“In our opinion, this was inevitable, in a system that has been trained to optimize the goal of producing one plausible-sounding word after another rather than actually engage with the meaning of language.”
Our reading, the sentence that was checked:
ChatGPT's errors were inevitable because it is trained to produce plausible next words rather than engage with meaning.
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 3.
- 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-ed0205-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)
From the speaker
- In our opinion
- this was inevitable
- trained to optimize the goal
- producing one plausible-sounding word after another
- rather than actually engage with the meaning of language
Our interpretation
- The plain empirical reading is that the training objective made the preceding error families unavoidable, rather than merely increasing their likelihood.
- The introductory opinion label does not remove the causal commitment made by 'inevitable'.
- The assessment concerns the cause of launch-era ChatGPT's errors. Later research may address the mechanism but cannot automatically establish what caused this particular system's failures.
- I do not equate predicting the next token with a demonstrated absence of semantic representations.
Counts as supported if: A causal experiment or applicable theoretical result establishes that ChatGPT's actual training process necessarily produces the identified failures because it optimizes plausible continuation without the relevant semantic processing.
Counts as contradicted if: An applicable causal result or counterexample establishes that the same relevant objective can avoid the identified failures, or that their occurrence depends on another factor rather than the asserted necessary mechanism.
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
Observed mistakes do not establish that the proposed cause made them inevitable. ChatGPT's documented training included human-feedback stages, and the relevant theoretical research proves a narrower necessity result under additional assumptions. That research does not extend its statistical argument to arithmetic. These findings limit the explanation without conclusively disproving the causal claim about the launch system.
Limitations
- The claim combines a causal assertion with an undefined account of meaning.
- A theoretical result about a class of models is not automatically a causal diagnosis of ChatGPT.
- Evidence of increased error risk would not establish inevitability.
Evidence (2 items)
Context only: OpenAI,
reward models
Why it matters here: The methods section describes human demonstrations, ranked responses and reinforcement learning. This complicates an account based solely on plausible next-word pretraining, but does not establish whether the resulting system engages with meaning or whether its errors were avoidable.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: arXiv; Adam Tauman Kalai and Santosh S. Vempala,
may mitigate
Why it matters here: The paper proves a necessity result for arbitrary facts under a calibration condition, while expressly excluding systematic facts such as arithmetic from that particular statistical argument. It neither establishes the full causal claim nor proves that launch-era ChatGPT could avoid every listed failure.
- Publisher
- arXiv; Adam Tauman Kalai and Santosh S. Vempala
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
An operational definition of semantic engagement and an applicable causal experiment or theorem establishing whether the specified errors necessarily follow from the launch system's training process.
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
In our opinion, this was inevitable, in a system that has been trained to optimize the goal of producing one plausible-sounding word after another rather than actually engage with the meaning of language.
- Where
- paragraph 3
- Published
- 2023-01-09
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-large-language-models-like-chatgpt-ed0205-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-ed0205-a1.