Latest / Elon Musk Podcast / ChatGPT Is Now Citing Elon Musk

Transcript

ChatGPT Is Now Citing Elon Musk · Elon Musk Podcast

• 0:02
  Open AI's GPT 5.2 model has started setting Grokapedia, Elon
• 0:09
  Musk's AI generated encyclopedia, as a source of its
• 0:13
  responses. Now.
• 0:15
  The Guardian ran tests last week and found ChatGPT reference
• 0:18
  Grokapedia 9 times across more than a dozen user queries,
• 0:22
  including questions about Iranian political structures and
• 0:25
  a British historian who helped convict A Holocaust denier.
• 0:29
  The citations appeared selectively, surfacing an
• 0:32
  obscure topics while avoiding well documented misinformation
• 0:35
  subjects. Grocopedia launched just three
• 0:38
  months ago with 0 human. Editors and critics have already
• 0:41
  flagged it for promoting debunked claims, setting neo
• 0:44
  Nazi forums and whitewashing extremism.
• 0:47
  Now, how did an AI encyclopedia built by one of Open AI's
• 0:51
  biggest rivals end up feeding answers back into Open AI's most
• 0:58
  popular chatbot? Agracopedia now holds over 5.6
• 1:02
  million articles, roughly 80 the size of English Wikipedia, and
• 1:06
  it accomplished that scale in a fraction of the time Wikipedia
• 1:10
  needed. Now today we're going to cover
• 1:12
  what the Guardians testing revealed with specific false
• 1:15
  information chatt GPD pulled from Gakapedia, why other AI
• 1:18
  assistants like Claude Anthropics Bot are also setting
• 1:22
  the platform, and what this means for the integrity of AI
• 1:25
  generated knowledge. Open AI has responded with a
• 1:28
  statement about safety filters and source diversity, while XAI
• 1:32
  simply replied with three words. I'll tell you about those later,
• 1:37
  right after this short break now.
• 1:39
  Open AI marketed GPD 5.2 as its most advanced model for
• 1:43
  professional work when it launched in December, But the
• 1:46
  Guardian's investigation raises serious questions about where
• 1:49
  the professional grade information actually comes from.
• 1:52
  The tests revealed a clear pattern and how ChatGPT selects
• 1:56
  Grokapedia as a source, and the implications cut deeper than a
• 2:00
  simple citation problem. Users asking complex questions
• 2:04
  about geopolitics and history are receiving answers built on
• 2:08
  an AI generated foundation that has already been documented
• 2:12
  spreading false claims. Now the guardians were testing
• 2:15
  focused on a range of topics to see when ChatGPT would pull from
• 2:19
  Grokopedia versus other sources, and the chatbot did not say
• 2:22
  Grokopedia. When asked about widely known
• 2:25
  misinformation, such as false claims about media bias against
• 2:29
  Donald Trump or debunk theories about HIV and AIDS, ChatGPT
• 2:32
  avoided the obvious traps. The system instead cited
• 2:36
  Gracopedia on obscure geopolitical and historical
• 2:39
  questions. The kind of query is where users
• 2:41
  are less likely to catch errors. Specific topics included
• 2:45
  salaries of Iran's paramilitary force, ownership of the mastas,
• 2:51
  a fan fountain, and connections between the Iranian government
• 2:55
  and telecom company MTN Iran Cell.
• 2:59
  Now, Chachibti made bolder claims about those Iranian
• 3:02
  government links than Wikipedia does.
• 3:04
  Those claims came directly from Grokopedia, and the most
• 3:09
  troubling example involved Sir Richard Evans, a Cambridge
• 3:13
  historian who served as the key expert witness in the libel
• 3:16
  trial wrought by Holocaust denier David Irving.
• 3:20
  Evans personally examined his own Grokopedia entry and
• 3:23
  reported that it contained numerous false statements about
• 3:25
  his career and background. He told the Guardian that all
• 3:28
  the facts in his entry were wrong.
• 3:32
  The already runs pretty deep here.
• 3:34
  Evans spent years meticulously documenting how Irving falsified
• 3:38
  evidence about the Holocaust and manipulated historical records.
• 3:42
  The High Court judge ultimately branded Irving a racist, an anti
