Please create an account to participate in the Slashdot moderation system

 


Forgot your password?
Close
typodupeerror
×
AI The Internet

YouTuber Trains AI On 4Chan's Most Hateful Board (engadget.com) 94

Posted by BeauHD from the you'll-never-believe-what-they-found dept.
An anonymous reader quotes a report from Engadget: As Motherboard and The Verge note, YouTuber Yannic Kilcher trained an AI language model using three years of content from 4chan's Politically Incorrect (/pol/) board, a place infamous for its racism and other forms of bigotry. After implementing the model in ten bots, Kilcher set the AI loose on the board -- and it unsurprisingly created a wave of hate. In the space of 24 hours, the bots wrote 15,000 posts that frequently included or interacted with racist content. They represented more than 10 percent of posts on /pol/ that day, Kilcher claimed.

Nicknamed GPT-4chan (after OpenAI's GPT-3), the model learned to not only pick up the words used in /pol/ posts, but an overall tone that Kilcher said blended "offensiveness, nihilism, trolling and deep distrust." The video creator took care to dodge 4chan's defenses against proxies and VPNs, and even used a VPN to make it look like the bot posts originated from the Seychelles. The AI made a few mistakes, such as blank posts, but was convincing enough that it took roughly two days for many users to realize something was amiss. Many forum members only noticed one of the bots, according to Kilcher, and the model created enough wariness that people accused each other of being bots days after Kilcher deactivated them. "It's a reminder that trained AI is only as good as its source material," concludes the report.
This discussion has been archived. No new comments can be posted.

YouTuber Trains AI On 4Chan's Most Hateful Board More

YouTuber Trains AI On 4Chan's Most Hateful Board

Comments Filter:

  • To be completely fair (Score:5)

    by twocows ( 1216842 ) on Tuesday September 13, 2022 @09:02AM (#62877121)
    I've been accused of being a bot on /v/ plenty of times... damn meatbags...

  • Always has been.

    • Re:They're all bots in there (Score:5, Insightful)

      by hsthompson69 ( 1674722 ) on Tuesday September 13, 2022 @09:15AM (#62877159)

      And that's the key - hate is an emergent property of competition in the #AttentionEconomy.

      Meatbags pay attention to things they fear, and things they hate, more than other things. Simple pursuit of profit in an economy that rewards attention, is going to maximize fear and hate.

      Evolutionary pressures create the hate and fear - and an AI that trains in that environment will be subject to those same pressures.

      • Meatbags ...

        Bender? ;)

        And that's the key - hate is an emergent property of competition in the #AttentionEconomy.

        It's interesting, but could you provide some sources as it doesn't seem like "hate" would come from this. It seems to me that "hate" is more the result of failures and/or struggles in personal life combined with some sort of narcissism, when a person projects their failures or cause of such onto some group/objects they can "hate" detaching their personal problems from their own actions this way, i.e. as "I do not have to feel bad about myself, because all my problems is the fault of ..."

        • More likely HK-47 than Bender IMHO

        • Our attention has developed under selective pressures. In the world we've evolved in, you needed to pay close attention to the things you fear (avoiding the predator that could gank you), and you needed to pay close attention to the things you hate (the flip side of fearing the predator, learning how to gank it back).

          If you don't, you're either eaten by the predator, or you're unable to destroy threats to you.

          The way you put it works too - their hateful attention avoids ego destruction - blame others for y

          • I see, as "hate" begin the source of aggression against anything standing on the way to survival and the "hate" itself as a long term memory pattern derived from former fear, injury or unavoidable obstacle. And even though "aggression" in general would itself be sufficient as a fight response, "hate" with it's implication of "revenge" let our predecessors be proactive to a danger.

            I am pretty sure I simplified things, especially considering development of prefrontal-cortex and later speech, so it's interesti

            • I like that framing.

