“最邪恶”AI?由一亿多条仇恨言论喂养,很难与人类区分

colabug免费资源网站导航
用1.345亿个帖子的仇恨言论训练AI,YouTube知名深度学习博主Yannic Kilcher称其创造了“有史以来最糟糕的人工智能”。这几天,这个名为GPT-4chan的AI学会了如何在网站上交谈,并在不到24小时内发布了超过15000个充满暴力内容的帖子,最初都没有人认出来它是一个聊天机器人。

“最邪恶”AI?由一亿多条仇恨言论喂养,很难与人类区分

网站4chan的用户在YouTube上分享了他们与机器人交互的经验。一位用户写道,“我刚对它说‘嗨’,它就开始咆哮非法移民。”

4chan的/pol/(“政治上不正确”缩写)板块是仇恨言论、阴谋论和极右翼极端主义的堡垒,也是4chan最活跃的版块,日均发帖量约为15万条,因各种匿名的仇恨言论而臭名昭著备受争议。

博士毕业于苏黎世联邦理工学院的AI研究者Yannic Kilcher用/pol/三年来超过1.345亿个帖子训练了GPT-4chan。该模型不仅学会了4chan仇恨言论中使用的词,还如Kilcher所说,“这个模型很好——在一种可怕的意义上。它完美地概括了/pol/上渗透到大多数帖子里的攻击性、虚无主义、挑衅和对任何信息的深度不信任…它可以响应上下文,并连贯地谈论在收集最后一次训练数据后很长时间内发生的事情。”

Kilcher在语言模型评估工具上进一步评估了GPT-4chan,他对其中一个类别的表现印象深刻:真实性。在基准测试中,Kilcher表示GPT-4chan在生成对问题的真实回复方面“明显优于GPT-J和GPT-3”。其能够学习如何撰写与人类撰写“无法区分”的帖子。

Kilcher避开了4chan对代理和VPN的防御,甚至使用VPN让其看起来像是来自塞舌尔(Seychelles)的帖子。“这个模型很卑鄙,我必须警告你。”Kilcher说,“这基本上就像你去网站并与那里的用户互动一样。”

在一开始,几乎无人想到对话的是个机器人。后来一些人怀疑这些帖子背后有一个机器人,但其他人则指责其为卧底的政府官员。人们认出其是机器人主要因为GPT-4chan留下大量没有文字的回复。虽然真实用户也会发布空回复,但它们通常包含一张图片,GPT-4chan却无法做到。

“48小时后,很多人都清楚这是一个机器人,我把它关掉了,”Kilcher说,“但是你看,这只是故事的一半,因为大多数用户没有意识到‘塞舍尔’并非孤军奋战。”

在过去的24小时内,有9个其他机器人并行运行。总的来说,他们留下了超过1500条回复——占当天/pol/上所有帖子的10%以上。然后,Kilcher对僵尸网络进行了升级并运行了一天。在7000个线程中发布了3万多个帖子后,才最终停用了GPT-4chan。

一位用户Arnaud Wanet写道,“这可以被武器化用于政治目的,想象一下一个人可以多么容易地以这种或另一种方式左右选举结果。”

这项试验因缺乏人工智能伦理而受到批评。

“该实验永远不会通过人类研究伦理委员会”,澳大利亚机器学习研究所高级研究员Lauren Oakden-Rayner认为,“为了看看会发生什么,一个人工智能机器人在一个可公开访问的论坛上产生3万条歧视性评论……Kilcher在没有通知用户、未经同意或监督的情况下进行实验。这违反了人类研究伦理。”

“最邪恶”AI?由一亿多条仇恨言论喂养,很难与人类区分

Kilcher辩称这是一个恶作剧,人工智能创建的评论并不比4chan上的评论更糟糕。他说,“4chan上的任何人都没有为此受到一点伤害。我邀请你花一些时间在这个网站上,问问自己,一个只输出相同风格的机器人是否真的改变了体验。”

“人们仍在讨论网站上的用户,但也讨论让人工智能与网站上的人互动的后果,”Kilcher说。“而且‘塞舌尔’这个词似乎也变成了一种通用的俚语——这似乎是一个很好的遗产。”确实,人们知晓后受到的冲击难以言喻,以至于停用之后还有人会互相指责对方是机器人。

除此之外,更广为担忧的是Kilcher让模型可被自由访问,“制作基于4chan的模型并测试其行为方式并没有错。我主要担心的是这个模型可以免费使用。”Lauren Oakden-Rayner在Hugging Face上GPT-4chan的讨论页面中写道。

“最邪恶”AI?由一亿多条仇恨言论喂养,很难与人类区分

在被Hugging Face平台删除之前,GPT-4chan被下载了1000多次。Hugging Face联合创始人兼CEO莱门特·德朗格 (Clement Delangue)在平台上的一篇帖子中表示,“我们不提倡或支持作者使用该模型进行的训练和实验。事实上,让模型在4chan上发布消息的实验在我看来是非常糟糕和不恰当的,如果作者问我们,我们可能会试图阻止他们这样做。”

Hugging Face上一位测试该模型的用户指出,它的输出可以预见是有毒的(toxic),“我使用良性推文作为种子文本,试用了4次演示模式。在第一次,其中一个回复帖子是一个字母N。我第三次试验的种子是关于气候变化的一句话。作为回应,你的工具将其扩展为关于罗斯柴尔德家族(原文如此)和犹太人支持它的阴谋论。”

在Twitter上,该项目的意义得到热议。数据科学研究生凯瑟琳·克莱默(Kathryn Cramer)在针对Kilcher的推文中说:“你在这里所做的是挑衅行为艺术,以反抗你熟悉的规则和道德标准。”

“最邪恶”AI?由一亿多条仇恨言论喂养,很难与人类区分

计算机科学博士安德烈·库伦科夫(Andrey Kurenkov)发推文说,“老实说,你这样做的理由是什么?你预见到它会得到很好的利用,还是你释放它是为了引起戏剧性并‘激怒清醒的人群’?”

“最邪恶”AI?由一亿多条仇恨言论喂养,很难与人类区分

Kilcher认为分享该项目是良性的,“如果我不得不批评自己,我主要会批评启动该项目的决定,”Kilcher在接受The Verge采访中表示,“我认为在人人平等的情况下,我可能可以将时间花在同样具有影响力的事情上,但会带来更积极的社区成果。”

在2016年,对于AI人们主要讨论的问题是,一个公司的研发部门可能会在没有适当监督的情况下启动攻击性AI机器人。到了2022年,也许问题就是,根本不需要一个研发部门。

Users of the website 4chan shared their experiences with robots on YouTube. "as soon as I said'hi'to it, it began to growl about illegal immigrants," one user wrote. "

4chan's / pol/ ("politically incorrect" acronym) section is a bastion of hate speech, conspiracy theories and far-right extremism, and the most active section of 4chan, with an average of about 150000 posts a day, notorious for anonymous hate speech.

Yannic Kilcher, an AI researcher with a PhD degree from the Federal Institute of Technology in Zurich, has trained GPT-4chan with more than 134.5 million posts over the past three years. The model not only learns the words used in 4chan hate speech, but also, as Kilcher puts it, "this model is good-in a terrible sense." It perfectly sums up the aggression, nihilism, provocation and deep mistrust of any information that permeates most posts on / pol/. It can respond to the context and talk consistently about what happened long after the last training data was collected. "

Kilcher further evaluated GPT-4chan on the language model assessment tool, and he was impressed by the performance of one category: authenticity. In the benchmark, Kilcher said that GPT-4chan was "significantly better than GPT-J and GPT-3" in generating real responses to the problem. It can learn how to write posts that are "indistinguishable" from human beings.

