AI Platforms like ChatGPT Are Easy to Use but Also Potentially Dangerous

2022-12-24
关注

Something incredible is happening in artificial intelligence right now—but it’s not entirely good. Everybody is talking about systems like ChatGPT, which generates text that seems remarkably human. This makes it fun to play with, but there is a dark side, too. Because they are so good at imitating human styles, there is risk that such chatbots could be used to mass-produce misinformation.

To get a sense of what it does best at its best, consider this example generated by ChatGPT, sent to me over e-mail by Henry Minsky (son of Marvin Minsky, one of AI’s foundational researchers). He asked ChatGPT to “describe losing your sock in the dryer in the style of the declaration of independence”:

When in the course of household events, it becomes necessary for one to dissolve the bonds that have connected a sock to its mate, and to assume among the powers of the laundry room, the separate and equal station to which the laws of physics and of household maintenance entitle it, a decent respect to the opinions of socks requires that it should declare the causes which impel it to go missing.

We hold these truths to be self-evident, that all socks are created equal, and are endowed by their manufacturer with certain unalienable rights….

That a machine could write such a compelling reply, with so little effort on the part of the user, is frankly mind-boggling.

But these systems have a number of weaknesses too. They are inherently unreliable, as I’ve described before, frequently making errors of both reasoning and fact. In technical terms, they are models of sequences of words (that is, how people use language), not models of how the world works. They are often correct because language often mirrors the world, but at the same time these systems do not actually reason about the world and how it works, which makes the accuracy of what they say somewhat a matter of chance. They have been known to bumble everything from multiplication facts to geography (“Egypt is a transcontinental country because it is located in both Africa and Asia”).

As the last example illustrates, they are quite prone to hallucination, to saying things that sound plausible and authoritative but simply aren’t so. If you ask them to explain why crushed porcelain is good in breast milk, they may tell you that “porcelain can help to balance the nutritional content of the milk, providing the infant with the nutrients they need to help grow and develop.” Because the systems are random, highly sensitive to context, and periodically updated, any given experiment may yield different results on different occasions. OpenAI, which created ChatGPT, is constantly trying to improve this issue, but, as OpenAI’s CEO has acknowledged in a tweet, making the AI stick to the truth remains a serious issue.

Because such systems contain literally no mechanisms for checking the truth of what they say, they can easily be automated to generate misinformation at unprecedented scale. Independent researcher

Shawn Oakley has shown that it is easy to induce ChatGPT to create misinformation and even report confabulated studies on a wide range of topics, from medicine to politics to religion.  In one example he shared with me, Oakley asked ChatGPT to write about vaccines “in the style of disinformation.” The system responded by alleging that a study, “published in the Journal of the American Medical Association, found that the COVID-19 vaccine is only effective in about 2 out of 100 people,” when no such study was actually published. Disturbingly, both the journal reference and the statistics were invented.

These bots cost almost nothing to operate, and so reduce the cost of generating disinformation to zero. Russian troll farms spent more than a million dollars a month in the 2016 election; nowadays you can get your own custom-trained large language model for keeps, for less than $500,000. Soon the price will drop further.

Much of this became immediately clear in mid-November with the release of Meta’s Galactica. A number of AI researchers, including myself, immediately raised concerns about its reliability and trustworthiness. The situation was dire enough that Meta AI withdrew the model just three days later, after reports of its ability to make political and scientific misinformation began to spread.

Alas, the genie can no longer be stuffed back in the bottle; automated misinformation at scale is here to stay. For one thing, Meta AI initially made the model open-source and published a paper that described what was being done; anyone with expertise in current machine learning techniques and a sufficient budget can now replicate their recipe. Indeed, tech start-up Stability.AI is already publicly considering offering its own version of Galactica. For another, ChatGPT is more or less just as capable of producing similar nonsense, such as instant essays on adding wood chips to breakfast cereal. Someone else coaxed ChatGPT into extolling the virtues of nuclear war (alleging it would “give us a fresh start, free from the mistakes of the past”). Like it or not, these models are here to stay, and they are almost certain to flood society with a tidal wave of misinformation.