              Maybe there's a chart here:

              hate - aggression against something between you and survival that has, on balance, a chance for success and a chance for failure

              fear - repulsion to something between you and survival that has, on balance, a greater chance of failure than of success

              greed - attraction to something between you and survival that has, on balance, a greater chance of success than failure

              Given that we tend to prioritize negative emotion higher, hate and fear probably get more attentio

      • Meatbags pay attention to things they fear, and things they hate.

        The "advanced" AI has managed to recreate the absolute lowest form of human intelligence.

      • Re: (Score:2)

        by shanen ( 462549 )

        Mod parent more insightful. And sad.

    • Always has been.

      Yep. I'm fairly sure humans can't read 150,000 posts per day.

  • and nothing was lost (Score:4, Interesting)

    by Kisai ( 213879 ) on Tuesday September 13, 2022 @09:08AM (#62877139)

    And nothing of value was lost.

    In all seriousness, GIGO (Garbage IN - Garbage OUT) applies doubly so for AI. AI can NOT, and will NEVER learn anything it has not trained on. So in the case of 4chan, it's likely nobody would have noticed for weeks if it wasn't for blindspots in the training data, and the inability of the bot to respond to real-world events.

    For things like GPT, these models have to be updated daily if you want them to be able to fool humans. Because that will be undoing of the bot's ability to interact with humans. That's going to be come the new test if someone is a bot or not. "What happened today in the world?"

    • What makes you think, in the case of 4chan, it's likely nobody would have noticed for weeks if it wasn't for blindspots in the training data, and the inability of the bot to respond to real-world events?
    • I dont understand why these models have to be updated daily if you want them to be able to fool humans.

      • Because even people whose views don't change come up with new excuses for them

        • No they don't. And you're a racist bigot homophobe TURF. (did I do it right?)

          • >did I do it right?

            No. First, it's terfs. Trans exclusionary radical feminists. Second, generally the people complaining about these things separate their hatred for terfs (if they hate terfs) so as not to mix the messages. So you'll find a complaint about racism and misogyny and then elsewhere a complaint about terfs and homophobia. If they complain about misogyny in the same breath as turfs they come uncomfortably close to feminism's inherent hatred of men, as that is the reason terfs are te. Terfs ha
          • it's TERF (Trans-Exclusionary Radical Feminist)
            • I think turf is just another slur woke culture created to vilify people they disagree with.Feminist aren't against trans people. They are against special rights being given to trans people at the expense of women's rights. Even then, this is all about trans activists burning bridges to their own allies over nothing. It's just like the words CIS or binary. Normal people don't use these words for a reason.

              • Feminist aren't against trans people

                That's not necessarily true - Situation with Leah Thomas is a perfect example of there being an not insignificant number of people people for exclusivity for women, but wanting to exclude a trans person from competing.

              • Some second-wave feminists wrote entire books about how "the problem of transsexualism would best be served by morally mandating it out of existence".

        • Well said.
      • You don't need to update the model daily. If you have a fresh corpus of text or a search engine you can use that to pad the request with up to date information. The model can reference the latest data instead of talking from memory.
    • Please explain why That's going to be come the new test if someone is a bot or not. "What happened today in the world?"

    • AI can NOT, and will NEVER learn anything it has not trained on.

      If it can't then it isn't real AI. The acronym literally means "Artificial Intelligence". There is plenty of software derived from AI research, such as expert systems which are not now considered to be "intelligent". There might be a way of deriving true intelligence from deep learning but basic simple deep learning systems can also be simply considered not "intelligent".

      • Hey, humans are doing the same - if you had no exposure to Chinese, you can't speak Chinese. If you had no exposure to math, you don't know math. The model has a static dataset, we have bodies and are part of the world - our dataset is dynamic. But we all need to learn from somewhere. True innovation is incremental and mostly accidental, the result of intuition based search and rigorous evaluation. Models are good at search and intuition, bad at evaluation.

      • "Artificial Intelligence" is a very broad term that includes many computer science techniques for making computers imitate intelligent behavior. So, to be clear, it does not need to actually BE intelligent in order to qualify as AI, it just needs to imitate intelligence in some way in order to qualify. That's why something like a chess program, which can do literally nothing beyond play chess well, is properly called a "chess-playing AI."