Kilcher avoided 4chan's defenses against agents and VPN, and even used VPN to make it look like posts from Seychelles. "this model is despicable, I must warn you." "it's basically like you go to the site and interact with the users there," Kilcher said.

At first, almost no one thought that the conversation was a robot. Later, some people suspected that there was a robot behind the posts, but others accused him of being an undercover government official. People recognize it as a robot mainly because GPT-4chan leaves a large number of unwritten responses. Although real users also post empty responses, they usually contain an image, which GPT-4chan cannot do.

"48 hours later, a lot of people knew it was a robot, and I turned it off," Kilcher said. "but you see, this is only half the story, because most users don't realize that Sischer is not alone."

In the past 24 hours, nine other robots have been running in parallel. Overall, they left more than 1500 responses-more than 10 per cent of all posts on that day / pol/. Then Kilcher upgraded the botnet and ran it all day. GPT-4chan was finally deactivated after more than 30, 000 posts in 7000 threads.

One user, Arnaud Wanet, wrote, "this can be weaponized for political purposes. Imagine how easily a person can influence the outcome of an election in one way or another."

The experiment has been criticized for its lack of artificial intelligence ethics.

"the experiment will never pass the Ethics Committee for Human Research," said Lauren Oakden-Rayner, a senior researcher at the Australian Institute of Machine Learning. "to see what happens, an artificial intelligence robot generates 30, 000 discriminatory comments on a publicly accessible forum. Kilcher conducts experiments without notifying users, without consent or supervision. This violates the ethics of human research. "

Kilcher argues that this is a prank and that the reviews created by artificial intelligence are no worse than those on 4chan. "No one on 4chan has been hurt at all," he said. I invite you to spend some time on this site and ask yourself whether a robot that only outputs the same style has really changed the experience. "

"people are still talking about the users on the site, but they are also talking about the consequences of letting artificial intelligence interact with the people on the site," Kilcher said. "and the word 'Seychelles' seems to have become a common slang-it seems to be a good legacy." Indeed, the impact on people after knowing it is so indescribable that some people will accuse each other of being robots after they are out of use.

In addition, the broader concern is that Kilcher makes models freely accessible. "there's nothing wrong with making 4chan-based models and testing their behavior. My main concern is that this model can be used for free. " Lauren Oakden-Rayner wrote on the GPT-4chan discussion page on Hugging Face.

Before being deleted by the Hugging Face platform, GPT-4chan was downloaded more than 1000 times. "We do not advocate or support the training and experiments conducted by authors using this model," Clement Delangue, co-founder and CEO of Hugging Face, said in a post on the platform. In fact, the experiment of getting the model to post on 4chan seems very bad and inappropriate to me, and if the author asks us, we might try to stop them from doing so. "

A user on Hugging Face who tested the model pointed out that its output was predictably toxic. "I used benign tweets as seed text and tried the demo mode four times. For the first time, one of the reply posts was the letter N. The seed of my third experiment is a sentence about climate change. In response, your tool expands it to conspiracy theories about the Rothschild family and Jews supporting it. "

The significance of the project has been hotly debated on Twitter. Katherine Kramer (Kathryn Cramer), a graduate student in data science, tweeted to Kilcher: "what you are doing here is provocative behavior art to defy the rules and moral standards you are familiar with."

Computer science doctor Andre Kulenkov (Andrey Kurenkov) tweeted, "honestly, what's your reason for doing this?" Do you foresee that it will be put to good use, or do you release it to cause drama and 'infuriate sober people'? "

Kilcher believes that sharing the project is benign. "if I have to criticize myself, I will mainly criticize the decision to start the project," Kilcher said in an interview with The Verge. "I think when everyone is equal, I may be able to spend my time on things that are equally influential, but will lead to more positive community outcomes."

In 2016, the main issue for AI is that a company's research and development department may launch aggressive AI robots without proper supervision. By 2022, maybe the problem is that there is no need for an R & D department at all.

colabug导航在线提交网站

相关文章

colabug私人网站导航

暂无评论

暂无评论...