“This would be very bad news, not just in terms of human and economic costs (which would mean more underpaid layers doing awful work) but also in terms of trustworthiness, because it would be mean that we have zero guarantee that any new iteration of a large language model is going to safe.”
Our reading, the sentence that was checked:
If every LLM update needs full RL retraining, there would be more underpaid labeling work and zero guarantee that any new LLM iteration is safe.
Assessment: Unresolved No deadline stated on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Why *is* Bing so reckless?, paragraph 34.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-why-is-bing-so-reckless-6d2d51-a2
We recorded 12 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 2023-02-20. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Conditional forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No deadline stated
- Deadline state on 2026-10-29
- No deadline stated (the release’s reference date)
Conditions
- if every LLM update requires complete RL retraining (antecedent, stated by the speaker)
From the speaker
- Maybe every new update of a large language model will require a complete retraining of the reinforcement learning module.
- more underpaid layers doing awful work
- zero guarantee that any new iteration of a large language model is going to safe
Our interpretation
- This is a conditional prediction, not an assertion that complete retraining is already necessary.
- I read 'layers' as a typographical error referring to human labelers.
- The labor prediction means that repeated complete retraining would increase poorly paid human labeling involving distressing material.
- In context, the safety prediction means that the previous version's safety training would provide no reliable assurance of the new version's safety. It does not mean that every new version necessarily causes harm.
- The antecedent requires evidence of necessity across updates. Companies choosing to retrain some models would not establish it.
Counts as supported if: First establish that every LLM update requires complete RL retraining. Then establish both an attributable increase in poorly paid, distressing labeling work and the inability to carry reliable safety assurances across versions.
Counts as contradicted if: With the antecedent established, evidence that the required retraining does not produce the predicted additional labor burden, or that reliable safety assurances carry across new versions, would contradict the corresponding consequence. Failure to establish the antecedent is not a failed prediction.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
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 deadline stated
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
The December 15, 2022 Constitutional AI paper describes training experiments without human labels identifying harmful outputs [e1]. The November 2, 2024 Rule Based Rewards paper proposes preference modeling using AI feedback and a small amount of human data [e2]. These methods qualify the proposed link between retraining and additional human labeling [e1][e2]. Neither study establishes that every LLM update requires complete reinforcement-learning retraining, or settles the predicted labor and safety consequences under that condition [e1][e2].
Limitations
- The evidence does not establish the universal retraining requirement [e1][e2].
- The studies do not measure additional poorly paid, distressing labeling work attributable to that requirement [e1][e2].
- They do not establish or refute reliable safety assurances across versions under that condition [e1][e2].
Evidence (2 items)
Context only: arXiv; Yuntao Bai and colleagues,
without any human labels
Why it matters here: This experiment provides a material qualification to the proposed labor mechanism: harmful-output labels need not all come from humans. It establishes neither the necessity of complete retraining for every update nor the consequences of that condition.
- Publisher
- arXiv; Yuntao Bai and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Tong Mu and colleagues,
utilizes AI feedback
Why it matters here: The proposed method further qualifies the assumption that repeated safety training necessarily entails substantially more human labeling. It supplies no wage records or universal result about retraining requirements.
- Publisher
- arXiv; Tong Mu and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Training records and reproducible research establishing that every LLM update requires complete reinforcement-learning retraining, followed by linked labor records and cross-version safety evaluations establishing both predicted consequences or decisively refuting either consequence under that condition.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Why *is* Bing so reckless?
This would be very bad news, not just in terms of human and economic costs (which would mean more underpaid layers doing awful work) but also in terms of trustworthiness, because it would be mean that we have zero guarantee that any new iteration of a large language model is going to safe.
- Where
- paragraph 34
- Published
- 2023-02-20
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-why-is-bing-so-reckless-6d2d51-a2: 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.