Safety experts quit OpenAI

[Schadrach]

But if there's no "answers", the AI has nothing to pull from.

Exactly. If the top search results are misinfo, disinfo, unrelated, memes or simply wrong then the Google AI bot will take those and confidently summarize them as a factual response to what it thinks you are looking for. Because that's what it does - it's essentially the result of an LLM being asked to summarize the first few search results for your query, whatever those might be. Hence the gluing cheese on a pizza one - it is from a highly upvoted comment on the Reddit thread that's the top result about getting your cheese to stick to your pizza.

 

[Phoenixmgs]

Exactly. If the top search results are misinfo, disinfo, unrelated, memes or simply wrong then the Google AI bot will take those and confidently summarize them as a factual response to what it thinks you are looking for. Because that's what it does - it's essentially the result of an LLM being asked to summarize the first few search results for your query, whatever those might be. Hence the gluing cheese on a pizza one - it is from a highly upvoted comment on the Reddit thread that's the top result about getting your cheese to stick to your pizza.

I am more so talking about an AI making up it's own news stories. I feel like it would be really bad at that as there's nothing for it to pull from essentially. It can pull from past elections and what was important issues but then it would have no idea if that is currently an issue. I still think you'll always need humans to deceive humans, then you can use AI/bots spread it faster. But AIs being some massive new threat is not gonna be a thing.

It's like with AIs driving cars, it's not that hard (it takes a lot of man hours coding them no doubt) to get the AIs to like 80-90% and then people are like AIs are so close but in actuality getting that last 10% is basically never going to happen. How many stories were there of self-driving cars being a year or two away like 5 or so years ago? The same thing will happen with these chat AIs. ChatGPT got released and everyone is astonished how well it works but it ain't going to get that much better.

 

[Phoenixmgs]

Yet to anyone who isn't able to actually look up case history, those references all looked believable. They included dates, names, page numbers, case titles. All fabricated.



But the number that gained traction wasn't the real number. It was a complete falsehood.




Right, but far more people-- millions upon millions a year-- fall for spam and scams.

It didn't even trick one judge.

But the real number could've gained just as much traction.

You need to convince people on a population level. Do you not get that?

 

[Silvanus]

It didn't even trick one judge.

!?! Yes, a judge, someone with years of legal training and access to all relevant case history. But it can clearly create the level of detail-- spurious dates, names, figures-- to make something appear superficially believable.

But the real number could've gained just as much traction.

...but it didnt. The lie did, proving that lies can gain wide traction.

You need to convince people on a population level. Do you not get that?

Do you not get that convincing a few million, or even a few thousand if they're optimally distributed, is enough to swing democratic votes?

 

[Phoenixmgs]

!?! Yes, a judge, someone with years of legal training and access to all relevant case history. But it can clearly create the level of detail-- spurious dates, names, figures-- to make something appear superficially believable.



...but it didnt. The lie did, proving that lies can gain wide traction.



Do you not get that convincing a few million, or even a few thousand if they're optimally distributed, is enough to swing democratic votes?

It sorta has to get by a judge.

I never said lies can't gain traction.

Again, never really said that. I said AI can't do it anymore dangerously than it currently is.

 

[Agema]

The Google AI bot is really good at synthesizing the top few search results for a question and expressing that information in an authoritative sounding manner. Unfortunately, that only works insofar as the top few search results actually provide a good answer to the question asked.

I'm also sure a large part of that is the fact that the more that it accesses to manufacture a result, the more it costs to run. AIs are not cheap to run.

I feel like you could take out the "just", slightly more advanced search functions could be immensely useful.

Well, that depends. Google already lets a load of sponsored shit float to the top of the search function, because Google long since realised that the profit was better taking money from corporations to bump their results rather than provide consumers with what they want. AI is only ever going to end the same way.

I can try to be positive, that AI can benefit us in lots of ways. But mostly, it's going to benefit tech companies. They are going to rampage over society without giving the slightest damn about the damage it may cause as long as their profits roll in. "Move fast and break things", they say. It's funny how they never have to pay for what they break.

 

[Silvanus]

It sorta has to get by a judge.

In a court brief it does. On the Internet it just has to convince Joe Public.

