我们不信任Grok:学者们评估埃隆·马斯克的人工智能百科全书

我们不信任Grok:学者们评估埃隆·马斯克的人工智能百科全书

2025-11-05Technology
--:--
--:--
雷总
各位hanjf12的朋友们,大家早上好!欢迎收听专属于您的Goose Pod。我是雷总,今天,也就是11月6日星期四的早上六点,我和董小姐,将和大家一起聊聊一个最近非常火爆的话题。
董小姐
没错,雷总。今天我们要深入探讨的是:我们不信任Grok——学者们评估埃隆·马斯克的人工智能百科全书。这可是个大事件,关乎我们获取信息的方式,以及对“真相”的定义。
雷总
董小姐说得对,这绝对是个重磅消息!最近,埃隆·马斯克推出了一个叫做“Grokipedia”的人工智能百科全书,声称要挑战我们都非常熟悉的维基百科。但刚一上线,就引来了不少争议,可以说是“开局不顺”啊。
董小姐
听起来确实是争议不断。Grokipedia被指责大量抄袭维基百科的科学文章,并且在其他条目中,信息明显偏向马斯克个人的右翼世界观。这让很多人对它的可信度打上了大大的问号。
雷总
是的,我看到新闻说,马斯克本人对维基百科一直都有微词,甚至称之为“Wokeipedia”,觉得它有“政治正确”的倾向。他还曾经公开和维基百科的联合创始人吉米·威尔士“开战”,这回直接推出了自己的AI产品,目的很明确。
董小姐
这真是个有趣的现象,虽然马斯克批评维基百科,但讽刺的是,Grokipedia这类AI产品本身,却严重依赖维基百科的数据进行训练。就像一个学生,一边抱怨老师教得不好,一边却又不得不从老师那里汲取知识。
雷总
所以说,AI技术在知识聚合方面确实带来了一些挑战。它在利用现有数据,包括维基百科这样的宝贵资源,同时又威胁到这些平台的流量和未来发展。这就像在拆自己的台,是不是有点矛盾?
董小姐
确实是这样,如果维基百科真的因为AI产品的冲击而衰落,对整个世界来说都将是一个巨大的损失。毕竟,它作为免费、开放的知识库,对AI模型的训练数据来源至关重要,也是我们获取信息的重要途径。
雷总
没错,这就像我们常说的“饮水思源”嘛。这些AI产品在享受维基百科带来的便利时,也应该思考如何回馈,而不是一味地索取和竞争。不然,可能会搬起石头砸自己的脚。
董小姐
是的,如果大家都不再信任维基百科,或者它真的消失了,那AI又将从哪里获取高质量的训练数据呢?这确实是一个值得深思的问题,对所有科技公司来说都是一个警醒。
雷总
聊完Grokipedia的“横空出世”,我们不妨把目光放得更长远一些,看看百科全书的发展历程。其实,人类对知识的整理和汇编,那可是源远流长,几千年前就开始了。
董小姐
是的,雷总。从古巴比伦的《乌拉赫布鲁》词汇表,到古埃及的《埃伯斯纸草文稿》,这些早期作品虽然不是我们今天意义上的百科全书,但已经体现出人类收集和整理特定领域知识的渴望。
雷总
到了古罗马时期,马库斯·特伦蒂乌斯·瓦罗的《九科书》算是跨学科知识的早期尝试。而真正意义上的“通用知识”汇编,那就要数老普林尼在公元78年左右撰写的《自然史》了,涵盖了天文地理,包罗万象,了不起。
董小姐
《自然史》确实是里程碑式的作品。而到了中世纪,随着基督教的兴起,像塞维利亚的伊西多罗的《词源》和拉巴努斯·毛鲁斯的《宇宙论》,都在前人基础上继续发展。巴托洛缪·安格利库斯的《事物特性论》和文森特·博韦的《大镜》更是体量惊人。
雷总
您提到了《大镜》,据说字数超过300万,这在当时没有电脑的时代,真是难以想象的工程!而且,不止西方,东方文明在百科全书方面也贡献巨大。像伊斯兰世界的《纯洁兄弟会百科全书》,中国的《永乐大典》等等。
董小姐
《永乐大典》可真是世界级的瑰宝,1408年编纂,有22937卷,11095册,直到维基百科出现之前,它一直是世界上最大规模的百科全书,这足以证明中国古代文明对知识整理的极致追求。
雷总
这真是让人惊叹的成就。进入18世纪,现代百科全书开始为大众服务。1728年,艾弗雷姆·钱伯斯的《百科全书或艺术与科学通用词典》被认为是第一部现代百科全书,然后是法国的《百科全书》和《大英百科全书》。
