Pause building, or at least deploying, incorrigible generative AI until reliability and technical and political safeguards are ensured; reliable AI may continue, and he rejects a permanent blanket ban. [1][2][3]
Initial statement. Pause building unreliable, incorrigible generative AI until technical and political safeguards ensure trustworthiness; reliable AI may continue, and he rejects a permanent blanket ban.
Source [1] · Essay
The Case for Boycotting Generative AI
Attributed to: Gary Marcus · Gary Marcus / Marcus on AI
Statement
“Until such time as we have better answers to that both technically and politically, we should stop building such things, and stop giving them access to the internet.”
Paragraph beginning Until such time, following the definition of incorrigible AI. · View original source ↗
Source context. Marcus explicitly calls for pausing construction or at least deployment, then directly says stop building until better technical and political answers. He rejects permanent bans on AI or superintelligence and allows reliable deterministic approaches. He describes a recent revision from an OpenAI-specific suspension to a broader response to incorrigible AI. Embedded Hubinger, Pachocki, Thompson and Newport quotations are separate speakers.
Scope
Incorrigible generative AI that cannot be trusted or aligned to human values, particularly probabilistic LLM-based systems with internet access.
Conditions
Pause building, or at least deployment, until technical and political answers establish reliability and trustworthiness; permit reliable alternatives and lift the ban if frontier labs get their house in order.
Stated rationale
Marcus explicitly calls for pausing construction or at least deployment, then directly says stop building until better technical and political answers. He rejects permanent bans on AI or superintelligence and allows reliable deterministic approaches. He describes a recent revision from an OpenAI-specific suspension to a broader response to incorrigible AI. Embedded Hubinger, Pachocki, Thompson and Newport quotations are separate speakers.
Source details
primary · Original authored essay visibly dated September 9, 2026; written statement and publication date coincide. Independently inspected complete original, including earlier skeptical view of extinction, China coordination and final paragraph acknowledging his change of mind. This is conditional PAUSE, not permanent STOP and not a ban on all AI.
Discovered 4 Oct 2026.
Published 9 Sep 2026.
Accessed 4 Oct 2026.
Source [2] · Essay
The new Sanders-Casar Ban Artificial Superintelligence Act – and why I oppose it
Attributed to: Gary Marcus · Gary Marcus / Marcus on AI
Statement
“…a permanent, unilateral ban on all research into superhuman AI is too broad”
Source context. His opposition is to permanence, unilateral scope and all-research breadth. He expressly says a temporary pause may be needed and endorses a safety-conditioned alternative quoted from Aguirre. Sanders and Curran embedded tweets are not Marcus statements.
Scope
Permanent unilateral prohibition of all research into superhuman AI.
Conditions
Opposes a permanent unilateral prohibition; allows a temporary pause, potentially years, until improved regulation and alignment; favors independent demonstration of control for stronger systems.
Stated rationale
His opposition is to permanence, unilateral scope and all-research breadth. He expressly says a temporary pause may be needed and endorses a safety-conditioned alternative quoted from Aguirre. Sanders and Curran embedded tweets are not Marcus statements.
Source details
primary · Original authored essay visibly dated September 3, 2026. Deliberate contrary evidence: opposition to this bill does not contradict the later conditional suspension. Quoted Aguirre mechanism is attributed separately in surrounding notes.
Discovered 4 Oct 2026.
Published 3 Sep 2026.
Accessed 4 Oct 2026.
Source [3] · Essay
BREAKING: Florida seeks injunction against OpenAI
Attributed to: Gary Marcus · