“What these systems do, no more and no less, is to put together sequences of words, but without any coherent understanding of the world behind them, like foreign language Scrabble players who use English words as point-scoring tools, without any clue about what that mean.”
Our reading, the sentence that was checked:
LLMs only string together sequences of words and have no coherent understanding of the world behind them.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, New adventures in AI hype: “Our language models are so ‘conscious’ we need to give them rights”, paragraph 7.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-new-adventures-in-ai-hype-our-language-cd38ec-a1
We recorded 7 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-04-25. 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
- argument from 2022 that Marcus says still applies to today's LLMs (qualifier, stated by the speaker)
From the speaker
- no more and no less
- put together sequences of words
- without any coherent understanding of the world behind them
- applies as much to today’s LLM’s as those of three years ago
Our interpretation
- The plain reading is that these models lack meaningful, coherent representations of situations and relationships, even when their language sounds convincing.
- Understanding is assessed through generalization, consistency and interventions on internal representations. Fluent answers alone do not establish it.
- The context explicitly connects June 2022 systems with those available on April 25, 2025. Later systems cannot settle this historical claim.
- Understanding is distinct from sentience. Evidence about one does not establish the other.
Counts as supported if: Representative evaluations establish that the relevant models rely on surface associations without coherent representations supporting generalization across related situations.
Counts as contradicted if: Reproducible evaluations and causal interventions demonstrate coherent, generalizable representations of situations and relationships in a model covered by the claim.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
Research documents both meaningful intermediate representations and failures of consistency. Anthropic's interventions challenge a literal claim that models merely associate words, while independent evaluations support doubts about coherent generalization. Neither settles the broader assertion about understanding without choosing a threshold Marcus did not supply. These are competing pieces of evidence about one disputed property, not established true and false components warranting a mixed score.
Limitations
- The positive mechanistic evidence comes from the model developer and has acknowledged methodological limits.
- The independent study covers particular tasks and models.
- No conclusion about sentience follows.
Evidence (2 items)
Supports the claim: Keyon Vafa and colleagues, arXiv,
Such incoherence creates fragility
Why it matters here: The experiments support concerns about coherence and transfer, including in general-purpose LLMs tested on logic puzzles. They do not establish that every relevant model lacks all meaningful understanding.
- Publisher
- Keyon Vafa and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: Anthropic,
combining independent facts
Why it matters here: Researchers identified intermediate geographic concepts and changed the answer by intervening on those concepts. This challenges the strongest word-association-only reading, but the authors acknowledge incomplete coverage and possible artifacts in their method.
- Where
- Multi-step reasoning
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
Replicated tests of models available by the post date that combine novel-task generalization, consistency under interventions and validated internal representations, with an explicit standard for coherent understanding.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, New adventures in AI hype: “Our language models are so ‘conscious’ we need to give them rights”
What these systems do, no more and no less, is to put together sequences of words, but without any coherent understanding of the world behind them, like foreign language Scrabble players who use English words as point-scoring tools, without any clue about what that mean.
- Where
- paragraph 7
- Published
- 2025-04-25
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-new-adventures-in-ai-hype-our-language-cd38ec-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-new-adventures-in-ai-hype-our-language-cd38ec-a1.