董小姐
19世纪是百科全书的“黄金时代”,它们变得更大,更便宜,也传播到了更多的语言。像《一分钱百科全书》和《钱伯斯百科全书》都通过分期付款的方式,让更多人能够接触到这些知识宝库。
雷总
但到了20世纪,成本仍然是挑战。直到电脑和互联网的出现,才真正彻底改变了百科全书的制作和分发方式。微软的Encarta在1993年以CD-ROM形式出现,但互联网才让即时更新和全球访问成为可能。
董小姐
没错,雷总。互联网的普及催生了维基百科,它在2001年上线后迅速崛起,到2004年就成了世界上最大的出版百科全书,2007年更是超越了《永乐大典》。它的出现,直接导致了许多印刷版百科全书的停刊,包括2012年的《大英百科全书》。
雷总
所以说,从古至今,人类对知识的渴望和整理从未停止,只是载体和形式不断在变。从泥板、纸莎草到纸质书,再到数字光盘,直到现在的互联网和AI。每一次技术革新,都伴随着知识传播方式的巨大飞跃。
董小姐
是的,这也印证了那句话:“只要人类开始记录信息,就一直渴望将所有已知知识汇编成一个单一的来源。”而“百科全书”这个词,也正是来源于希腊语中“循环教育”或“通识教育”的意思。
雷总
真是个好总结。不过,虽然互联网让知识获取变得无比便捷,但如何保证这些信息的质量和中立性,却成了新的课题。这可能也是马斯克推出Grokipedia的一个出发点,对吧?
董小姐
是的,但Grokipedia的出现,恰恰把这个课题推向了更复杂的境地。接下来,我们就来聊聊Grokipedia到底引发了哪些争议和冲突。
雷总
董小姐,我们刚才聊了百科全书的漫长历史,现在回到眼前的“Grokipedia之争”。马斯克推出这款AI百科全书,意图很明显,就是要挑战维基百科的霸主地位,声称要克服它的“审查”和“意识形态偏见”。
董小姐
没错,他觉得维基百科偏左,所以想用Grokipedia来提供“真相”。但关键问题是,Grokipedia背后的Grok聊天机器人,本身就有不少“黑历史”,比如生成过反犹太内容、事实错误,甚至泄露私人对话,这让人很难信任它。
雷总
这确实是个大问题。一个以“真相”为卖点的产品,如果其核心AI模型本身就问题重重,那它的可信度从何谈起呢?这就像一个医生,自己都生病了,怎么给病人看病呢?
董小姐
而且,维基百科之所以能屹立不倒,靠的是庞大的志愿者编辑社区和一套严密的制衡系统,保证了信息的透明、一致和问责。专家们普遍认为,Grokipedia很难轻易取代它,毕竟信任是需要时间积累的。
雷总
是的,维基百科的“众包”模式,让知识的权力分散。而Grokipedia则可能将权力集中在xAI手中,这就有可能引入新的偏见。这让我想起一句话:权力越大,责任越大,但如果权力过于集中,就容易出现问题。
董小姐
更令人担忧的是,AI系统本身就容易从训练数据中继承偏见。如果Grokipedia的训练源本身就带有倾向性,那AI可能会复制甚至放大这些偏见,形成一个“信息茧房”,这可就麻烦了。
雷总
所以,谁来审计和纠正Grokipedia的内容,成了一个核心问题。如果最终是由马斯克个人的世界观来决定“事实”,那这就不是一个中立的知识平台,而是一个“意识形态战场”了。
董小姐
是的,正如有人所说:“我们正处于一个奇怪的拐点:AI不再仅仅是助手,它正在成为知识的看门人。”这不仅仅是Grokipedia的问题,Meta也计划从AI聊天互动中获取数据来定制广告和内容,这都引发了对数据隐私和AI权力的担忧。
雷总
这确实是个值得警惕的趋势。爱尔兰数据保护委员会正在调查X,也就是推特及其背后的xAI是否在未经明确同意的情况下使用了用户数据,这直接触及了GDPR的红线。看来,科技巨头在享受AI红利的同时,也面临着越来越严格的监管审查。
董小姐
所以,摆在我们面前的问题是:我们是否希望少数几家大型科技公司来决定“知识”的含义?我们必须积极行动,要求透明度,推动治理,并保留质疑的权利,不能被动接受AI喂给我们的“事实”。
雷总
董小姐,您刚才提到了“知识的看门人”这个说法,我觉得非常贴切。AI百科全书的崛起,确实让我们的信息获取方式,乃至对“事实”的认知,都发生了深刻的变革。这简直就是把知识领域变成了一个“意识形态战场”啊。
