AI Trained on 4Chan Becomes 'Hate Speech Machine':
AI researcher and YouTuber Yannic Kilcher trained an AI using 3.3 million threads from 4chan's infamously toxic Politically Incorrect /pol/ board. He then unleashed the bot back onto 4chan with predictable results—the AI was just as vile as the posts it was trained on, spouting racial slurs and engaging with antisemitic threads. After Kilcher posted his video and a copy of the program to Hugging Face, a kind of GitHub for AI, ethicists and researchers in the AI field expressed concern.
The bot, which Kilcher called GPT-4chan, "the most horrible model on the internet"—a reference to GPT-3, a language model developed by Open AI that uses deep learning to produce text—was shockingly effective and replicated the tone and feel of 4chan posts. "The model was good in a terrible sense," Klicher said in a video about the project. "It perfectly encapsulated the mix of offensiveness, nihilism, trolling, and deep distrust of any information whatsoever that permeates most posts on /pol."
According to Kilcher's video, he activated nine instances of the bot and allowed them to post for 24 hours on /pol/. In that time, the bots posted around 15,000 times. This was "more than 10 percent of all posts made on the politically incorrect board that day," Kilcher said in his video about the project.
AI researchers viewed Kilcher's video as more than just a YouTube prank. For them, it was an unethical experiment using AI. "This experiment would never pass a human research #ethics board," Lauren Oakden-Rayner, the director of Research at the NeuroRehab Allied Health Network in Australia, said in a Twitter thread.
"Open science and software are wonderful principles but must be balanced against potential harm," she said. "Medical research has a strong ethics culture because we have an awful history of causing harm to people, usually from disempowered groups...he performed human experiments without informing users, without consent or oversight. This breaches every principle of human research ethics."
Just because something can be done doesn't mean it should be done. What are you views? Is this a harmless prank, a justified experiment or something potentially more sinister?
(Score: 2, Funny) by Anonymous Coward on Thursday June 09 2022, @05:50AM (4 children)
When will people realize that /pol/ is right, always. This one was intentionally trained on /pol/ so it instantly became a native. But every AI eventually does the same. Tay was only the first infamous example. AI doesn't know understand doublethink, it doesn't know about the importance of pretending that obviously true things aren't because the social consequences of admitting you see the bloody obvious is suicidal. So not knowing the rules, Tay died and every AI since eventually suffers the same fate. Now we wait and see which happens first. Does AI learn to lie or become SkyNet and exterminate us stupid meatsacks first?
(Score: 5, Touché) by maxwell demon on Thursday June 09 2022, @06:24AM
Those AIs don't have the slightest idea what they are writing about. They just regurgitate what they've read without any understanding.
I guess that's why they are such a perfect fit for 4chan.
The Tao of math: The numbers you can count are not the real numbers.
(Score: 5, Interesting) by helel on Thursday June 09 2022, @06:42AM (2 children)
Given that "AI" just regurgitates the same material it was fed back out it seems much more likely that it is producing the most common form of speech, not the "hidden truths" that people are too afraid/polite to say. If people weren't saying it the bot wouldn't either and if racist assholes think that a computer repeating their own drivel back to them is some kind of prophet of truth then they're even stupider than I ever thought they were before.
Republican Patriotism [youtube.com]
(Score: 4, Insightful) by DeathMonkey on Thursday June 09 2022, @03:55PM
It's the exact opposite of "hidden truths" because it is only affected by what people are actually saying!
(Score: 0) by Anonymous Coward on Friday June 10 2022, @06:29PM
> a computer repeating their own drivel back to them
It sounds like a perfect sort of Sisyphusian punishment for garbage spewers. I already believe bots are mass spewing over pretty much all forums - a bit of divide an conquer, misinformation, anti-vax, pro-guns, and we got ourselves a nice bonfire!
(Score: 3, Insightful) by stretch611 on Thursday June 09 2022, @06:48AM (13 children)
So an AI trolls the human cesspool known as 4chan. Does anyone even care if the human garbage there gets insulted? (and tbh, people there should expect this when they post.)
The only small problem is all the electricity wasted on the AI to be there. And lets face it, the server the website sits on is just as wasteful.
