“The only thing keeping it from being even more toxic and deceitful than it already is, is a system called Reinforcement Learning by Human Feedback, and “OpenAI” has been very closed about exactly how that really works.”
Our reading, the sentence that was checked:
The only thing keeping ChatGPT from being even more toxic and deceitful than it already is is RLHF, and OpenAI has been very closed about exactly how RLHF works.
Assessment: Contradicted Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Inside the Heart of ChatGPT’s Darkness, paragraph 31.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-inside-the-heart-of-chatgpts-darkness-1f22cd-a1
We recorded 11 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 2023-02-11. 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)
From the speaker
- The only thing keeping it from being even more toxic and deceitful
- Reinforcement Learning by Human Feedback
- OpenAI has been very closed about exactly how that really works
Our interpretation
- ChatGPT means the deployed product, including its content safeguards, rather than only the underlying model weights.
- RLHF includes the usual demonstration, preference-ranking, and reinforcement-learning pipeline. Supervised fine-tuning is not counted as a separate safeguard merely to manufacture a contradiction.
- The exclusivity claim requires RLHF to be the sole effective protection against additional harmful behavior.
- Closed about exactly how is read as withholding ChatGPT-specific implementation details, not as saying no general RLHF explanation existed.
- The relevant implementation is the publicly deployed system from launch through the post.
Counts as supported if: Evidence establishes that only RLHF restrained the deployed system's harmful behavior and that ChatGPT-specific implementation details were substantially undisclosed.
Counts as contradicted if: Documentation establishes an additional deployed protective mechanism, or sufficiently complete ChatGPT-specific disclosures contradict the secrecy claim. A contradiction of exclusivity does not by itself disprove limited disclosure.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
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
The claim of exclusivity is contradicted on the deployed-product reading. Contemporary reporting with OpenAI confirmation identifies toxicity detection and data filtering, while the launch documentation describes separate moderation. General RLHF methods were also public. This does not disprove the narrower complaint that exact ChatGPT-specific implementation details remained undisclosed.
Limitations
- This rejects the assertion that RLHF was the only safeguard; it does not reject every part of the disclosure criticism.
- Documentation establishes additional protective mechanisms, not their comparative effect sizes.
- A reading confined to the underlying model's learned behavior would require further causal evidence.
- The company confirmation is available through original news reporting rather than a standalone published statement.
Evidence (3 items)
Goes against the claim: TIME,
remove toxic data
Why it matters here: OpenAI's confirmation describes a toxicity detector incorporated into ChatGPT and efforts to filter training material. These are protections beyond reinforcement learning; the report does not quantify their separate effects.
- Publisher
- TIME
- Kind of source
- News report
- Describes events of
- Public from
Context only: OpenAI,
Proximal Policy Optimization
Why it matters here: OpenAI described the feedback-training pipeline and separately stated that its Moderation API warned or blocked unsafe content. The page establishes public method disclosure and corroborates additional safeguards, but is not a controlled efficacy comparison.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: OpenAI researchers via arXiv,
human preferences
Why it matters here: The paper explains the predecessor's demonstration, ranking, reward-model, and reinforcement-learning procedure. It counters a reading that the general method was secret, but cannot establish complete disclosure or causal effects for ChatGPT itself.
- Publisher
- OpenAI researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Contemporaneous ChatGPT ablations would establish each mechanism's causal contribution. A dated inventory of implementation disclosures would settle the remaining disclosure question.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Inside the Heart of ChatGPT’s Darkness
The only thing keeping it from being even more toxic and deceitful than it already is, is a system called Reinforcement Learning by Human Feedback, and “OpenAI” has been very closed about exactly how that really works.
- Where
- paragraph 31
- Published
- 2023-02-11
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-inside-the-heart-of-chatgpts-darkness-1f22cd-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 marcus-inside-the-heart-of-chatgpts-darkness-1f22cd-a1.