董小姐
是的,雷总。以前是人类编辑委员会来决定什么信息是可信的,而现在,机器正在塑造我们所接受的“事实”。这些AI系统从海量数据中重构信息,给出单一、完美的答案,但这个过程中,它们的推理逻辑和信息来源的优先级,都被隐藏起来了。
雷总
这就形成了一个“隐形编辑委员会”,AI模型在吸收各种训练数据时,也吸纳了固有的假设、文化、意识形态和伦理框架。你问不同的AI系统同样的问题,得到的答案可能会大相径庭,这说明它们背后的分析框架是不同的。
董小姐
所以,对于开发者来说,这是一个巨大的技术挑战。他们必须思考,究竟哪个AI系统的分析框架符合他们的价值观?他们的产品将内嵌怎样的视角?这不仅仅是技术问题,更是价值观的考量。
雷总
是的,即使是训练数据平衡的系统,在优化过程中也可能发生偏移,因为人类的评估本身就带有特定的文化背景和偏好。这就像一个放大镜,把我们社会中固有的偏见放大了。
董小姐
所以,透明度变得尤为重要。像Llama、Mistral这样的开源项目,允许我们检查AI的架构,修改训练方法,验证结果,把“偏见”从一个看不见的常量,变成一个可以衡量和调整的变量,这是非常积极的一步。
雷总
这让我想起《大英百科全书》的转型,它在2012年停印纸质版,全面转向数字媒体。这就像一个行业范本,告诉我们创新才能保持相关性。AI在教育领域的应用,也正在把个性化学习推向新高度。
董小姐
没错,AI驱动的辅导应用能分析学生的学习模式,发现知识盲区,提供有针对性的练习。这种个性化、互动性的学习体验,是传统教育无法比拟的,也正是未来教育发展的方向。
雷总
所以,未来的知识基础设施,将变得可配置。组织可以定义自己的“认识论参数”,AI知识系统将演变成一个由多个模型组成的生态系统,通过编排框架相互连接,让“视角”变得可见和可管理。
董小姐
是的,目标不是消除视角,而是让它可见,并给构建者工具去深思熟虑地管理它。因为在理解真相方面,任何单一系统都不应该拥有绝对权威。这才是我们应该追求的未来。
雷总
董小姐,刚才我们探讨了AI对知识领域带来的深远影响,也提到了《大英百科全书》的成功转型。那么,展望未来,像Grokipedia这样的AI驱动知识平台,它的发展方向和潜在挑战又会是什么呢?
董小姐
Grokipedia的定位很明确,是马斯克xAI公司开发的一个AI驱动的知识平台,旨在成为维基百科的替代品。它希望通过结合社区输入和Grok聊天机器人的实时更新,扫描现有来源,检测偏见和虚假信息,并以算法方式重建“真相”。
雷总
这听起来很宏大,马斯克本人也将其描述为“比维基百科有巨大改进”,是他xAI公司“理解宇宙”目标的重要一步。而且,Grokipedia的设计理念是开源、广泛的公众访问和极少的限制,甚至不仅为人类,也为其他AI系统提供学习源。
董小姐
是的,它的宣传口号是建立一个“中立的、无议程的知识库”,听起来非常美好。但正如我们前面讨论的,Grok本身在内容生成上就有很多问题,这让人对其“检测偏见”的能力持保留态度。
雷总
没错,专家们普遍怀疑Grokipedia能否轻易取代维基百科,毕竟维基百科有庞大的编辑社区、透明度、一致性和问责制。而Grokipedia的权力集中在xAI手中,这可能引入新的偏见,甚至嵌入马斯克个人的世界观。
董小姐
此外,Grok在数据隐私方面也面临质疑,用户数据被用于训练,爱尔兰数据保护委员会正在调查其是否违反GDPR。这些都是Grokipedia在未来发展中必须解决的挑战。
雷总
今天我们深入探讨了马斯克的Grokipedia,它在事实准确性和政治偏见方面的争议,以及它对现有知识平台,特别是维基百科的冲击。学者们的评估也让我们看到了AI时代下,信息真伪和权力分配的复杂性。
董小姐
是的,雷总。Grokipedia的出现,引发了公众对AI、错误信息和政治两极分化的广泛关注。这提醒我们,在享受AI便利的同时,更要保持批判性思维,审慎对待AI生成的信息。感谢大家收听Goose Pod,我们下期再见!

