“Its output is more fluent, and lots of individual mistakes have been addressed.”
Our reading, the sentence that was checked:
Since 2019, GPT's output has become more fluent and many individual mistakes have been fixed.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Gaslighting and reality in AI, paragraph 29.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-gaslighting-and-reality-in-ai-fa52f0-a3
We recorded 5 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 2023-02-02. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- 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
- compared with GPT-2 in 2019 (qualifier, stated by the speaker)
From the speaker
- Its output is more fluent
- lots of individual mistakes have been addressed
- What’s changed since 2019?
Our interpretation
- The comparison is between GPT-2 in 2019 and its GPT successors available by February 2, 2023.
- The plain reading is smoother, more natural writing and a substantial reduction in errors, not elimination of all errors.
- Addressed is read as reduced or corrected in later outputs; it does not require proof that developers individually patched particular examples.
- Direct GPT-2/GPT-3 comparisons can establish improvements achieved within this interval without relying on later models.
Counts as supported if: Comparable evaluations show more natural writing and meaningful reductions in multiple error categories in GPT successors relative to GPT-2.
Counts as contradicted if: Comparable evaluations show no meaningful improvement in writing fluency or error incidence, or show systematic deterioration on those dimensions.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
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
A 2021 study comparing GPT-2 and GPT-3 found that untrained evaluators distinguished GPT-3 text from human writing at chance level, supporting improved naturalness [e1]. A separate 2021 comparison found reduced coverage of encyclopedic, commonsense and incoherence error annotations in GPT-3 relative to GPT-2 under matched sampling settings [e2]. These results support reductions in several kinds of mistakes within the claimed period [e2]. Off-prompt, math, and grammar and usage errors plateaued, so the gains were uneven [e2].
Limitations
- Humanlikeness is an indirect measure of fluency and does not establish factual correctness [e1].
- The error comparison concerns generated English news, and results depend on sampling settings [e2].
- These studies do not directly evaluate the January 30, 2023 ChatGPT update [e1, e2].
- Reduced error coverage does not establish that developers individually patched particular mistakes or permanently eliminated them [e2].
Evidence (2 items)
Supports the claim: arXiv; Elizabeth Clark and colleagues,
evaluators distinguished between GPT3- and human-authored text at random chance level.
Why it matters here: Under matched prompting, evaluators distinguished GPT-2 from human text more accurately than GPT-3 from human text across stories, news, and recipes. This supports improved naturalness, although humanlikeness is an indirect fluency measure.
- Publisher
- arXiv; Elizabeth Clark and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Yao Dou and colleagues,
These error categories decrease with in-domain training (Grover) and larger model size (GPT-3).
Why it matters here: Matched GPT-2/GPT-3 comparisons find reductions in encyclopedic, commonsense, and incoherence errors. Other categories plateau or behave less consistently, and generation settings affect the results.
- Publisher
- arXiv; Yao Dou and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Gaslighting and reality in AI
Its output is more fluent, and lots of individual mistakes have been addressed.
- Where
- paragraph 29
- Published
- 2023-02-02
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-gaslighting-and-reality-in-ai-fa52f0-a3: Current. Outcome: Supported. 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-gaslighting-and-reality-in-ai-fa52f0-a3.
HeldBy this site
The final-text check failed: citations_correct.
About assessment marcus-gaslighting-and-reality-in-ai-fa52f0-a2, which was not published.