Why ChatGPT can’t be trusted with breaking news
We recorded 7 claims from this post. 2 were checked against evidence. 5 can’t be checked. 6 more are held. Read the post at its source.
Claims from this post (7)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“The unreliability of LLMs in the face of novelty is one of the core reasons why businesses shouldn’t trust LLMs”
Our reading: Businesses should not trust LLMs, in large part because LLMs are unreliable in the face of novelty.
No horizon
Qualifier: because of unreliability in the face of novelty
Assessment: Can’t be checkedSaid
Observation
“Pure LLMs are inevitably stuck in the past, tied to when they are trained, and deeply limited in their inherent abilities to reason, search the web, “think” critically etc.”
Our reading: Pure LLMs are inevitably stuck in the past of their training and deeply limited in inherent abilities to reason, search the web, and think critically.
No horizon Evidence as of
Limitation: The cited study concerns GPT-3 in 2020, not the capabilities of LLMs generally in January 2026 [e1].
Assessment: UnresolvedSaid
Observation
“Of course, the LLM companies have armies of humans; in my own brief experiments a couple hours later, OpenAI had apparently already patched this particular problem.”
Our reading: LLM companies have armies of humans, and when Marcus ran his own brief experiments a couple of hours later, OpenAI had apparently already patched this particular problem.
No horizon Evidence as of
Limitation: Private experiments may have occurred without a public record.
Assessment: UnresolvedSaid
Causal claim
“But nobody should be surprised by these errors; they emanate from the same dynamics as hallucinations: pure LLMs correlate words in their training, but lack stable, revisable world models.”
Our reading: Breaking-news errors in LLMs arise, like hallucinations, because pure LLMs correlate words but lack stable, revisable world models.
No horizon Evidence as of
Limitation: The relevant studies use different models and tasks from the reported news incident.
Assessment: Can’t be checkedSaid
Forecast
“(As I wrote here recently, the far more likely outcome is a tie.)”
Our reading: The US-China generative AI race will far more likely end in a tie than in either side dominating.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“But the idea of using them to plot strategy in rapidly-changing environments like war is laughable.”
Our reading: The idea of using LLMs to plot strategy in rapidly-changing environments like war is laughable.
No horizon
Scope: rapidly-changing environments like war
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“Relying on systems that can barely keep up with the publicly reported news, let alone stuff where human correction may be a lot harder to come by, is hardly the way to win a war.”
Our reading: Relying on systems that can barely keep up with publicly reported news, let alone matters where human correction may be a lot harder to come by, is hardly the way to win a war.
No horizon
Qualifier: let alone stuff where human correction may be a lot harder to come by
Held from this post (6)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The two reviews read the claim differently
- The criteria are ambiguous. Our reading, not assessed: GenAI companies will endlessly play catch-up, using many humans to apply undisclosed patches to their LLMs as each new problem arises.
- The two reviews read the claim differently
- The record does not meet the current rules. Our reading, not assessed: Neither the US nor China has any real chance of completely dominating generative AI.
- The criteria are ambiguous. Our reading, not assessed: Capabilities beyond word correlation are bolted onto LLMs with variable reliability.
- The two reviews reached different assessments. Our reading, not assessed: ChatGPT's denial of Maduro's capture was a fabrication, and the underlying LLM cannot on its own consult reputable news organizations that had already reported the story.