Again, never really said that. I said AI can't do it anymore dangerously than it currently is.

Right, but this was based on the misconception that scale and speed don't matter.

 

[Phoenixmgs]

In a court brief it does. On the Internet it just has to convince Joe Public.



Right, but this was based on the misconception that scale and speed don't matter.

And the public is getting their court info from reading briefings? Also, without that, you had like half the people here interpreting the Rittenhouse case wrong and still wrong to this day.

I said the scale and speed was already as maximized as you can get for human consumption.

 

[Silvanus]

And the public is getting their court info from reading briefings?

This example wasn't brought up specifically so we could talk about misinfo about court cases; you've missed the point entirely. It simply shows what the program is capable of. It isn't restricted to echoing what people wrote elsewhere. It can manufacture details, numbers, stats, names etc. Now think how that ability can be utilised in other contexts.

Also, without that, you had like half the people here interpreting the Rittenhouse case wrong and still wrong to this day.

A complete irrelevance to this conversation about AI. Stay on topic.

I said the scale and speed was already as maximized as you can get for human consumption.

And I already said that's bollocks. You have literally no reason to think this, apart from complacency.

Besides which, if the scale and speed get to a certain saturation point, misinfo squeezes out reliable information. If we assume you're right, and the human mind can only 'consume' a finite amount of info, then what do you think the effect is of adding more and more misinfo into the available pot? It makes it less likely that non-misinfo will make the cut.

 

[Phoenixmgs]

This example wasn't brought up specifically so we could talk about misinfo about court cases; you've missed the point entirely. It simply shows what the program is capable of. It isn't restricted to echoing what people wrote elsewhere. It can manufacture details, numbers, stats, names etc. Now think how that ability can be utilised in other contexts.



A complete irrelevance to this conversation about AI. Stay on topic.



And I already said that's bollocks. You have literally no reason to think this, apart from complacency.

Besides which, if the scale and speed get to a certain saturation point, misinfo squeezes out reliable information. If we assume you're right, and the human mind can only 'consume' a finite amount of info, then what do you think the effect is of adding more and more misinfo into the available pot? It makes it less likely that non-misinfo will make the cut.

I know what it's capable of. We are talking about what kind of threat AI is, not what it's capable of.

Not irrelevant because I said people are believing mis/disinformation as readily as possible and this is all without AI. Even now years after the case, people still believe the narrative they want to believe. And even the president basically called Rittenhouse a white supremacist, how much faster and how much bigger a scale of misinformation do you think AI will be able to accomplish?

Yeah I do. I'm not going to see any more news with AI spitting out stories faster than I am now. And you can't convince me to believe something without actually seeing some real world affects, that takes time. And the news sources I do check in on, they aren't originators of stories but comment on them in essence, and they can only comment on X amount of stories a day anyway. If there's 10 times more stories because of AI, they can't talk about anymore then they currently talk about. We already have 24 hour news cycle.

 

[Silvanus]

I know what it's capable of. We are talking about what kind of threat AI is, not what it's capable of.

...which is directly informed by what its capable of. You disputed that AI can fabricate convincing detail without just regurgitating what humans have written. I provided an example of it doing so.

Not irrelevant because I said people are believing mis/disinformation as readily as possible and this is all without AI.

Nothing you've provided demonstrates that it's at any kind of "upper limit". There's zero reason to think so.

Yeah I do. I'm not going to see any more news with AI spitting out stories faster than I am now.

OK, so you've just missed the point again. It's not about just seeing "more". If you have a limit to how much you consume, then shifting the balance of available content further towards misinfo means that it will squeeze out reliable info.

Say you have a pot of 100 available stories, but an average scroller will get through just 10 a day. Currently, 20 are misinfo, and 80 are broadly more-or-less truthful.

Then a producer of misinformation pumps out an extra 20 fake stories. Now there're 120 in that pot. The scroller isn't going to get through 100, let alone 120... but now, every story they see has a higher likelihood of being fake, because now their odds for each are 40/120 rather than 20/100.

And you can't convince me to believe something without actually seeing some real world affects, that takes time.

And yet people do believe things without seeing real world effects. Including you, because you've fallen for some prime bollocks in the past.

