“It is therefore impossible for external scientists to determine to what extent any specific correct answer merely reflects some very similar examples in the training data.”
Our reading, the sentence that was checked:
Because ChatGPT's training corpus has not been made available to the public, it is impossible for external scientists to determine to what extent any specific correct answer merely reflects some very similar examples in the training data.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Large Language Models like ChatGPT say The Darnedest Things, paragraph 13.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-large-language-models-like-chatgpt-9f6100-a1
We recorded 10 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 2023-01-09. 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)
Conditions
- training corpus not made available to the general public (qualifier, stated by the speaker)
- external scientists (scope, stated by the speaker)
From the speaker
- the training corpus is not been made available to the general public
- therefore impossible for external scientists
- determine to what extent any specific correct answer
- merely reflects some very similar examples in the training data
Our interpretation
- The plain reading is that lack of corpus access prevents outside researchers from establishing how much a particular correct ChatGPT answer depends on similar training examples.
- Determine means a scientifically justified finding, not absolute certainty; merely detecting a familiar topic is insufficient.
- The present-tense claim concerns the ChatGPT system and research situation at publication. It is not treated as a forecast that no future method could ever work.
- The January 9, 2023 snapshot is used for this ongoing causal assertion.
Counts as supported if: Evidence specific to the then-current ChatGPT system establishes both unavailable corpus access and a methodological barrier preventing answer-specific training-dependence assessment, rather than merely documenting difficulty.
Counts as contradicted if: A validated method applicable to the then-current ChatGPT system establishes that a particular correct answer depends on similar training examples without access to OpenAI's corpus. Later results must demonstrably apply to that system and period to be decisive.
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 research found illustrates both the attribution problem and ways outsiders can detect memorization. Neither settles the historical causal claim as written. The book study explicitly limits its membership conclusions. Later extraction experiments challenge a timeless impossibility reading, but they do not establish what could be determined about the January 2023 system. Lack of corpus access alone is insufficient proof of impossibility.
Limitations
- A timeless reading would give the later extraction evidence greater weight.
- Finding reproduced text is different from measuring the cause of an arbitrary correct answer.
- Both cited studies concern later research and are context rather than decisive outcome evidence.
Evidence (2 items)
Context only: arXiv; Kent K. Chang and colleagues,
fundamentally unknowable outside of OpenAI.
Why it matters here: The authors infer familiarity with books while acknowledging that their method cannot verify actual training membership. This illustrates the distinction Marcus raises, but does not establish a January 2023 impossibility result.
- Publisher
- arXiv; Kent K. Chang and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Milad Nasr, Nicholas Carlini, and collaborating researchers,
we extract training data from ChatGPT.
Why it matters here: Researchers identified long ChatGPT outputs matching public text without possessing OpenAI's corpus. This challenges an unrestricted claim that outsiders can never identify memorization, but concerns later experiments and does not settle attribution for the January system.
- Publisher
- Milad Nasr, Nicholas Carlini, and collaborating researchers
- Kind of source
- Independent measurement
- Describes events of
- Public from
What would settle it
A validated attribution or extraction study tied to the January 2023 ChatGPT system, or a rigorous result establishing the claimed identification barrier under its actual access conditions.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Large Language Models like ChatGPT say The Darnedest Things
It is therefore impossible for external scientists to determine to what extent any specific correct answer merely reflects some very similar examples in the training data.
- Where
- paragraph 13
- Published
- 2023-01-09
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-large-language-models-like-chatgpt-9f6100-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.
About assessment marcus-large-language-models-like-chatgpt-9f6100-a1.