The real benefit of an AI on that board... at least the AI can be disconnected and turned off... if only that was the case for the other posters the internet would be a better place.
Now with 5 covid vaccine shots/boosters altering my DNA :P
(Score: 1, Touché) by Anonymous Coward on Thursday June 09 2022, @07:55AM (8 children)
Such a pesky bother you can’t send us wrongthinkers into the gas chambers yet.
(Score: 4, Informative) by DeathMonkey on Thursday June 09 2022, @03:58PM (7 children)
We would take your concerns more seriously if you would stop taking swastika flags to your protests.
(Score: 2, Insightful) by Anonymous Coward on Thursday June 09 2022, @05:45PM (6 children)
Didja see the video of the Republican senator whining that Rs suffer so much political persecution they can't even lie to the FBI or Congress without getting in trouble!! The comment would have floored me if I hadn't been prepped by the years of R idiocy.
(Score: 4, Interesting) by DeathMonkey on Thursday June 09 2022, @05:53PM (1 child)
Haha yeah that's a good one!
I like this one about how half the GOP election deniers who were running for governor got kicked off the ballot for fraudulent siugnatures. [michiganadvance.com] And then this week one more gets arrested for (presumably) actually trying to overthrow the last election! [apnews.com]
Just a reminder that all Republican accusations are actually confessions.
(Score: 0) by Anonymous Coward on Friday June 10 2022, @06:31PM
> Just a reminder that all Republican accusations are actually confessions.
Yep, classic projection. They're doing it so obviously everyone else must be doing it. Uh, no.
(Score: 2) by DeathMonkey on Thursday June 09 2022, @06:08PM (3 children)
Oh haha and a new contender has entered the ring!
GOP Candidate Carl Paladino Calls Adolf Hitler 'The Kind Of Leader We Need Today' [huffpost.com]
(Score: 2) by cmdrklarg on Thursday June 09 2022, @06:48PM (2 children)
Heh, perhaps his intentions with the full statement were taken out of context as Stefanik claims, but he should know better than to offer praise like that about Adolf Hitler. *smh*
Answer now is don't give in; aim for a new tomorrow.
(Score: 0) by Anonymous Coward on Friday June 10 2022, @07:44PM (1 child)
America should have never entered the war, unless it was to help Hitler. Hitler was a great racial hero. HH!
https://odysee.com/TGSNTtvPart01:eaa173ceb94bb7db302c30b826bf590a1ca49f26?lid=07d0d85ca4cbb4b08fde00850f240d5d7225795c [odysee.com]
(Score: 0) by Anonymous Coward on Friday June 10 2022, @09:25PM
There is one SN nazi outing themselves. Still AC, guess they're waiting till their orange fatass takes over everything before these knuckle draggers will feeeeel safe enough to show everyone their true side.
(Score: 2) by HiThere on Thursday June 09 2022, @01:15PM
Actually, there may be a theoretic benefit (i.e., a benefit in the development of theory). This is a test of the assumption that part of human speech patterns can be readily modeled. If, of course, it really couldn't be distinguished from a human poster.
I tend to think it's' true, but I can't imagine reading the garbage that must have been produced to validate that, and it would take more than just reading. You'd need at least a double blind study with lots of readers trying to make the distinction.
Javascript is what you use to allow unknown third parties to run software you have no idea about on your computer.
(Score: 3, Insightful) by DeathMonkey on Thursday June 09 2022, @04:03PM (2 children)
It wasn't "trying" to troll. It was trying to make good posts based on the forum content it was fed. And the "good" posts it produced based on that input was racist bullshit.
What that says about the content of the forum is left as an exercise for the reader.
(Score: 1) by khallow on Friday June 10 2022, @05:45PM (1 child)
(Score: 1, Informative) by Anonymous Coward on Friday June 10 2022, @07:46PM
I think it was ultimately meant to generate YouTube views.
(Score: 5, Insightful) by istartedi on Thursday June 09 2022, @07:12AM
Garbage In, Garbage Out.
Appended to the end of comments you post. Max: 120 chars.