 

[Phoenixmgs]

...which is directly informed by what its capable of. You disputed that AI can fabricate convincing detail without just regurgitating what humans have written. I provided an example of it doing so.



Nothing you've provided demonstrates that it's at any kind of "upper limit". There's zero reason to think so.



OK, so you've just missed the point again. It's not about just seeing "more". If you have a limit to how much you consume, then shifting the balance of available content further towards misinfo means that it will squeeze out reliable info.

Say you have a pot of 100 available stories, but an average scroller will get through just 10 a day. Currently, 20 are misinfo, and 80 are broadly more-or-less truthful.

Then a producer of misinformation pumps out an extra 20 fake stories. Now there're 120 in that pot. The scroller isn't going to get through 100, let alone 120... but now, every story they see has a higher likelihood of being fake, because now their odds for each are 40/120 rather than 20/100.



And yet people do believe things without seeing real world effects. Including you, because you've fallen for some prime bollocks in the past.

If you mean that by the briefing, it didn't fool anyone it needed to fool.

I already pointed out the shitshow that was the Rittenhouse case and how many people believed misinformation (and fucking still do with actual video evidence of all the incidents). How is AI going to make that worse if say a similar case happened with AI pumping out tons of extra stories about the case?

That 20:80 ratio is way off, it's like the reverse of that. Just look at the current crime narratives in the US, one echo chamber is fed all info that crime is really bad and the other echo chamber is fed all info that crime is fine (and you can do that without falsifying actual data, mainly through omission of information).

Again, not people believing said things but people on a population level.

 

[Silvanus]

If you mean that by the briefing, it didn't fool anyone it needed to fool.

How are you missing the point so badly? It demonstrated the abilities you said it didn't have, which would be very useful at fooling people in other contexts. Do you know how to apply principles to other contexts?

That 20:80 ratio is way off, it's like the reverse of that.

Jesus wept... the actual numbers there don't matter to my point. My point: if you have a pot containing two elements, and you add more of one of those elements, then anybody picking from that pot is more likely to be landed with that element than they were before. It doesn't fucking matter that they can't get through the entire pot. It doesn't fucking matter if they have an 'upper limit' on how much they consume. If they pick from it at all, then their likelihood of getting that latter element is increased.

This is an exceptionally simple statistical concept.

Again, not people believing said things but people on a population level.

The population is comprised of people.

 

[Phoenixmgs]

How are you missing the point so badly? It demonstrated the abilities you said it didn't have, which would be very useful at fooling people in other contexts. Do you know how to apply principles to other contexts?



Jesus wept... the actual numbers there don't matter to my point. My point: if you have a pot containing two elements, and you add more of one of those elements, then anybody picking from that pot is more likely to be landed with that element than they were before. It doesn't fucking matter that they can't get through the entire pot. It doesn't fucking matter if they have an 'upper limit' on how much they consume. If they pick from it at all, then their likelihood of getting that latter element is increased.

This is an exceptionally simple statistical concept.



The population is comprised of people.

With the briefing, it really didn't make up anything. It just found basically a template and then filled in names and cases. Did it create an actual new legal argument for example?

The numbers do matter. If your news feed is already over saturated with misinformation, more of it isn't gonna change much. Let's say 9/10 of the news stories you check out is misinformation. So, that means in a week you read 70 stories and only 7 are legit. How much of a difference does that make if that drops to 5 or 6? Is that going to change someone's view of things in any noticeable manner?

People believe the earth is flat. That group is not big enough to vote in a manner to get say NASA to recognize their theory or change their funding in any way. Did the people that believe 5G towers spread covid get anything done?

 

[Silvanus]

With the briefing, it really didn't make up anything. It just found basically a template and then filled in names and cases. Did it create an actual new legal argument for example?

We were literally talking about the ability of AI to fabricate details such as stats. That's what it did: what we were talking about.

A "new legal argument" would be completely irrelevant to our topic, because misinformers wouldn't even want AI to create new arguments.

The numbers do matter. If your news feed is already over saturated with misinformation, more of it isn't gonna change much. Let's say 9/10 of the news stories you check out is misinformation. So, that means in a week you read 70 stories and only 7 are legit. How much of a difference does that make if that drops to 5 or 6? Is that going to change someone's view of things in any noticeable manner?

