“Software like LaMDA simply doesn’t; it doesn’t even try to connect to the world at large, it just tries to be the best version of autocomplete it can be, by predicting what words best fit a given context.”
Our reading:
Language models like LaMDA do not connect their representations to the world; they are just autocomplete that predicts which words fit a context.
Assessment: Unresolved
What happened
Was it true for the period described?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
The evidence does not settle whether the named systems connect language to real-world referents as of 25 April 2025 [e1–e3]. The LaMDA paper, available on 10 February 2022, documents next-token pretraining; this establishes its training objective, not the absence of semantic representations [e1]. The July 2020 position paper argues that training on linguistic form cannot yield meaning, but does not empirically test the named systems [e2]. Anthropic's 27 March 2025 report describes interventions on Claude's geographical representations that changed its answers, challenging a simple word-association account without independently establishing semantic grounding [e3]. LaMDA's source-attribution measures also do not settle whether its design seeks semantic reference, so the assertion that these systems make no such effort remains unresolved [e1].
Caveats
- The evidence supplies no decisive test separating semantic reference from sophisticated linguistic associations [e1–e3].
- The Claude findings come from its developer, cover selected tasks, and use a method acknowledged to capture only part of the computation and potentially introduce artifacts [e3].
- The evidence does not establish the full capabilities or design of the unspecified Claude version intended by the claim.
Evidence (3 items)
Context only: arXiv; Google authors,
LaMDA was pre-trained to predict the next token in a text corpus.
Why it matters: This establishes the prediction objective, not the absence of semantic representations. The paper also describes fine-tuning and external tools, but its source-based groundedness does not by itself settle Marcus's semantic claim.
- Publisher
- arXiv; Google authors
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Association for Computational Linguistics,
trained only on form
Why it matters: This provides a substantive argument consistent with Marcus's distinction between word sequences and meaning. It predates the named systems and is not an empirical demonstration that they lack semantic understanding; availability is conservatively recorded at the end of the proceedings' dated conference.
- Publisher
- Association for Computational Linguistics
- Kind of source
- Peer-reviewed paper
- About events on
- Published
Context only: Anthropic,
combining independent facts
Why it matters: The reported interventions on Claude's geographical representations challenge a simple word-association account. They do not independently establish semantic grounding, and Anthropic acknowledges that its method captures only part of the computation and can introduce artifacts.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Where we had to interpret
- The ordinary reading in context concerns semantic understanding: whether these systems relate language to the things and situations it describes. It is not merely a claim that they cannot retrieve external text.
- The autocomplete description asserts a limitation on what the systems learn and do, not merely the use of a next-token prediction objective. That objective alone cannot establish the asserted limitation.
- Evidence of semantic grounding need not establish consciousness or require human-like sensory experience. Conversely, retrieving sources or producing fluent answers does not by itself establish semantic understanding.
- I assess the ongoing assertion as of the post date, 2025-04-25. There is no regular reporting period for this property. The June 2022 passage supplies historical context; earlier studies are retained as context rather than assigned new event dates.
- The words 'doesn’t even try' concern whether the system's design seeks semantic connection. The developer's use of the label 'groundedness' is insufficient unless its operational meaning matches that connection.
Right if: Evidence about the named systems establishes that their apparent reference to things and situations is confined to associations among linguistic forms, without representations that track and are used to reason about those referents. Design evidence must also support the asserted absence of an effort toward semantic connection. A prediction objective or isolated failures alone would not establish this.
Wrong if: Reproducible behavioral and mechanistic evidence establishes that a named system uses representations tied to real-world referents or situations across changes in wording and context, addressing explanations based only on surface associations. For the design assertion, evidence must show an effort toward semantic reference, rather than merely source attribution. Retrieval alone does not meet either test.
We wrote this test after the outcome was known.
What would settle it
Controlled behavioral and mechanistic studies of the relevant versions showing whether representations track independently specified real-world referents and causally support reasoning across changes in wording and context. These studies must distinguish semantic reference from linguistic association. Design documentation must separately establish whether the systems pursue semantic connection beyond source attribution; evidence supporting categorical absence must go beyond isolated failures.
We found 8 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, New adventures in AI hype: “Our language models are so ‘conscious’ we need to give them rights”, paragraph 9.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2025-04-25.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, New adventures in AI hype: “Our language models are so ‘conscious’ we need to give them rights”
Software like LaMDA simply doesn’t; it doesn’t even try to connect to the world at large, it just tries to be the best version of autocomplete it can be, by predicting what words best fit a given context.
- Where
- paragraph 9
- Published
- 2025-04-25
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.