“It autocompletes sentences, some of those completions are true given that the training data is so vast, and some aren’t, given that the training data aren’t infinite.”
Our reading, the sentence that was checked:
GPT autocompletes sentences; some outputs are true because training data is vast and some are false because training data is not infinite.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, A Tale of Two Liars, paragraph 6.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-a-tale-of-two-liars-356862-a1
We recorded 3 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 2023-11-22. 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
- It autocompletes sentences
- some of those completions are true
- the training data is so vast
- some aren’t
- the training data aren’t infinite
Our interpretation
- Autocompletes sentences means generating text by predicting successive tokens.
- Vast but finite training data is read as broad but incomplete factual coverage.
- The claim attributes some correct and incorrect outputs to that coverage. It does not expressly say that data coverage is the only cause or that infinite data would guarantee truth.
- The relevant systems are GPT models available when the post appeared.
Counts as supported if: Model documentation confirms successive-token prediction, and evidence establishes that training-data coverage causally contributes to both correct answers and factual errors in the relevant GPT systems.
Counts as contradicted if: Documentation refutes the generation mechanism, or controlled evidence refutes the proposed contribution of training-data coverage. Other causes of errors do not alone contradict this nonexclusive claim.
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 autocomplete component is supported. Research also makes incomplete factual coverage a credible contributor to errors. However, the controlled coverage experiment examined another model, while the GPT-3 analysis used estimated training-data counts. Later theoretical work supplies context rather than a direct test of the relevant GPT systems. The complete causal claim therefore remains unresolved. Evidence of other error mechanisms would not, by itself, contradict his statement that coverage explains some outputs.
Limitations
- This assessment does not interpret finite data as the sole asserted cause.
- Training-data coverage, model scale, and total data volume are different variables.
- Theoretical results and experiments on other models cannot fully establish the GPT-specific attribution.
Evidence (3 items)
Supports the claim: OpenAI, arXiv,
predict the next token
Why it matters here: This supports the next-token prediction component. It does not establish why particular completions are true or false.
- Where
- Abstract
- Publisher
- OpenAI, arXiv
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Proceedings of Machine Learning Research,
reduce the dependence
Why it matters here: The study supports a role for factual coverage, including a controlled removal experiment on a separate model. Its GPT-3 results use estimated training-data counts, so they do not directly isolate the claimed cause in GPT.
- Publisher
- Proceedings of Machine Learning Research
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Nature,
reward unwarranted guessing.
Why it matters here: The paper explains how sparse factual support can produce errors and how evaluation incentives can preserve guessing. This supports the plausibility of a coverage mechanism without directly establishing its contribution in the GPT systems discussed in 2023.
- Where
- Abstract
- Publisher
- Nature
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
What would settle it
GPT-specific controlled experiments that vary relevant factual coverage while holding model size, training objectives, prompting, and evaluation constant.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, A Tale of Two Liars
It autocompletes sentences, some of those completions are true given that the training data is so vast, and some aren’t, given that the training data aren’t infinite.
- Where
- paragraph 6
- Published
- 2023-11-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-a-tale-of-two-liars-356862-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.