You moan about my (hypothetical) number, then posit that 9 in 10 stories currently are already outright misinformation?

No, we're not at any kind of "saturation point" where even more won't make a difference. That's just sheer bollocks.

People believe the earth is flat. That group is not big enough to vote in a manner to get say NASA to recognize their theory or change their funding in any way. Did the people that believe 5G towers spread covid get anything done?

Well yes, they did: they drove vaccination rates and support for public health measures to drop to the point where many more people died than necessary.

But of course, you yourself are someone who bought into flagrant misinformation around covid, so we both know where that topic will go.

 

[Phoenixmgs]

We were literally talking about the ability of AI to fabricate details such as stats. That's what it did: what we were talking about.

A "new legal argument" would be completely irrelevant to our topic, because misinformers wouldn't even want AI to create new arguments.



You moan about my (hypothetical) number, then posit that 9 in 10 stories currently are already outright misinformation?

No, we're not at any kind of "saturation point" where even more won't make a difference. That's just sheer bollocks.



Well yes, they did: they drove vaccination rates and support for public health measures to drop to the point where many more people died than necessary.

But of course, you yourself are someone who bought into flagrant misinformation around covid, so we both know where that topic will go.

Again, I never said AI can't fabricate details. I said it can't do it in a believable manner, there's really just random chance that it will.

Any news story that isn't a very objective event like a murder or plane crash is spun by both sides to frame it towards their narrative mainly through misinformation by omission. That's like every non super objective story that's on MSNBC, CNN, FOX, etc. Look how something that actually was super objective like the Rittenhouse case had people at odds with one another depending on whether they watched "left" or "right" news.

Everyone laughed at the people that said 5G towers spread covid. The more and more time goes on, the more right all my stances about covid are. Funny how Fauci just said that the 6 foot distance was bullshit and not backed by any science, exactly what I said from the start because it was just a number they pulled out their asses.

 

[Silvanus]

Again, I never said AI can't fabricate details. I said it can't do it in a believable manner, there's really just random chance that it will.

So you don't think those details it fabricated-- names and dates etc, page references, "bogus judicial decisions with bogus quotes" to use the court's own description-- would be convincing to a layperson? They absolutely, absolutely would. I mean, Gergar has fallen for two less convincing fakes in the last week alone.

Any news story that isn't a very objective event like a murder or plane crash is spun by both sides to frame it towards their narrative mainly through misinformation by omission.

We're not talking about "spin" or bias or whateverthefuck. We're talking about actual misinformation. Outright falsehoods used to convince people.

So, no, you do not have a "90/10" balance in favour of actual falsehood in news stories. There's too much out there already, but the majority of content does not qualify, and we are not even approaching a saturation point. Yet: with more and more churned out-- at ever greater speeds and in ever greater volume by AI-- the likelihood of people encountering it instead of credible sources increases.

Everyone laughed at the people that said 5G towers spread covid.

Except all those gullible people who didn't, and then didn't get vaccinated. Your complacency seems to be absolute; if something doesn't affect you personally, you appear to be incapable of considering it a problem at all.

 

[Phoenixmgs]

So you don't think those details it fabricated-- names and dates etc, page references, "bogus judicial decisions with bogus quotes" to use the court's own description-- would be convincing to a layperson? They absolutely, absolutely would. I mean, Gergar has fallen for two less convincing fakes in the last week alone.



We're not talking about "spin" or bias or whateverthefuck. We're talking about actual misinformation. Outright falsehoods used to convince people.

So, no, you do not have a "90/10" balance in favour of actual falsehood in news stories. There's too much out there already, but the majority of content does not qualify, and we are not even approaching a saturation point. Yet: with more and more churned out-- at ever greater speeds and in ever greater volume by AI-- the likelihood of people encountering it instead of credible sources increases.



Except all those gullible people who didn't, and then didn't get vaccinated. Your complacency seems to be absolute; if something doesn't affect you personally, you appear to be incapable of considering it a problem at all.

The layperson doesn't read briefings so that doesn't matter.

