Scores by kind of claim
Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.
-
Factual claims
Statements about the present or the past, checked against the record: did they hold up?
76 70 82 -
Predictions
Statements about events outside the person's control, checked once they fell due: did they come true?
60 42 78
All types combined: 74 (likely 68 to 80, high confidence), from 79 claims: 71 factual claims, 8 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.
Eliezer Yudkowsky's claims
46 claims
“Noting that all of these things are presently being trained on computers that are just like on the internet, which is like not a very smart life decision for us as a species.”
Eliezer Yudkowsky, interviewee 30 March 2023 Lex Fridman Podcast #368 Eliezer Yudkowsky transcript (Podscript)
As of March 2023, leading AI systems are being trained on computers that are connected to the internet.
Mostly true True in substance, wrong in a detail.
The second judge would not resolve this claim; the verdict shown is the first judge's.
What would count as true or false
Judge as of 30 March 2023. The claim is that leading AI systems were trained on computers that were connected to the internet rather than air-gapped. True: credible reporting, lab documentation or expert sources show that frontier labs (for example OpenAI, Google, Anthropic) trained their leading models on internet-connected cloud or datacenter infrastructure, with no air-gapping, as of then. Mostly true: this holds generally, with minor exceptions or some network isolation that falls short of air-gapping. Mixed: practices clearly varied across the leading labs, with roughly half using isolated setups. Mostly false: most leading training runs were done on air-gapped or strongly isolated clusters, and internet connectivity was only partial or indirect. False: evidence shows the leading systems were trained on air-gapped computers.
Evidence
- OpenAI says it will no longer train its AI models on the live internet, citing 'significant risks' | TechCrunch TechCrunch · 23 October 2026
“Neither Anthropic or OpenAI has gone this far, but simply building effective restraints to keep models off the internet has also been seen as impractical or likely to compromise the efficacy of training runs.”
- OpenAI says it will no longer train its AI models on the live internet, citing 'significant risks' | TechCrunch TechCrunch · 23 October 2026
“The decision to train these models in sandboxes cut off from the internet is notable because employees at frontier labs have insisted that doing so would be nearly infeasible.”
- How Microsoft's bet on Azure unlocked an AI revolution - Source Microsoft News (Source) · 17 March 2023
“The two companies began working in close collaboration to build supercomputing resources in Azure that were designed and dedicated to allow OpenAI to train an expanding suite of increasingly powerful AI models.”
- As OpenAI releases GPT-4, Microsoft details Azure AI infrastructure behind it DatacenterDynamics
“While a lot of details about the model are shrouded in mystery, what is known is that OpenAI uses the cloud service of investor Microsoft to train its systems.”
Reasoning
Contemporaneous reporting shows OpenAI trained its models, including the GPT-4 generation, on Microsoft's Azure cloud supercomputing infrastructure (E3, E4). Later TechCrunch reporting says OpenAI only in October 2026 decided to stop training on the live internet. It adds that frontier-lab employees had insisted that training in sandboxes cut off from the internet would be nearly infeasible, and that building restraints to keep models off the internet was seen as impractical (E1, E2). Together these indicate leading models were not trained on air-gapped systems, though no source directly describes the network isolation in place at every lab as of March 2023.
Cite this claim
Eliezer Yudkowsky, “As of March 2023, leading AI systems are being trained on computers that are connected to the internet.” (said 30 March 2023). Verdict: mostly true. Reliability API, claim c_5065b7943237, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_5065b7943237
Claims without a verdict
Is a verdict here wrong, or do you speak for Eliezer Yudkowsky? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.