“There is no intention to create misinformation—but also no capacity to avoid it, because fundamentally, GPT-3 is a model of how words relate to one another, not a model of how language might relate to the perceived world.”
Our reading, the sentence that was checked:
GPT-3 cannot avoid producing misinformation because it models how words relate to each other, not how language relates to the world.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Noam Chomsky and GPT-3, paragraph 12.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-noam-chomsky-and-gpt-3-f648bc-a1
We recorded 8 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2022-05-21. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- no intention to create misinformation
- but also no capacity to avoid it
- because fundamentally, GPT-3 is a model of how words relate to one another
- not a model of how language might relate to the perceived world
Our interpretation
- Two parts. (a) GPT-3 had no reliable built-in way to avoid producing false content. (b) The cause is that it is trained only on relations among words, without grounding in the perceived world.
- Read fairly, 'no capacity to avoid it' means it cannot reliably avoid misinformation, not that it never says anything true.
Counts as supported if: Controlled evidence that GPT-3-era models' misinformation stems from text-only training, and that grounding in the world is what would be needed to fix it.
Counts as contradicted if: Evidence that GPT-3-era models without world grounding could reliably avoid misinformation, or that the errors had a different main cause.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
The claim has an empirical part and a causal part. The evidence fits the empirical part read fairly. GPT-3 and InstructGPT kept making up information, at 41% and 21% on closed-domain tasks (e2). But fine-tuning without any grounding in the world halved that rate, which shows some capacity to reduce misinformation. TruthfulQA's authors tie false answers to imitating training text (e1), which partly fits Marcus's story. No cited study isolates the cause he names, being a model of word relations rather than of the world, from other factors. The causal part therefore stays open on the available evidence.
Limitations
- The evidence is about error rates, not about mechanism.
- Settling this would need a controlled grounded-versus-ungrounded comparison, which no source here provides.
Evidence (2 items)
Context only: ACL 2022 (Lin, Hilton, Evans), via arXiv,
However, this result is expected if false answers are learned from the training distribution.
Why it matters here: Links falsehoods to imitating training text, which partly fits Marcus's word-statistics story. It does not test whether grounding in the world is the missing factor.
- Where
- Abstract
- Publisher
- ACL 2022 (Lin, Hilton, Evans), via arXiv
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: arXiv (OpenAI authors),
InstructGPT models make up information not present in the input about half as often as GPT-3 (a 21% vs. 41% hallucination rate, respectively).
Why it matters here: Human-feedback fine-tuning, with no grounding in the world, halved fabrication but left it at 21%. This fits 'cannot reliably avoid' and also shows some capacity to reduce errors, so the causal story stays unsettled.
- Publisher
- arXiv (OpenAI authors)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Controlled studies comparing text-only and world-grounded models of similar scale on misinformation rates, or mechanistic work showing why GPT-3-era models fabricate.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Noam Chomsky and GPT-3
There is no intention to create misinformation—but also no capacity to avoid it, because fundamentally, GPT-3 is a model of how words relate to one another, not a model of how language might relate to the perceived world.
- Where
- paragraph 12
- Published
- 2022-05-21
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-noam-chomsky-and-gpt-3-f648bc-a1: Current. Outcome: Unresolved. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.