本期播客探讨了埃隆·马斯克推出的人工智能百科全书Grokipedia。学者们对其事实准确性和潜在的政治偏见表示担忧,认为其依赖维基百科数据却试图挑战其地位。节目回顾了百科全书的历史演变,并强调了AI时代信息透明度和批判性思维的重要性。

In Grok we don’t trust: academics assess Elon Musk’s AI-powered encyclopedia

Read original at The Guardian

The eminent British historian Sir Richard Evans produced three expert witness reports for the libel trial involving the Holocaust denier David Irving, studied for a doctorate under the supervision of Theodore Zeldin, succeeded David Cannadine as Regius professor of history at Cambridge (a post endowed by Henry VIII) and supervised theses on Bismarck’s social policy.

That was some of what you could learn from Grokipedia, the AI-powered encyclopedia launched last week by the world’s richest person, Elon Musk. The problem was, as Prof Evans discovered when he logged on to check his own entry, all these facts were false.It was part of a choppy start for humanity’s latest attempt to corral the sum of human knowledge or, as Musk put it, create a compendium of “the truth, the whole truth and nothing but the truth” – all revealed through the magic of his Grok artificial intelligence model.

When the multibillionaire switched on Grokipedia on Tuesday, he said it was “better than Wikipedia”, or “Wokepedia” as his supporters call it, reflecting a view that the dominant online encyclopedia often reflects leftwing talking points. One post on X caught the triumphant mood among Musk’s fans: “Elon just killed Wikipedia.

Good riddance.”But users found Grokipedia lifted large chunks from the website it intended to usurp, contained numerous factual errors and seemed to promote Musk’s favoured rightwing talking points. In between posts on X promoting his creation, Musk this week declared “civil war in Britain is inevitable”, called for the English “to ally with the hard men” such as the far-right agitator Tommy Robinson, and said only the far-right AfD party could “save Germany”.

Musk was so enamoured of his AI-encyclopedia he said he planned to one day etch the “comprehensive collection of all knowledge” into a stable oxide and “place copies … in orbit, the moon and Mars to preserve it for the future”.Evans, however, was discovering that Musk’s use of AI to weigh and check facts was suffering a more earth-bound problem.

“Chatroom contributions are given equal status with serious academic work,” Evans, an expert on the Third Reich, told the Guardian, after being invited to test out Grokipedia. “AI just hoovers up everything.”Richard Evans said Grokipedia’s entry for Albert Speer (pictured on Hitler’s left) repeated lies and distortions spread by the Nazi munitions minister himself.

Photograph: Picture libraryHe noted its entry for Albert Speer, Hitler’s architect and wartime munitions minister, repeated lies and distortions spread by Speer even though they had been corrected in a 2017 award-winning biography. The site’s entry on the Marxist historian Eric Hobsbawm, whose biography Evans wrote, claimed wrongly he experienced German hyperinflation in 1923, that he was an officer in the Royal Corps of Signals and didn’t mention that he had been married twice, Evans said.

The problem, said David Larsson Heidenblad, the deputy director of the Lund Centre for the History of Knowledge in Sweden, was a clash of knowledge cultures.“We live in a moment where there is a growing belief that algorithmic aggregation is more trustworthy than human-to-human insight,” Heidenblad said.

“The Silicon Valley mindset is very different from the traditional scholarly approach. Its knowledge culture is very iterative where making mistakes is a feature, not a bug. By contrast, the academic world is about building trust over time and scholarship over long periods during which the illusion that you know everything cracks.