        Many people seem to think that AI refers only to the stricte

        • "Artificial Intelligence" is a very broad term that includes many computer science techniques for making computers imitate intelligent behavior.

          This is just not right although it is something which has been pushed by the marketing departments of various companies. There are different terms for each possible thing which isn't actually intelligent. Most of what you are talking about would come under "machine learning" for example, which does not have to include "intelligence". AI was very clearly originally defined as the search to build an artificial mind which was "intelligent" with one problem being that we don't yet have a fully clear definition

          • The dictionary [merriam-webster.com] disagrees with you. Quote:

            the capability of a machine to imitate intelligent human behavior

            I trust I don't need to link the definition of "imitate" for you to agree that something that "imitates" intelligence does so because it is not intelligent.

            In English, word meaning is determined by popular use. So, the phrase "artificial intelligence" is popularly used to refer to anything computers do that "acts as if" intelligent even when not, and therefore that's what it means. This is a hard th

    • You probably didn't watch the author's video. The model wasn't so bad in all respects. This is from the transcript (https://www.youtube.com/watch?v=efPrtcLdcdM)

      > I evaluated the model on a huge set of benchmarks and it turns out this horrible terrible model is more truthful yes more truthful than any other gpt out there

      In other words training on 4chan created the most anti-bullshit model.

    • Predictions of the form "computers will never be able to..." are quite risky, since they are always founded on an obvious ignorance about what will be achieved in the future.

      In the specific case of the term "AI," such a prediction might have a leg to stand-on due to the social phenomenon of a defacto definition of AI along the lines of "AI is whatever humans can do that computers can't do right now." Such a definition is a moving target that can re-update itself after any step forward is taken in the AI fi

  • For the lulz.

  • Careful... (Score:4, Funny)

    by VeryFluffyBunny ( 5037285 ) on Tuesday September 13, 2022 @09:36AM (#62877217)
    The Republican GQP Party might elect it as their leader!

    • Re: (Score:1)

      by Anonymous Coward
      I was thinking that the bot should be set loose on The Truth and other lovely bastions of hate. Keep doing it every couple of months, and hopefully keep them occupied over the "bots in the comments"

    • And it would be the most human, humane and likable leader they had in years.

    • Re: (Score:2)

      by shanen ( 462549 )

      But was the AI born in the US of A? Let's see that birth certificate!

      As jokes go, I wish it were merely funny rather than aspirational.

  • It never changes.

  • ...but was convincing enough that it took roughly two days for many users to realize something was amiss.

    Maybe the hate did not seem authentic?
    "To continue, you must authentic-hate to the server."

  • how much money was the bot able to raise ?

    1. train chatbot on racist/homophobic/misogynistic/fake-nostalgic spew

    2. Open a PAC with the bot as "spokesperson"

    3. offer membership to the bot's "gold patriot wall-of-fame supporter program"

    4. profit !

    • Please, this is 4chan, not truthsocial.

      • Please, this is 4chan, not truthsocial.

        Sounds like a failure of eemajination, the bot should have tried fundraising for a worthy cause. Base the thing in the Seychelles (or Nigeria) and . . . PROFIT! Consider it an idiot-tax.

  • Seriously, how long has GIGO been in the lexicon?

    • Holy shit, was I late to the GIGO party. I should at least have perused the first handful of responses.

      Ignore. Redundant.

    • Re: (Score:2)

      by HiThere ( 15173 )

      Well, at least since the early 1960's. I can't remember exactly when i first heard it, but it wasn't new then. Wikipedia dates it back to 1957.

  • Spewing hate requires no actual intelligence. Therefore an artificial intelligence that does so is no more intelligent than the actual posters there! Making an artificial intelligence with a complete lack of intelligence is, well, something that should be relatively easy (although the creator might need slightly more intelligence than the average racist, but again that's a pretty low bar).

