“These models are not "learning" or, let's be honest, "training" on this data, because that's not how they work — they're using mathematics to plagiarize it based on the likelihood that somebody else's answer is the correct one.”
Our reading, the sentence that was checked:
LLMs do not learn from data; they plagiarize it mathematically, going by the likelihood that someone else's answer is correct.
Assessment: Contradicted Evidence as of
Where and when Ed Zitron wrote it
Where's Your Ed At, They're Looting The Internet, paragraph 42.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment wyea-the-great-looting-of-the-internet-8e8846-a1
We recorded 21 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 2024-04-15. 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
- not "learning"
- not "training" on this data
- that's not how they work
- using mathematics to plagiarize it
- based on the likelihood that somebody else's answer is the correct one
Our interpretation
- I assess the technical account of models developed and available before 2024-04-15, including their preceding training and documented behavior.
- Learning means fitting statistical structure from data that can support responses beyond verbatim reproduction; it does not require consciousness or human-like understanding.
- The categorical contrast is read as claiming that copying others' answers describes the operating mechanism instead of statistical learning.
- The ethical characterization of training as plagiarism is not independently scored as a technical fact.
- There is no periodic reporting interval for this mechanism; the historical interval covers the pre-post models and their training, without an invented start date.
Counts as supported if: Technical documentation and controlled tests show that the relevant models operate through the described answer-copying mechanism and do not learn structure that supports generalization.
Counts as contradicted if: Technical records establish parameter training for predictive learning rather than selection of an existing answer by its probability of correctness, with empirical evidence of capabilities beyond verbatim reproduction.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Contradicted
The dated evidence goes against what the claim committed to, under the criteria frozen before grading.
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
As a technical description, the claim is contradicted. GPT-4's documented training predicts subsequent tokens; it does not select somebody else's answer according to its probability of being correct. Controlled analogy tests also provide evidence of learned performance beyond verbatim reproduction. Memorization and extraction of training text are real, and broader claims of human-like reasoning face material challenges. Neither qualification establishes that models do not statistically learn.
Limitations
- The result uses the ordinary machine-learning meaning of learning.
- It does not establish consciousness, human understanding, or reliable reasoning.
- It does not deny memorization, unattributed copying, or possible copyright infringement.
- The analogy findings are qualified by published counterexamples.
Evidence (4 items)
Goes against the claim: OpenAI, arXiv,
predict the next token
Why it matters here: The documented objective is predictive model training. A probability over next tokens is not a probability that somebody else's complete answer is correct.
- Where
- Abstract
- Publisher
- OpenAI, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: Taylor Webb, Keith J. Holyoak, and Hongjing Lu, arXiv,
GPT-3 matched or exceeded human performance for zero-shot Digit Matrices.
Why it matters here: The researchers constructed new text-based matrix problems and tested responses without task examples. Their findings support learned capabilities beyond simply reproducing a stored answer, although they do not establish unrestricted or human-like reasoning.
- Publisher
- Taylor Webb, Keith J. Holyoak, and Hongjing Lu, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Nicholas Carlini and coauthors, arXiv,
memorize and leak individual training examples
Why it matters here: Controlled extraction attacks recovered verbatim training material from GPT-2. This substantiates copying risks, but memorization can coexist with statistical learning and does not establish the exclusive mechanism asserted.
- Publisher
- Nicholas Carlini and coauthors, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Damian Hodel and Jevin West, arXiv,
simplest variations of letter string problems
Why it matters here: This response challenges broad claims of human-like analogical reasoning using modified letter-string tasks. It limits generalization from e2, without showing that next-token models do not learn or that all outputs are copied answers.
- Publisher
- Damian Hodel and Jevin West, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Where's Your Ed At, They're Looting The Internet
These models are not "learning" or, let's be honest, "training" on this data, because that's not how they work — they're using mathematics to plagiarize it based on the likelihood that somebody else's answer is the correct one.
- Where
- paragraph 42
- Published
- 2024-04-15
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment wyea-the-great-looting-of-the-internet-8e8846-a1: Current. Outcome: Contradicted. 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 wyea-the-great-looting-of-the-internet-8e8846-a1.