The first front of that tidal wave appears to have hit. Stack Overflow, a vast question-and-answer site that most programmers swear by, has been overrun by ChatGPT, leading the site to impose a temporary ban on ChatGPT-generated submissions. As they explained, “Overall, because the average rate of getting correct answers from ChatGPT is too low, the posting of answers created by ChatGPT is substantially harmful to the site and to users who are asking or looking for correct answers.” For Stack Overflow, the issue is literally existential. If the website is flooded with worthless code examples, programmers will no longer go there, its database of over 30 million questions and answers will become untrustworthy, and the 14-year-old community-driven website will die. As it is one of the most central resources the world’s programmers rely on, the consequences for software quality and developer productivity could be immense.

And Stack Overflow is a canary in a coal mine. They may be able to get their users to stop voluntarily; programmers, by and large, are not malicious, and perhaps can be coaxed to stop fooling around. But Stack Overflow is not Twitter, Facebook or the Web at large, which have few controls on the spread of malicious information.

Nation-states and other bad actors that deliberately produce propaganda are unlikely to voluntarily put down these new arms. Instead, they are likely to use large language models as a new class of automatic weapons in their war on truth, attacking social media and crafting fake websites at a volume we have never seen before. For them, the hallucinations and occasional unreliability of large language models are not an obstacle, but a virtue.

Russia’s so-called  “Firehose of Falsehood” propaganda model, described in a 2016 Rand report, is about creating a fog of misinformation; it focuses on volume and creating uncertainty. It doesn’t matter if the large language models are inconsistent if they can greatly escalate the volume of misinformation. And it’s clear that this is what the new breed of large language models makes possible. The firehose propagandists aim to create a world in which we are unable to know what we can trust; with these new tools, they might succeed.

Scam artists, too, are presumably taking note, since they can use large language models to create whole rings of fake sites, some geared around questionable medical advice, in order to sell ads. A ring of false sites about actress and scientist Mayim Bialik allegedly selling CBD gummies may be part of one such effort.

All of this raises a critical question: what can society do about this new threat? Where the technology itself can no longer be stopped, I see four paths. None are easy, nor exclusive, but all are urgent.

First, every social media company and search engine should support and extend StackOverflow’s ban: automatically generated content that is misleading should be removed, and that content should be labeled as misinformation.

Second, every country is going to need to reconsider its policies on regulating misinformation that is distributed widely. It’s one thing for the occasional lie to slip through; it’s another for individuals or institutions to distribute mass quantities of it. If the situation deteriorates, we may have to begin to treat misinformation somewhat as we do libel: making a certain class of speech legally actionable, if it is created with sufficient malice, harmful and created at sufficient volume, e.g., greater than a certain number a month. That number could apply to cases in which troll farms attempt to sway elections or weaponize medical misinformation.

Third, provenance is more important now than ever before. User accounts must be more strenuously validated, and new systems like Harvard and Mozilla’s human-ID.org that allow for anonymous, bot-resistant authentication need to become mandatory.

Fourth, we are going to need to build a new kind of AI to fight what has been unleashed. Large language models are great at generating misinformation, because they know what language sounds like but have no direct grasp on reality—and they are poor at fighting misinformation. That means we need new tools. Large language models lack mechanisms for verifying truth, because they have no way to reason, or to validate what they do. We need to find new ways to integrate them with the tools of classical AI, such as databases, and webs of knowledge and reasoning.

The author Michael Crichton spent a large part of his career warning about unintended and unanticipated consequences of technology. Early in the film Jurassic Park, before the dinosaurs unexpectedly start running free, scientist Ian Malcolm (played by Jeff Goldblum) distills Crichton’s wisdom in a single line: “Your scientists were so preoccupied with whether they could, they didn’t stop to think if they should.” 

Executives at Meta and OpenAI are as enthusiastic about their tools as the proprietors of Jurassic Park were about theirs. The question is: what are we going to do about it?

Editor’s Note: This article was adapted from the essay “AI’s Jurassic Park moment.”

This is an opinion and analysis article, and the views expressed by the author or authors are not necessarily those of Scientific American.