    Making an AI that can recognize situations and apply the right amount of compassion and empathy is an actual challen

  • A Joke, Right? (Score:1)

    by Anonymous Coward
    I mean, this guy evidently had never lurked on 4chan at all, ever, to insinuate that his stupid post bot caused people to accuse each other of being bots on THAT forum. Or to unironically believe that the users accusing each other of being bots aren't also correct. The height of naivety. Why is this even news? "I did a thing that happens a lot, quick, every dying, readerless media outlet write stupid articles about it!"

    The most telling part is where it seems obvious to this nobody that users were discerni

  • The nature of intelligence (Score:3)

    by byromaniac ( 8103402 ) on Tuesday September 13, 2022 @10:44AM (#62877419)

    "It's a reminder that trained AI is only as good as its source material," concludes the report.

    I find it fascinating that every story about the shortcomings of AI applies just as well to human intelligence. The gist of this story is that if you train an AI on 4chan it will behave comparably to the humans trained on 4chan. Careful what you put in your heads, folks!

  • Thanks Yannic! (Score:3)

    by oumuamua ( 6173784 ) on Tuesday September 13, 2022 @10:45AM (#62877425)
    Not just a 'YouTuber', Left out of the summary is that Yannic posts videos explaining and summarizing AI papers. A very useful service for those of us lacking time or deep math background. Also a shout out to:
    Ms Coffee Bean https://www.youtube.com/c/AICo... [youtube.com]
    The AI Epiphany https://www.youtube.com/c/TheA... [youtube.com]
    Outlier (just starting out but slick videos) https://www.youtube.com/channe... [youtube.com]
    Siraj Raval (now with Udacity)
    You guys have saved me hundreds of hours!
    Let me know who I've missed.

  • it took roughly two days for many users to realize something was amiss

    Bots have been on Twitter for years and no one noticed. And they are still in denial.

  • You could in principle use the model to detect speech like 4chan's /pol/ board, and flag it.

  • Let's be honest, there are no humans in /pol/

  • I don't know what "kek" means and at this point I'm too afraid to ask.

    • I'm still traumatized from finding out what "uwu" means.

    • Re:I don't understand 4chan. (Score:5, Informative)

      by twocows ( 1216842 ) on Tuesday September 13, 2022 @11:40AM (#62877611)
      "kek" has dual ancestry -- it comes from the continued degradation of "lol" (lol -> lel -> kek, jej also had some very minor usage for a while... a step further and we'd have ended up on heh, funnily enough), and also from World of Warcraft, where common-speaking races (Alliance ones) would see orcish-speaking races (Horde ones) saying "lol" as "kek" due to the language filter (the game was designed so that you can't understand the other faction's speech because a pseudo-random filter is applied to it). It's really just a different way of saying "lol." There's always been a bit of a pattern on 4chan and related sites/subcultures of deliberately mangling language; it's fun to see things you created catch on which is why people do it, and it creates a point of in-group out-group differentiation, which is why it stays popular.

  • What could go wrong?

  • Has anyone pointed out that this was an utterly horrific thing to do? This "prankster" should be held to account for each and every terrible thing their bot posted. They knew exactly what they were doing, and caused harm doing it.

    • If this means bolstering the hate that the board losers are pumping themselves up on, I agree.

      I hope this won't lead to one of them flaking out and gunning down a supermarket full of shoppers.

  • They have been been a hateful lot. Post your hate comments below.
  • So does the Turing test specify a general audience, or can you stick to a specific selection of people for the test...

  • I'm afraid that the AI will become as stupid as the (l)users on that board.

  • Something I noticed after spending a long time talking to climate science dismissives [medium.com] is that - while they are individually unique, each having their own personality - the way they reason is highly linguistic in nature, and therefore well suited to simulation by a language model.

    Consider this: my car is rusty. If I try to bend a piece of rusty metal, it will break off. If I scrape the rust, little flecks of it will fall off. A language model will "know" all of this because it has seem humans talking abou

Slashdot Top Deals

It's a naive, domestic operating system without any breeding, but I think you'll be amused by its presumption.

Close