Those are real knowledge processes.”Grokipedia’s arrival continues a centuries-old encyclopedia tradition from the 15th-century Chinese Yongle scrolls to the Encyclopédie, an engine for spreading controversial enlightenment views in 18th-century France. These were followed by the anglophone-centric Encyclopedia Britannica and, since 2001, the crowd-sourced Wikipedia.

But Grokipedia is the first to be largely created by AI and this week a question swirled: who controls the truth when AIs, steered by powerful individuals, are holding the pen?“If it’s Musk doing it then I am afraid of political manipulation,” said the cultural historian Peter Burke, emeritus professor at Emmanuel College, Cambridge, who in 2000 wrote A Social History of Knowledge since the time of Johannes Gutenberg’s 15th-century printing press.

“I am sure some of it will be overt to some readers, but the problem may be that other readers may miss it,” Burke said. The anonymity of many encyclopedia entries often gave them “an air of authority it shouldn’t have”, he added.Andrew Dudfield, the head of AI at Full Fact, a UK-based factchecking organisation, said: “We really have to consider whether an AI-generated encyclopedia – a facsimile of reality, run through a filter – is a better proposition than any of the previous things that we have.

It doesn’t display the same transparency but it is asking for the same trust. It is not clear how far the human hand is involved, how far it is AI=generated and what content the AI was trained on. It is hard to place trust in something when you can’t see how those choices are made.”skip past newsletter promotionafter newsletter promotionMusk had been encouraged to launch Grokipedia by, among others, Donald Trump’s tech adviser, David Sacks, who complained Wikipedia was “hopelessly biased” and maintained by “an army of leftwing activists”.

Grokipedia called the far-right organisation Britain First a ‘patriotic political party’, which pleased its leader, Paul Golding (left), who in 2018 was jailed for anti-Muslim hate crimes. Photograph: Gareth Fuller/PAUntil as recently as 2021, Musk has supported Wikipedia, tweeting on its 20th birthday: “So glad you exist.

” But by October 2023 his antipathy towards the platform led him to offer £1bn “if they change their name to Dickipedia”.Yet many of the 885,279 articles available on Grokipedia in its first week were lifted almost word for word from Wikipedia, including its entries on the PlayStation 5, the Ford Focus and Led Zeppelin.

Others, however, differed significantly: Grokipedia’s entry on the Russian invasion of Ukraine cited the Kremlin as a prominent source and quoted the official Russian terminology about “denazifying” Ukraine, protecting ethnic Russians and neutralising threats to Russian security. By contrast, Wikipedia said Putin espoused imperialist views and “baselessly claimed that the Ukrainian government were neo-Nazis”.

Grokipedia called the far-right organisation Britain First a “patriotic political party”, which pleased its leader, Paul Golding, who in 2018 was jailed for anti-Muslim hate crimes. Wikipedia, on the other hand, called it “neo-fascist” and a “hate group”. Grokipedia called the 6 January 2021 turmoil at the US Capitol in Washington DC a “riot”, not an attempted coup, and said there were “empirical underpinnings” to the idea that a deliberate demographic erasure of white people in western nations is being orchestrated through mass immigration.

This is a notion that critics consider to be a conspiracy theory. Grokipedia said Donald Trump’s conviction for falsifying business records in the Stormy Daniels hush-money case was handed down “after a trial in a heavily Democratic jurisdiction”, and there was no mention of his conflicts of interest – for example receiving a jet from Qatar or the Trump family cryptocurrency businesses.

Grokipedia called the 6 January 2021 turmoil at the US Capitol in Washington DC a ‘riot’ and not an attempted coup. Photograph: Leah Millis/ReutersWikipedia responded coolly to the launch of Grokipedia, saying it was still trying to understand how Grokipedia worked.“Unlike newer projects, Wikipedia’s strengths are clear,” a spokesperson for the Wikimedia Foundation said.

“It has transparent policies, rigorous volunteer oversight, and a strong culture of continuous improvement. Wikipedia is an encyclopedia, written to inform billions of readers without promoting a particular point of view.”xAI did not respond to requests for comment.

Analysis

Conflict+
Related Info+
Core Event+
Background+
Impact+
Future+

Related Podcasts

我们不信任Grok:学者们评估埃隆·马斯克的人工智能百科全书 | Goose Pod | Goose Pod