. And if it does then obviously it'll be accidentally-on-purpose / the-chinese'-fault.

[Mr. Dazzle]

I saw an interesting vid the other day about AI playing Go. It was big news a few years ago when a computer beat the top human Go player, it was supposedly the last great challenge of computer game playing. A bit like IBMs Deep Blue and Chess back in the 90s.

Recently an amateur player managed to beat one of the top AI Go players, a piece of software which is supposedly unbeatable and even more advanced than the one which can beat the world number 1. The human didn't just win, he won 14 games to 15.

How? He played using a strategy which would have been obvious to a human but runs contrary to the "correct" way to play. The AI can't beat it because the computer has no fundamental understanding of what the game is. It just knows that by making certain moves it wins, based on its training data. It has no intuitive concept of the actual point of the game. It's extremely narrow minded, like all AI at the moment.

[Mr. Dazzle]

What I think I do know, is that AI is going to be lacking in the emotional department. You can't teach emotions, you can't teach (for a machine to recognise) gut feelings and you can't teach a machine fear, trepidation, concern etc.

I wouldn't be so sure about this BTW.

AI software doesn't really work like any other kind. Its built on nueral nets just like our brains use and indeed the whole concept behind this came about from studying brains. You don't teach it rules like a traditional bit of software and a central theme in AI is that no-one really knows how the nets develop.

Who's to say that as these nets get more complicated that feelings won't emerge, just like they do in our brains?

Really good video on how the nets work;

[Count Steer]

I was reclining on my hand embroidered, silk, Persian coverlet, contemplating things in general and my own in particular* when it occurred to me that if AIs are going to replace so many people, fewer actual people will be creating content for these AIs to soak up, they'll all just be reading each others output eventually. Well, that and bike forums, Instagram cat pictures and Faceberk.

The future starts to look like a bit of a funny farm.

*50 gold Internet points if anyone knows that particular reference.

[Count Steer]

I've been following a discussion about ChatGPT and cryptic crossword clues. Someone gave 3 clues to a particular word. ChatGPT gave a very erudite analysis of them and came up with 3 (incorrect) solutions.

They then gave it the same 3 with the solution and asked it to assess and rate them. Again, it gave a very credible looking response but failed to parse the clues properly (ie the word was 'sparkle' and one clue referenced cryptically the horse Arkle. It failed to spot this and came up with a totally fallacious but convincingly written parsing).

The problem or one of the problems is that the responses read very convincingly. It's a bit like a politician...say something positively and convincingly enough and you're more than half way to establishing it as true. People question the validity of some Wikipedia content because people can't be trusted but if a computer comes up with something they'll probably take it hook, line and sinker.

[Count Steer]

I think we need to worry about who controls the AI first before worrying about the AI being 'motivated' to establish control over people but I take your point if they can erm, evolve, so quickly Skynet may not be that far off.

Currently they seem only able to answer as if they are certain they are correct, there's no shade in the responses. It's like 'this is true' every time, even adding in references (which may not in fact be real) because the aim is to create a convincing response rather than a correct one. They can't seem to create one that says 'Hmmm, I'm not absolutely sure about this but...' or 'this is said with a certainty of 63% because the following data is unavailable, pleads provide x, y or z in order to increase accuracy'.

It all seems a bit gung ho.

[Count Steer]

Well I am being deliberately polite with Chat GPT (just in case) but it did once apologise to me during a chat regarding climate change. I'll see if I can dig out the transcript. It's quite interesting...

edit> Damn. No it's been auto deleted and I have swapped systems twice since then so even if I have a transcript, it will be buried. Something along the lines of "will global warming during this interglacial period delay the return to the ice age proper"....

It was all too keen to deliver the typical doom and gloom narrative without investigating any alternative.

Shame, it would have been interesting to see.

The other thing that intrigues me is the 'evolutionary driver' for AI. People keep saying they will 'evolve' quickly. There's a misconception that evolution is driven by some quest for perfection/survival when it's actually an observation of the outcome of multiple, almost random, failures. Something happens over generations of assorted changes in the DNA. Anything that doesn't work dies out. What we see are the ones that just happened to not fail but it looks like there's a driving force pushing things to towards survival in any particular environment. There isn't there's just a huge pile of failures that we don't really see. It takes a lot of random changes, generations and time and most changes fail. How's that work with AI? Sure they sort of adapt very quickly but, there aren't many, and failures could be catastrophic (for the AIs and everything else).

I suppose the only drivers are the parameters and criteria that the developers put in them. ie 'success looks like this'. So again, it's the owners of the AI that are going to be the problem. The AIs just keep adjusting to meet the criteria that humans put into them. I guess it's evolution of a sort.

[Mr. Dazzle]

I posted in another thread i think (scrolled up but can't see it), but check out 3Blue1Brown's videos on YouTube about how Nueral Nets work.

In fact check out loads of his videos if you wanna understand some mathematical concepts. He's really good at explaining stuff intuitively. His video on Fourier transforms gave me a way better insight on them even though I've been using them for years. Ditto on where e comes from.

Re: Evolution, I saw another really interesting point on AI primers/targets on a video about AI and economics. This guys reckons "AI primer writer" will be a big career in future, I find it hard to disagree with that assessment. I.e. knowing what question to ask the AI and how will be a job.

[ChrisW]

Re: Evolution, I saw another really interesting point on AI primers/targets on a video about AI and economics. This guys reckons "AI primer writer" will be a big career in future, I find it hard to disagree with that assessment. I.e. knowing what question to ask the AI and how will be a job.

This really does remind of Neuromancer.

[Count Steer]

Thanks. I think that sums it up really. It does its best to come up with credible answers that satisfy the question but it only has 'limited' data to work with. A lot of the learning process for humans is finding out, in a non lethal way, that you've got it wrong. Who tells an AI 'This is incorrect?' What would need to happen for an AI to say 'I'm sorry, I was wrong'?

It would be interesting, as it develops, to ask 'If this is true, you and the rest of us will cease to exist, what do you think we should do about it?'

[Mr. Dazzle]

Sounds like the AI is apologising for Screwd's misunderstanding, which means it's reached the level of a politician.