1
00:00:02,960 --> 00:00:09,240
Open AI's GPT 5.2 model has 
started setting Grokapedia, Elon

2
00:00:09,240 --> 00:00:13,760
Musk's AI generated 
encyclopedia, as a source of its

3
00:00:13,760 --> 00:00:15,040
responses. 
Now. 

4
00:00:15,040 --> 00:00:18,320
The Guardian ran tests last week
and found ChatGPT reference 

5
00:00:18,320 --> 00:00:22,480
Grokapedia 9 times across more 
than a dozen user queries, 

6
00:00:22,720 --> 00:00:25,480
including questions about 
Iranian political structures and

7
00:00:25,480 --> 00:00:29,000
a British historian who helped 
convict A Holocaust denier. 

8
00:00:29,480 --> 00:00:32,200
The citations appeared 
selectively, surfacing an 

9
00:00:32,200 --> 00:00:35,680
obscure topics while avoiding 
well documented misinformation 

10
00:00:35,680 --> 00:00:38,160
subjects. 
Grocopedia launched just three 

11
00:00:38,160 --> 00:00:41,360
months ago with 0 human. 
Editors and critics have already

12
00:00:41,360 --> 00:00:44,340
flagged it for promoting 
debunked claims, setting neo 

13
00:00:44,340 --> 00:00:47,320
Nazi forums and whitewashing 
extremism. 

14
00:00:47,960 --> 00:00:51,960
Now, how did an AI encyclopedia 
built by one of Open AI's 

15
00:00:51,960 --> 00:00:58,120
biggest rivals end up feeding 
answers back into Open AI's most

16
00:00:58,120 --> 00:01:02,480
popular chatbot? 
Agracopedia now holds over 5.6 

17
00:01:02,480 --> 00:01:06,520
million articles, roughly 80 the
size of English Wikipedia, and 

18
00:01:06,760 --> 00:01:10,320
it accomplished that scale in a 
fraction of the time Wikipedia 

19
00:01:10,320 --> 00:01:12,280
needed. 
Now today we're going to cover 

20
00:01:12,280 --> 00:01:15,040
what the Guardians testing 
revealed with specific false 

21
00:01:15,040 --> 00:01:18,720
information chatt GPD pulled 
from Gakapedia, why other AI 

22
00:01:18,720 --> 00:01:22,680
assistants like Claude 
Anthropics Bot are also setting 

23
00:01:22,680 --> 00:01:25,400
the platform, and what this 
means for the integrity of AI 

24
00:01:25,400 --> 00:01:28,280
generated knowledge. 
Open AI has responded with a 

25
00:01:28,280 --> 00:01:32,280
statement about safety filters 
and source diversity, while XAI 

26
00:01:32,640 --> 00:01:36,600
simply replied with three words.
I'll tell you about those later,

27
00:01:37,000 --> 00:01:39,880
right after this short break 
now. 

28
00:01:39,880 --> 00:01:43,720
Open AI marketed GPD 5.2 as its 
most advanced model for 

29
00:01:43,720 --> 00:01:46,760
professional work when it 
launched in December, But the 

30
00:01:46,760 --> 00:01:49,800
Guardian's investigation raises 
serious questions about where 

31
00:01:49,800 --> 00:01:52,440
the professional grade 
information actually comes from.

32
00:01:52,840 --> 00:01:56,560
The tests revealed a clear 
pattern and how ChatGPT selects 

33
00:01:56,560 --> 00:02:00,720
Grokapedia as a source, and the 
implications cut deeper than a 

34
00:02:00,720 --> 00:02:04,680
simple citation problem. 
Users asking complex questions 

35
00:02:04,680 --> 00:02:08,960
about geopolitics and history 
are receiving answers built on 

36
00:02:08,960 --> 00:02:12,080
an AI generated foundation that 
has already been documented 

37
00:02:12,320 --> 00:02:15,680
spreading false claims. 
Now the guardians were testing 

38
00:02:15,680 --> 00:02:19,320
focused on a range of topics to 
see when ChatGPT would pull from

39
00:02:19,320 --> 00:02:22,880
Grokopedia versus other sources,
and the chatbot did not say 

40
00:02:22,880 --> 00:02:25,520
Grokopedia. 
When asked about widely known 

41
00:02:25,520 --> 00:02:29,000
misinformation, such as false 
claims about media bias against 

42
00:02:29,000 --> 00:02:32,960
Donald Trump or debunk theories 
about HIV and AIDS, ChatGPT 

43
00:02:32,960 --> 00:02:36,320
avoided the obvious traps. 
The system instead cited 

44
00:02:36,320 --> 00:02:39,720
Gracopedia on obscure 
geopolitical and historical 

45
00:02:39,720 --> 00:02:41,560
questions. 
The kind of query is where users

46
00:02:41,560 --> 00:02:45,120
are less likely to catch errors.
Specific topics included 

47
00:02:45,120 --> 00:02:51,280
salaries of Iran's paramilitary 
force, ownership of the mastas, 

48
00:02:51,280 --> 00:02:55,040
a fan fountain, and connections 
between the Iranian government 

49
00:02:55,040 --> 00:02:58,800
and telecom company MTN Iran 
Cell. 

50
00:02:59,360 --> 00:03:02,080
Now, Chachibti made bolder 
claims about those Iranian 

51
00:03:02,080 --> 00:03:03,920
government links than Wikipedia 
does. 

52
00:03:04,200 --> 00:03:09,760
Those claims came directly from 
Grokopedia, and the most 

53
00:03:09,760 --> 00:03:13,360
troubling example involved Sir 
Richard Evans, a Cambridge 

54
00:03:13,360 --> 00:03:16,840
historian who served as the key 
expert witness in the libel 

55
00:03:16,840 --> 00:03:19,640
trial wrought by Holocaust 
denier David Irving. 

56
00:03:20,640 --> 00:03:23,680
Evans personally examined his 
own Grokopedia entry and 

57
00:03:23,680 --> 00:03:25,800
reported that it contained 
numerous false statements about 

58
00:03:25,800 --> 00:03:28,840
his career and background. 
He told the Guardian that all 

59
00:03:28,840 --> 00:03:31,080
the facts in his entry were 
wrong. 

60
00:03:32,440 --> 00:03:34,200
The already runs pretty deep 
here. 

61
00:03:34,200 --> 00:03:38,360
Evans spent years meticulously 
documenting how Irving falsified

62
00:03:38,360 --> 00:03:42,040
evidence about the Holocaust and
manipulated historical records. 

63
00:03:42,440 --> 00:03:46,360
The High Court judge ultimately 
branded Irving a racist, an anti

64
00:03:46,360 --> 00:03:50,200
semite and falsifier of history 
based largely on Evans expert 

65
00:03:50,200 --> 00:03:53,120
testimony. 
ChatGPT cited that same 

66
00:03:53,120 --> 00:03:56,280
Gracopedia page when responding 
to user questions about Evans 

67
00:03:56,560 --> 00:04:00,080
effectively laundering debunked 
information through Open AI's 

68
00:04:00,080 --> 00:04:02,680
flagship product. 
The historian who exposed the 

69
00:04:02,680 --> 00:04:06,720
Holocaust deniers lies is now 
having lies spread about him 

70
00:04:06,960 --> 00:04:10,760
through AI systems that treat 
machine generated content as the

71
00:04:10,760 --> 00:04:13,360
equivalent of peer reviewed 
scholarship. 

72
00:04:14,200 --> 00:04:17,120
Now I've been going through all 
of our stats here at the show 

73
00:04:17,959 --> 00:04:23,360
and 39% of you are following the
channel and I'm so grateful for 

74
00:04:23,360 --> 00:04:29,240
you. 61% of you have listened to
the show before but haven't hit 

75
00:04:29,240 --> 00:04:32,360
the subscribe or follow button. 
If that's you, if you listen to 

76
00:04:32,360 --> 00:04:35,120
the show and you're into it, 
please take a second of your 

77
00:04:35,120 --> 00:04:36,800
time. 
You can hit the follow button. 

78
00:04:36,800 --> 00:04:39,480
I'm going to spend the next 10 
years doing this. 

79
00:04:39,760 --> 00:04:42,760
So that's what you get from me, 
better content over the next 10 

80
00:04:42,760 --> 00:04:44,400
years. 
All I need is a second of your 

81
00:04:44,400 --> 00:04:46,040
time. 
So thank you so much for doing 

82
00:04:46,040 --> 00:04:49,120
that. 
Agrocopedia launched on October 

83
00:04:49,120 --> 00:04:54,120
27th, 2025 as Elon Musk's answer
to what he calls Woke Pedia. 

84
00:04:54,400 --> 00:04:57,960
The platform uses XA is Grok 
large language model to generate

85
00:04:57,960 --> 00:05:00,360
articles rather than relying on 
human editors. 

