“If I can be blunt, it is an absolute embarrassment that a technology that has collectively cost about half a trillion dollars can’t do something as basic as (reliably) check its output against wikipedia or a CNN article that is handed on a silver plattter.”
Our reading, the sentence that was checked:
LLM technology, having cost about half a trillion dollars, cannot reliably check its output against Wikipedia or a supplied article.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Why DO large language models hallucinate?, paragraph 23.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-why-do-large-language-models-hallucinate-14b4ca-a1
We recorded 16 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 2025-05-05. 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)
From the speaker
- technology that has collectively cost about half a trillion dollars
- can’t do something as basic as (reliably) check its output
- wikipedia or a CNN article
- handed on a silver plattter
Our interpretation
- Assess the spending and reliability assertions together; exclude the judgement that the situation is embarrassing.
- Read the money figure as approximately USD 500 billion in cumulative costs attributable to LLM technology, not funding raised, valuations or future commitments.
- The surrounding post describes roughly eight years of expenditure. Use 2017 as the approximate starting year, with January 1 only as an inclusive year boundary, and end at the post date.
- For the ongoing capability assertion, use the latest pre-post evidence, including the April 2025 report linked by Marcus. Older capability studies are context.
- Read reliable checking as dependable detection or avoidance of factual mismatches against supplied sources, not perfect performance on every possible input. No numerical reliability threshold is imposed.
Counts as supported if: A defined, non-duplicative accounting of cumulative LLM costs supports approximately USD 500 billion by the post, and contemporary evaluations show material failures when checking against available or supplied sources.
Counts as contradicted if: Comparable expenditure records place cumulative costs clearly outside a defensible approximation to USD 500 billion, or representative contemporary evaluations establish dependable source checking under the claimed conditions.
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
The archived CNN benchmark supports concern about source fidelity: its best reported model still had a 15% error rate. That does not establish a general inability to check outputs, and research on explicit fact-checking supplies a qualification. The USD 500 billion assertion remains unverified. Available investment and data-centre figures measure different things, and the claim supplies no accounting basis. The compound claim therefore cannot be scored supported or contradicted.
Limitations
- The money component has an unknown accounting basis.
- The 60-question benchmark is small and uses an automated checking system.
- A grounding error rate is not automatically a self-verification failure rate.
- The cumulative cost window is approximate; contemporary capability is assessed using pre-post reports.
Evidence (4 items)
Context only: AIMultiple,
Our benchmark revealed that OpenAI GPT-4.5 has the lowest hallucination rate (i.e. highest accuracy rate) of 15%.
Why it matters here: The archived report describes 60 CNN-based questions per model and a best reported error rate of 15%, supporting the narrower concern about source fidelity. It tests answering from supplied material rather than a separate self-verification procedure and cannot settle the spending assertion.
- Where
- Results; Methodology
- Publisher
- AIMultiple
- Kind of source
- Independent measurement
- Describes events of
- Public from
Context only: Tang, Laban and Durrett, arXiv,
Training on this data teaches models to check each fact in the claim and recognize synthesis of information across sentences.
Why it matters here: MiniCheck directly studies checking generated claims against grounding documents and reports models reaching GPT-4-level performance. This qualifies a reading of complete incapability, but does not establish a reliability threshold or the performance of all systems in April 2025.
- Where
- Abstract, version 2
- Publisher
- Tang, Laban and Durrett, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Stanford HAI, AI Index,
Private investment in generative AI reached $33.9 billion in 2024, up 18.7% from 2023 and over 8.5 times higher than 2022 levels.
Why it matters here: This measures private investment in generative AI during one calendar year. It neither supports nor contradicts cumulative expenditure on LLM technology.
- Publisher
- Stanford HAI, AI Index
- Kind of source
- Independent measurement
- Measures
- Generative AI sector, Annual private investment received, USD billions, estimate, 2024-01-01 to 2024-12-31
- Describes events of
- Public from
Context only: International Energy Agency; European Commission-hosted copy,
Investment is surging in new data centres and the capital-intensive IT equipment used for training and running AI models.
Why it matters here: The section estimates annual data-centre investment exceeding USD 500 billion in 2024. Its infrastructure scope and annual period differ from cumulative LLM costs, so the similar number does not validate Marcus's figure.
- Publisher
- International Energy Agency; European Commission-hosted copy
- Kind of source
- Official statistics
- Measures
- Global data-centre sector, Annual data-centre capital investment, USD billions, estimate, 2024-01-01 to 2024-12-31
- Describes events of
- Public from
What would settle it
A documented cumulative LLM expenditure estimate with clear scope and accounting, together with representative pre-post tests of explicit checking against supplied sources under a stated reliability criterion.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Why DO large language models hallucinate?
If I can be blunt, it is an absolute embarrassment that a technology that has collectively cost about half a trillion dollars can’t do something as basic as (reliably) check its output against wikipedia or a CNN article that is handed on a silver plattter.
- Where
- paragraph 23
- Published
- 2025-05-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-why-do-large-language-models-hallucinate-14b4ca-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-why-do-large-language-models-hallucinate-14b4ca-a1.