“Human children are constantly learning new things, often from tiny amounts of data. There is literally nothing like that in the AI world.”
Our reading, the sentence that was checked:
No current AI system can learn new things from tiny amounts of data the way human children constantly do.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Three baffling claims about AI and machine learning in four days, statistical errors in top journals, and claims from Yann LeCun that you should not believe., paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-too-much-benefit-of-the-doubt-4cae0d-a1
We recorded 13 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 2022-10-16. 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
- Human children are constantly learning new things
- often from tiny amounts of data
- literally nothing like that in the AI world
Our interpretation
- The surrounding examples make the ordinary reading broad, continuing acquisition of unfamiliar skills from little new experience, comparable to a child.
- This is not automatically a claim that no narrow one-shot or few-shot learning method exists. The literal wording nevertheless makes such results material counterevidence.
- Prior learning is allowed for both children and AI. A large pretraining history alone does not disqualify a demonstration of learning from a few new examples.
- For this ongoing capability observation, I use the current reporting month through the post, October 1–16, 2022. Earlier experiments are contextual unless a source establishes the relevant October state.
Counts as supported if: Broad evaluations of systems existing in the target period establish a substantial gap from children's continuing, data-efficient learning across diverse unfamiliar tasks.
Counts as contradicted if: A system existing in the target period demonstrates continuing learning across diverse unfamiliar language, reasoning, and practical tasks from comparably small amounts of new experience.
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
Gato and GPT-3 make the phrase literally nothing like that too sweeping if it means no analogous few-shot learning. The surrounding passage, however, describes a child's broad and continuing ability to acquire unfamiliar skills. These papers do not demonstrate that full capability, and their limitations do not establish its absence across all systems. The contextual claim remains unresolved on the retrieved evidence.
Limitations
- The interpretation materially affects the result.
- Few-shot performance and persistent learning are different measurements.
- Neither source provides a matched comparison with a five-year-old across the stated abilities.
Evidence (2 items)
Context only: arXiv,
task learning
Why it matters here: Gato provides a concrete analogy to learning new tasks from limited data, including adaptation in robotics. Its reported experiments do not demonstrate the full breadth and continuing learning of a child, and precede the target snapshot.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv,
strong few-shot performance
Why it matters here: The paper reports both strong few-shot results and weaknesses in inference and reading-comprehension tasks. This challenges a blanket denial of analogous learning, but is not a matched test of broad, continuing childlike learning in October 2022.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A historical evaluation of October 2022 systems against children on diverse unfamiliar tasks, reporting new examples, prior experience, transfer, and retention. A clear narrower interpretation would permit a separate assessment of the literal few-shot-learning claim.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Three baffling claims about AI and machine learning in four days, statistical errors in top journals, and claims from Yann LeCun that you should not believe.
Human children are constantly learning new things, often from tiny amounts of data. There is literally nothing like that in the AI world.
- Where
- paragraph 15
- Published
- 2022-10-16
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-too-much-benefit-of-the-doubt-4cae0d-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-too-much-benefit-of-the-doubt-4cae0d-a1.