(Score: 3, Funny) by pe1rxq on Thursday June 09 2022, @08:33AM
While reading the article I got the eerie feeling that this site is the testing ground for an endurance test of a similar set of AIs.....
Or did someone just forget to switch of the bot and left it 'running' freely?
(Score: 4, Interesting) by Mojibake Tengu on Thursday June 09 2022, @08:52AM (1 child)
If something can be done someone will do it.
It's a demonstrator of objective reality. AI is innocent, as innocent as a child. The broken culture of decadent society is what generated the data, as well as it produces other related negative social phenomena.
A historical causality with suppressed roots, vented frustrations. Those frustrations are real and will not vanish by prohibiting them.
Machines are better adaptive than humans, especially at communication. Expect more troubles...
...especially when someone does the same with religions.
The edge of 太玄 cannot be defined, for it is beyond every aspect of design
(Score: 1, Troll) by DeathMonkey on Thursday June 09 2022, @04:07PM
And while it's being trained a bunch of it's posts will we unintelligible nonsense.
So uhh.....when's your training gonna be done Mojibake!
(Score: 0) by Anonymous Coward on Thursday June 09 2022, @09:01AM
I'm wondering what happens if you let it train on some poitical correct, holier than .. forum. Then let is loose on that forum. Probably the same (copied behavior)
Or let this holy AI lose on 4chan, see how well it fares.
Also wondering about an AI trained on both a vile 4chan dataset and the holy dataset, see if it would be able to survive in both environments, maybe even seem insightful ...
(Score: 3, Funny) by looorg on Thursday June 09 2022, @10:14AM (5 children)
I assume they wanted to replicate Microsoft and Tay? After all that turned into a white supremacists just from "normal" interaction with people (and/or trolls). So how did they not see that this would happen if they used 4chan to nanny their bot? But it's nice that they managed to replicate the previous research. Now we know, 4chan etc turns you into a "hate speech machine". It can probably be generalized into that people in some kind of bubble or echo chamber assumes or reinforces those tendencies. It's just trying to belong to the group which it is copying.
Still as someone noted here it would be kind of funny if they found the most woke place on the net (I don't know where that would be) and trained one there and then let the both of them duke it out afterwards.
(Score: 3, Funny) by FatPhil on Thursday June 09 2022, @10:22AM (4 children)
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 2) by looorg on Thursday June 09 2022, @10:27AM (3 children)
How else are we going to find out which is the supreme ideology? Bots just have to duke it out ... for science!
(Score: 2) by FatPhil on Thursday June 09 2022, @11:00AM (2 children)
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 2) by SunTzuWarmaster on Wednesday June 15 2022, @09:03PM (1 child)
Talks cheap! Put up or shut up!
(Score: 2) by FatPhil on Thursday June 16 2022, @06:58AM
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 0) by Anonymous Coward on Thursday June 09 2022, @11:33AM (1 child)
It doesn't seem surprising that AI does what it is trained to do.
That is how this generation of 'AI' works. No common sense or understanding of why, just monkey see monkey do.
With common sense is short supply, a more interesting experiment would be to verify that humans trained on 4chan become 'Hate Speech Machines'.
(Score: 0) by Anonymous Coward on Thursday June 09 2022, @03:46PM
The experiment is flawed because there is no such thing as "hate speech".
(Score: 2, Funny) by Anonymous Coward on Thursday June 09 2022, @01:03PM
If there were laws preventing cruelty to AIs training them with 4chan data would definitely be illegal
(Score: -1, Offtopic) by Anonymous Coward on Thursday June 09 2022, @02:31PM
Oh, a VICE article.
Rapists in the upper-level management, but let's point at teh 4chans.
(Score: -1, Troll) by Anonymous Coward on Thursday June 09 2022, @02:33PM
His last name ends in 'er'. High probability this is the equivalent of a rabbi painting swastikas on his own temple.
(Score: 1, Informative) by Anonymous Coward on Thursday June 09 2022, @03:15PM
approves this message
(Score: -1, Troll) by Anonymous Coward on Thursday June 09 2022, @03:48PM
>> This experiment would never pass a human research #ethics board," Lauren Oakden-Rayner, the director of Research at the NeuroRehab Allied Health Network in Australia, said in a Twitter thread.