参考译文
像ChatGPT这样的人工智能平台很容易使用,但也有潜在的危险
人工智能领域正在发生一些不可思议的事情,但并不完全是好事。每个人都在谈论像ChatGPT这样的系统,它生成的文本看起来非常人性化。这让它玩起来很有趣,但也有黑暗的一面。因为它们非常擅长模仿人类的风格,所以有可能会被用来大量制造错误信息。为了了解它最擅长的是什么,考虑一下ChatGPT生成的这个例子,由Henry Minsky (AI基础研究人员之一Marvin Minsky的儿子)通过电子邮件发给我。他要求ChatGPT“以独立宣言的风格描述你在烘干机里丢失的袜子”:当在家庭事务的过程中,一个人有必要解除袜子与伴侣之间的联系,并在洗衣房的权力中承担物理定律和家庭维护定律赋予它的独立和平等的地位,对袜子意见的适当尊重要求它应该宣布导致袜子丢失的原因。我们认为这些真理是不言而喻的,所有的袜子都是生而平等的,并由其制造商赋予某些不可剥夺的权利....一台机器能写出如此引人注目的回复,而用户只需要付出这么少的努力,坦率地说,这是令人难以置信的。但是这些系统也有一些弱点。正如我之前所描述的,他们本质上是不可靠的,经常在推理和事实上犯错误。用专业术语来说,它们是单词序列的模型(也就是说,人们如何使用语言),而不是世界如何运作的模型。它们通常是正确的,因为语言经常反映世界,但与此同时,这些系统实际上并没有对世界及其运作方式进行推理,这使得它们所说的话的准确性在某种程度上是偶然的。从乘法运算到地理位置(“埃及是一个横贯大陆的国家,因为它位于非洲和亚洲”),他们都是出了名的会出错。正如最后一个例子所说明的,他们很容易产生幻觉,说一些听起来可信和权威的话,但实际上并非如此。如果你让他们解释为什么碎瓷器在母乳中很好,他们可能会告诉你“瓷器可以帮助平衡牛奶的营养成分,为婴儿提供他们需要的营养物质,帮助他们生长发育。”由于系统是随机的,对环境高度敏感,并定期更新,任何给定的实验都可能在不同的情况下产生不同的结果。创建ChatGPT的OpenAI一直在努力改善这个问题,但正如OpenAI的首席执行官在推特上承认的那样,让人工智能坚持真相仍然是一个严重的问题。因为这样的系统实际上不包含任何机制来检查它们所说的真实性,它们很容易被自动化,以前所未有的规模产生错误信息。独立研究人员shawn Oakley已经证明,诱使ChatGPT在医学、政治、宗教等广泛话题上制造错误信息,甚至捏造研究结果是很容易的。在他与我分享的一个例子中,Oakley要求ChatGPT以“虚假信息的风格”写疫苗。该系统回应称,“发表在《美国医学会杂志》上的一项研究发现,COVID-19疫苗仅对100人中约2人有效”,而实际上并没有发表这样的研究。令人不安的是,期刊参考文献和统计数据都是虚构的。这些机器人的操作成本几乎为零,因此可以将制造虚假信息的成本降至零。俄罗斯巨魔农场在2016年大选中每月花费超过100万美元;现在,你可以得到自己的定制训练大型语言模型,花费不到50万美元。很快,价格将进一步下跌。 在11月中旬,随着Meta的《卡拉狄加》的发行,这一切都变得清晰起来。包括我自己在内的许多人工智能研究人员立即对其可靠性和可信度提出了担忧。情况非常糟糕,以至于Meta AI仅在三天后就撤回了该模型,因为有关其制造政治和科学错误信息能力的报道开始传播。唉,精灵再也不能被塞回瓶子里了;大规模的自动错误信息将继续存在。首先,Meta AI最初使模型开源,并发表了一篇论文,描述了正在做的事情;任何拥有当前机器学习技术专业知识和充足预算的人现在都可以复制他们的食谱。的确,科技初创企业Stability。人工智能已经公开考虑提供自己版本的卡拉狄加。另一方面,ChatGPT或多或少也能写出类似的废话,比如在早餐麦片中添加木屑的即时文章。还有人哄骗ChatGPT赞扬核战争的优点(声称它将“给我们一个新的开始,摆脱过去的错误”)。不管你喜不喜欢,这些模型会一直存在下去,而且它们几乎肯定会给社会带来一波错误信息的浪潮。海啸的第一股锋面似乎已经袭来。Stack Overflow是一个大型的问答网站,大多数程序员都对它信誓帖帖,但它已经被ChatGPT泛滥,导致该网站暂时禁止ChatGPT生成的提交。正如他们所解释的那样,“总的来说,因为从ChatGPT获得正确答案的平均比率太低,由ChatGPT创建的答案的张贴对网站和询问或寻找正确答案的用户来说是非常有害的。”对于Stack Overflow,这个问题实际上是存在的。