86
00:05:00,760 --> 00:05:03,840
Users can suggest corrections 
through a pop up form, but they 

87
00:05:03,840 --> 00:05:06,720
can't directly edit content the 
way Wikipedia allows you to. 

88
00:05:07,520 --> 00:05:10,920
Musk position the project as a 
bias free alternative that would

89
00:05:10,920 --> 00:05:12,800
purge propaganda from public 
knowledge. 

90
00:05:13,400 --> 00:05:15,640
The reality has looked pretty 
different though. 

91
00:05:17,000 --> 00:05:19,760
Analysis found Wikipedia 
articles aligned with Musk's 

92
00:05:19,760 --> 00:05:22,600
personal views on topics like 
gender identity. 

93
00:05:22,840 --> 00:05:26,720
Tesla Controversies and 
Neuralink and the Guardian doing

94
00:05:26,720 --> 00:05:29,400
their research here, and 
multiple other outlets 

95
00:05:29,400 --> 00:05:32,560
documented pages that promote 
white nationalist talking 

96
00:05:32,560 --> 00:05:36,560
points, praise neo Nazis, cast 
Holocaust deniers in positive 

97
00:05:36,560 --> 00:05:39,760
terms, and frame far right 
figures as victims of 

98
00:05:39,840 --> 00:05:43,680
institutional suppression. 
The scale of Rockopedia's growth

99
00:05:43,760 --> 00:05:47,520
presents its own problems too. 
Platform reached 5.6 million 

100
00:05:47,520 --> 00:05:52,000
articles by early 2026, jumping 
from 800,000 at launch to nearly

101
00:05:52,000 --> 00:05:55,120
80% the volume of English 
Wikipedia in about 3 months. 

102
00:05:55,560 --> 00:05:59,760
Human editors and Wikipedia 
spent over 2 decades building 

103
00:05:59,760 --> 00:06:02,800
their corpus through careful 
research, citation and community

104
00:06:02,800 --> 00:06:05,080
debate. 
Grakopedia achieved similar 

105
00:06:05,080 --> 00:06:09,200
volume by having an AI system 
scrape existing content, rewrite

106
00:06:09,200 --> 00:06:12,640
those articles, and generate new
ones at machine speed. 

107
00:06:13,120 --> 00:06:16,880
Symmetry is copied Wikipedia 
near the word for word under a 

108
00:06:16,880 --> 00:06:19,960
Creative Commons license. 
While others introduced entirely

109
00:06:19,960 --> 00:06:23,200
new claims from questionable 
sources, researchers from 

110
00:06:23,200 --> 00:06:26,640
Grokopedia cited neo Nazi forums
and its reference lists. 

111
00:06:27,000 --> 00:06:30,080
Problem that becomes 
exponentially worse when other 

112
00:06:30,080 --> 00:06:34,040
AI systems start treating 
Grokopedia as a legit source. 

113
00:06:35,360 --> 00:06:39,440
Now, Jet GBT was not the only AI
assistant caught citing 

114
00:06:39,440 --> 00:06:41,920
Grokopedia. 
The Guardians testing found the 

115
00:06:41,920 --> 00:06:44,520
Anthropics. 
Claude, my favorite AI right 

116
00:06:44,520 --> 00:06:47,440
now, also referenced the 
platform when using questions 

117
00:06:47,440 --> 00:06:49,920
about global oil production and 
Scottish beer. 

118
00:06:50,640 --> 00:06:53,160
Little bit different, though. 
Anthropic did not respond to 

119
00:06:53,160 --> 00:06:58,000
requests for comment. 
Because that's what they usually

120
00:06:58,000 --> 00:06:59,880
do. 
Big corporations do not respond 

121
00:06:59,880 --> 00:07:01,400
because they just wanted to blow
over. 

122
00:07:01,840 --> 00:07:04,840
And the fact that multiple 
leading AI systems are pulling 

123
00:07:04,840 --> 00:07:08,840
from an encyclopedia built by a 
competitor and built by an AI 

124
00:07:09,240 --> 00:07:12,000
raises questions about how these
comanies curate and filter their

125
00:07:12,000 --> 00:07:15,400
web search sources. 
Open AI told the Guardian that 

126
00:07:15,400 --> 00:07:19,400
Chat GT aims to draw from a 
broad range of ublicly available

127
00:07:19,400 --> 00:07:22,440
sources and viewpoints, adding 
that safety filters reduce the 

128
00:07:22,440 --> 00:07:26,120
risk of servicing links 
associated with high severity 

129
00:07:26,240 --> 00:07:26,920
harms.