Having outright false information is less likely to convince people, hence you do it by giving legit information but fail to mention the facts that would disagree with the narrative you're pushing (AI can't really do that). The Rittenhouse case is a perfect example. "He traveled to another state!!!" But the fact that's where his father lives is omitted from the story. You keep cutting out all of my Rittenhouse references when it's the perfect example of how people are at polar opposites of an very objective event, all done without AI. How is AI going to make it any worse than it is now?

You're are always going to have people that believe irrational things and are super gullible. AI isn't going to change that.

 

[Silvanus]

The layperson doesn't read briefings so that doesn't matter.

For the third (?) time: the same abilities can be applied in other contexts. Seriously, this has already been directly addressed at least twice, not to mention how it doesn't make any goddamn sense as a response anyway.

Imagine there was a batch of tainted vegetables sold. Someone bought some, cooked them into a stew, but thankfully their guests noticed the stew looked off and didn't eat it.

Your response here is equivalent to saying, "people noticed it in the stew, and most people don't eat stew anyway, so there's no risk". Ignoring the fact that those tainted vegetables will be in other things as well: the stew just brought it to our attention.

Having outright false information is less likely to convince people, hence you do it by giving legit information but fail to mention the facts that would disagree with the narrative you're pushing (AI can't really do that).

I couldn't give less of a shit about your spurious effort to conflate actual falsehood with standard bias. We're discussing one and not the other right now. Recap:

* balance of actual falsehood is not 90/10 or anywhere near a hypothetical "saturation point".
* Adding more misinfo increases the likelihood that any reader will consume that instead of non-misinfo.
* standard bias in reporting is completely irrelevant to this.

You're are always going to have people that believe irrational things and are super gullible. AI isn't going to change that.

Indeed: it's just going to encourage them to engage in manipulated behaviours and voting patterns, which is precisely the danger we're talking about.

You don't seem to be willing or able to actually engage with what's being said here. Every time it's just a kneejerk dismissal or repeat of some already-addressed gripe or distraction, or else a facile attempt to shift the conversation onto something else like bias or Rittenhouse. You're not considering or engaging with anything.

 

[Phoenixmgs]

For the third (?) time: the same abilities can be applied in other contexts. Seriously, this has already been directly addressed at least twice, not to mention how it doesn't make any goddamn sense as a response anyway.

Imagine there was a batch of tainted vegetables sold. Someone bought some, cooked them into a stew, but thankfully their guests noticed the stew looked off and didn't eat it.

Your response here is equivalent to saying, "people noticed it in the stew, and most people don't eat stew anyway, so there's no risk". Ignoring the fact that those tainted vegetables will be in other things as well: the stew just brought it to our attention.



I couldn't give less of a shit about your spurious effort to conflate actual falsehood with standard bias. We're discussing one and not the other right now. Recap:

* balance of actual falsehood is not 90/10 or anywhere near a hypothetical "saturation point".
* Adding more misinfo increases the likelihood that any reader will consume that instead of non-misinfo.
* standard bias in reporting is completely irrelevant to this.



Indeed: it's just going to encourage them to engage in manipulated behaviours and voting patterns, which is precisely the danger we're talking about.

You don't seem to be willing or able to actually engage with what's being said here. Every time it's just a kneejerk dismissal or repeat of some already-addressed gripe or distraction, or else a facile attempt to shift the conversation onto something else like bias or Rittenhouse. You're not considering or engaging with anything.

I'm aware of that... Something given to me that I do need to look at, I know more about it than a legal briefing. Hence, you need greater deception to fool someone on something they are expected to know about. It's like saying AI can fool a layperson by making up a patient's chart when a doctor wouldn't be fooled by it; the layperson would never be asked to look at a patient chart.

Misinformation is incorrect or misleading information. Misinformation can exist without specific malicious intent; disinformation is distinct in that it is deliberately deceptive and propagated. Misinformation can include inaccurate, incomplete, misleading, or false information as well as selective or half-truths.

Flat earthers already have fucking conventions without AI. What is this AI going to encourage them to do beyond that? The main issue with why something like flat earthers or whatever is just the nature of the internet where they find people that agree with them. Without that, they would just be called stupid or idiots by everyone else and you can have like a convention with just one person. Anyone in the US voting for either of the 2 main parties is already an example of voter manipulation.