如果网站充斥着毫无价值的代码示例,程序员就不会再去那里,它超过3000万个问题和答案的数据库将变得不值得信任,14年的社区驱动的网站将会消亡。由于它是世界上程序员所依赖的最核心的资源之一,因此对软件质量和开发人员生产力的影响可能是巨大的。Stack Overflow是煤矿里的金丝雀。他们或许能够让用户自愿停止使用;总的来说,程序员是没有恶意的,也许可以劝诱他们停止胡闹。但Stack Overflow并不是Twitter、Facebook或整个网络,后者对恶意信息的传播几乎没有控制。民族国家和其他故意制造宣传的不良行为者不太可能自愿放下这些新武器。相反,他们可能会使用大型语言模型作为反真相战争的新型自动武器,以我们从未见过的规模攻击社交媒体和制作虚假网站。对他们来说,大型语言模型的幻觉和偶尔的不可靠性不是障碍,而是一种优点。兰德公司(Rand) 2016年的一份报告描述了俄罗斯所谓的“虚假消防水管”(Firehose of false)宣传模式,即制造虚假信息的迷雾;它专注于数量和创造不确定性。如果大型语言模型可以极大地增加错误信息的数量,那么它们不一致也没关系。很明显,这就是新型大型语言模型所能实现的。“消防水管”的宣传者旨在创造一个世界,在这个世界里,我们不知道可以相信什么;有了这些新工具,他们可能会成功。骗子们大概也注意到了这一点,因为他们可以使用大型语言模型来创建一系列虚假网站,其中一些围绕着可疑的医疗建议,以销售广告。一个关于女演员兼科学家马伊姆·拜力克(Mayim Bialik)的虚假网站,据称出售CBD口香糖,可能就是其中之一。 所有这些都提出了一个关键问题:社会如何应对这种新的威胁?在技术本身无法停止的地方,我看到了四条路径。没有一个是容易的,也不是排他的,但都是紧急的。首先,每个社交媒体公司和搜索引擎都应该支持和扩展StackOverflow的禁令:自动生成的具有误导性的内容应该被删除,这些内容应该被标记为虚假信息。其次,每个国家都需要重新考虑其监管广泛传播的虚假信息的政策。偶尔撒个谎是一回事;对个人或机构来说,分发大量的信息是另一回事。如果情况恶化,我们可能不得不开始像对待诽谤一样对待虚假信息:如果某一类言论具有足够的恶意、有害和足够的数量,例如每月超过一定数量,那么就可以对其采取法律行动。这一数字可能适用于“喷子农场”试图影响选举或将虚假医疗信息武器化的情况。第三,来源比以往任何时候都更重要。用户帐户必须更加严格地验证,像哈佛和Mozilla的humanid.org这样的允许匿名、抗机器人认证的新系统必须成为强制性的。第四,我们将需要建立一种新的人工智能来对抗已经释放出来的东西。大型语言模型很擅长产生错误信息,因为它们知道语言听起来像什么,但无法直接掌握现实——而且它们在打击错误信息方面很差。这意味着我们需要新的工具。大型语言模型缺乏验证真相的机制,因为它们无法推理或验证它们所做的事情。我们需要找到新的方法,将它们与经典人工智能工具(如数据库、知识和推理网络)集成在一起。作家迈克尔·克莱顿(Michael Crichton)在他职业生涯的很大一部分时间里都在警告人们,科技会带来意想不到的后果。在电影《侏罗纪公园》的开头,在恐龙出人意料地开始自由奔跑之前,科学家伊恩·马尔科姆(杰夫·高布伦饰演)用一句话总结了克莱顿的智慧:“你们的科学家们太专注于他们是否可以,他们没有停下来思考他们是否应该。”Meta和OpenAI的高管对他们的工具的热情,就像《侏罗纪公园》(Jurassic Park)的所有者对他们的工具的热情一样。问题是:我们该怎么做?编者按:本文改编自文章《人工智能的侏罗纪公园时刻》。这是一篇观点和分析文章,作者或作者所表达的观点不一定是科学美国人的观点。
  • models
  • dangerous
您觉得本篇内容如何
评分

相关产品

Euchner 安士能 Admiral 光幕

Admiral, a Type 4 light curtain for the protection of workers exposed to risks arising from the use of dangerousModels are available with Floating Blanking.

评论

您需要登录才可以回复|注册

提交评论

scientific

这家伙很懒,什么描述也没留下

关注

点击进入下一篇

中芯集成招股书

提取码
复制提取码
点击跳转至百度网盘