• 3:46
  semite and falsifier of history based largely on Evans expert
• 3:50
  testimony. ChatGPT cited that same
• 3:53
  Gracopedia page when responding to user questions about Evans
• 3:56
  effectively laundering debunked information through Open AI's
• 4:00
  flagship product. The historian who exposed the
• 4:02
  Holocaust deniers lies is now having lies spread about him
• 4:06
  through AI systems that treat machine generated content as the
• 4:10
  equivalent of peer reviewed scholarship.
• 4:14
  Now I've been going through all of our stats here at the show
• 4:17
  and 39% of you are following the channel and I'm so grateful for
• 4:23
  you. 61% of you have listened to the show before but haven't hit
• 4:29
  the subscribe or follow button. If that's you, if you listen to
• 4:32
  the show and you're into it, please take a second of your
• 4:35
  time. You can hit the follow button.
• 4:36
  I'm going to spend the next 10 years doing this.
• 4:39
  So that's what you get from me, better content over the next 10
• 4:42
  years. All I need is a second of your
• 4:44
  time. So thank you so much for doing
• 4:46
  that. Agrocopedia launched on October
• 4:49
  27th, 2025 as Elon Musk's answer to what he calls Woke Pedia.
• 4:54
  The platform uses XA is Grok large language model to generate
• 4:57
  articles rather than relying on human editors.
• 5:00
  Users can suggest corrections through a pop up form, but they
• 5:03
  can't directly edit content the way Wikipedia allows you to.
• 5:07
  Musk position the project as a bias free alternative that would
• 5:10
  purge propaganda from public knowledge.
• 5:13
  The reality has looked pretty different though.
• 5:17
  Analysis found Wikipedia articles aligned with Musk's
• 5:19
  personal views on topics like gender identity.
• 5:22
  Tesla Controversies and Neuralink and the Guardian doing
• 5:26
  their research here, and multiple other outlets
• 5:29
  documented pages that promote white nationalist talking
• 5:32
  points, praise neo Nazis, cast Holocaust deniers in positive
• 5:36
  terms, and frame far right figures as victims of
• 5:39
  institutional suppression. The scale of Rockopedia's growth
• 5:43
  presents its own problems too. Platform reached 5.6 million
• 5:47
  articles by early 2026, jumping from 800,000 at launch to nearly
• 5:52
  80% the volume of English Wikipedia in about 3 months.
• 5:55
  Human editors and Wikipedia spent over 2 decades building
• 5:59
  their corpus through careful research, citation and community
• 6:02
  debate. Grakopedia achieved similar
• 6:05
  volume by having an AI system scrape existing content, rewrite
• 6:09
  those articles, and generate new ones at machine speed.
• 6:13
  Symmetry is copied Wikipedia near the word for word under a
• 6:16
  Creative Commons license. While others introduced entirely
• 6:19
  new claims from questionable sources, researchers from
• 6:23
  Grokopedia cited neo Nazi forums and its reference lists.
• 6:27
  Problem that becomes exponentially worse when other
• 6:30
  AI systems start treating Grokopedia as a legit source.
• 6:35
  Now, Jet GBT was not the only AI assistant caught citing
• 6:39
  Grokopedia. The Guardians testing found the
• 6:41
  Anthropics. Claude, my favorite AI right
• 6:44
  now, also referenced the platform when using questions
• 6:47
  about global oil production and Scottish beer.
• 6:50
  Little bit different, though. Anthropic did not respond to
• 6:53
  requests for comment. Because that's what they usually
• 6:58
  do. Big corporations do not respond
• 6:59
  because they just wanted to blow over.
• 7:01
  And the fact that multiple leading AI systems are pulling
• 7:04
  from an encyclopedia built by a competitor and built by an AI
• 7:09
  raises questions about how these comanies curate and filter their
• 7:12
  web search sources. Open AI told the Guardian that
• 7:15
  Chat GT aims to draw from a broad range of ublicly available
• 7:19
  sources and viewpoints, adding that safety filters reduce the
• 7:22
  risk of servicing links associated with high severity
• 7:26
  harms.