“If we don’t know what is in the data sets, we can’t know what biases these systems will spread through society if they are widely used, we can’t know the extent to which individual content creatore are being exploited, and worse any outside efforts to mitigate risks are greatly undermined.”
Our reading, the sentence that was checked:
If we don't know what is in AI training data sets, we can't know what biases these systems will spread through society if widely used or the extent to which individual content creators are being exploited, and outside efforts to mitigate risks are greatly undermined.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Six months after the Pause letter, AI is still erratic — and still something we should still be deeply concerned about, paragraph 14.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-six-months-after-the-pause-letter-40c98f-a2
We recorded 9 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 2023-09-21. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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)
Conditions
- if we don't know what is in the data sets (antecedent, stated by the speaker)
- if they are widely used (qualifier, stated by the speaker)
From the speaker
- If we don’t know what is in the data sets
- we can’t know what biases these systems will spread through society
- if they are widely used
- the extent to which individual content creatore are being exploited
- any outside efforts to mitigate risks are greatly undermined
Our interpretation
- The ordinary reading is that training-data secrecy prevents reliable assessment of propagated biases and the extent of creator exploitation, and substantially weakens independent risk mitigation.
- Identifying particular biased outputs or particular training examples is relevant counterevidence, but does not by itself establish societal effects or the extent of exploitation.
- Creator exploitation includes uncompensated or unauthorized use and resulting harm; identifying dataset membership alone does not establish either legal infringement or economic harm.
- This is a general causal claim, so later research may test its mechanism. It is not scored as a forecast conditional on a particular deployment.
Counts as supported if: Independent audits or comparative studies show that withholding training-data information prevents the stated assessments and materially reduces outside mitigation effectiveness, with disclosure addressing those limitations.
Counts as contradicted if: Independent investigators reliably assess the stated societal biases and extent of creator exploitation without training-data disclosure, or comparisons show that nondisclosure does not substantially impair outside mitigation. Evidence settling only one component would warrant a component-level finding.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
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 May 29, 2024 paper describes how training-data access helps auditors trace problems and assess fair use [e1]. The September 17, 2024 study identifies linguistic discrimination in model responses through output-based evaluation [e2]. The February 21, 2026 study traces selected, marked training data through outputs under black-box access [e3]. These findings show benefits of data access and some investigation possible without disclosure, but do not isolate disclosure’s effect on assessing societal bias, the extent of creator exploitation, or outside mitigation effectiveness [e1–e3]. The full causal claim remains unresolved [e1–e3].
Limitations
- Output discrimination does not establish bias propagation through society under widespread use [e2].
- Tracing selected training examples does not establish the extent of unauthorized or uncompensated use or resulting harm to creators [e3].
- The evidence does not quantify how much training-data secrecy impairs outside mitigation or isolate it from other access restrictions [e1–e3].
Evidence (3 items)
Supports the claim: arXiv,
Access to data helps auditors trace problems and assess fair use.
Why it matters here: The paper identifies concrete audit benefits from access to training data, supporting the proposed mechanism. It does not establish that nondisclosure prevents all the claimed assessments or quantify a substantial effect on every outside mitigation effort.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv,
can perpetuate linguistic discrimination
Why it matters here: Researchers identified discrimination by prompting the models and evaluating their responses. This challenges the reading that no bias can be identified without dataset disclosure, but does not measure propagation through society under widespread deployment.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Nature Communications,
information isotopes
Why it matters here: The study demonstrates tracing selected, marked data through generated outputs without internal model access. This limits an absolute impossibility claim, but does not establish the full extent of creators’ uncompensated use or harm.
- Publisher
- Nature Communications
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
What would settle it
Independent comparative audits with and without training-data disclosure, holding access to weights, deployment records, and developer assistance constant, that measure reliable assessment of societal bias and creator exploitation and the effectiveness of outside mitigation.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Six months after the Pause letter, AI is still erratic — and still something we should still be deeply concerned about
If we don’t know what is in the data sets, we can’t know what biases these systems will spread through society if they are widely used, we can’t know the extent to which individual content creatore are being exploited, and worse any outside efforts to mitigate risks are greatly undermined.
- Where
- paragraph 14
- Published
- 2023-09-21
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-six-months-after-the-pause-letter-40c98f-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.
About assessment marcus-six-months-after-the-pause-letter-40c98f-a2.