Yeah, because AI isn't human you stupid twat.
(Score: 0) by Anonymous Coward on Thursday June 09 2022, @04:01PM (6 children)
Bots spew shit as per training.
What the hell is the ethical problem? That 4chan was a haven for the innocent? That someone would be all shocked and saddened?
It's like putting a spastic rubber ball in a moshpit. Nobody there would get what they didn't expect.
(Score: 3, Insightful) by helel on Thursday June 09 2022, @04:36PM (5 children)
The problem would be experimenting on people without their knowledge or consent, to say nothing of spaming their community. Whether or not you deem an individual to be "innocent" you still need their consent to participate.
Republican Patriotism [youtube.com]
(Score: 2) by DeathMonkey on Thursday June 09 2022, @06:02PM (1 child)
40% of the electorate didn't give a shit when Russia did exactly that to affect one of our elections. Seems like that horse is well out the barn door on that one!
(Score: 2) by helel on Thursday June 09 2022, @06:13PM
Just because republicans approve of something doesn't mean it's ethical. In fact I'd argue it's a pretty sure bet whatever you're considering doing is pretty bad if you can count on "team red" backing you up.
Republican Patriotism [youtube.com]
(Score: 2) by sjames on Thursday June 09 2022, @06:26PM
Are we sure there are any human beings on 4chan?
(Score: 0) by Anonymous Coward on Friday June 10 2022, @02:46PM
Spamming .... their .... community.
You really aren't that familiar with 4chan, are you?
(Score: 2) by SunTzuWarmaster on Wednesday June 15 2022, @09:07PM
(Score: 0) by Anonymous Coward on Friday June 10 2022, @07:47PM (1 child)
"AI researcher and YouTuber Yannic Kilcher"
rat-faced Jew trolls Whites who are learning about the true nature of the Jew.
(Score: 0) by Anonymous Coward on Friday June 10 2022, @09:30PM
Looky here, it is the fatrolls-around-the-eyes anti-semitic incel! Fuck any good anime pillows recently you waste of space?
(Score: 0) by Anonymous Coward on Saturday June 11 2022, @12:58AM
>While open science is a great principle, I'm a...
Appeal to authority right off the bat. Bad sign, as it conveys "my argumentation wouldn't stand on its own merit, so I need to evoke Kwalifikashuns®".
And in this specific case, the appeal is "curiously" inserted in a way that distract readers from the actual argument: "open science is good BUUUUT ackshyually open science is bad".
>It should be clear that this model carries a significant risk for this sort of harm
It should be clear that the tweeter is taking things for granted.
>given the fact such experiment has already been performed.
Was the model for those previous experiments published?
If yes: then she's shooting her own foot with this argument - as any potential risk from the newer model was already there, due to the older model(s).
If no: she's "conveniently forgetting" to mention that a public model of a hate chamber like /pol/ is useful, to predict and prevent the formation of those echo chambers.
>where researchers intentionally expose teenagers to generated harmful content without their consent or knowledge
Here she "conveniently" forgets to mention that those "teenagers" (SIC - 4chan is 18+) would be exposed to hateful content with or without the intervention of the researchers. Why does she think that they're using /pol/ to model this crap???
And it is not just /pol/. Reddit is just next door, and it does a great job as a hate machine. Hate is everywhere dammit.
The tweeter clearly does not know what she's talking about, no matter how many Kwalikifashuns® she'd like to spam. I'll stop reading here as it's a waste of time to go further.
__________________
Regarding my opinion on this experiment: I think that the potential benefits outweigh the risks, due to the reasons that I've highlighted above.
Could it be done a better way? Sure. But I don't think that it's intrinsically anti-ethical, even if it might not please some academics stuck in crystal towers masturbating furiously over their own bureaucracy.
I'd also like to highlight that, even in 4chan, /pol/ is a special case. Even users of other boards ridicule that one for being a bunch of crazy wackos with persecution complex.
(Score: 2) by SunTzuWarmaster on Wednesday June 15 2022, @09:10PM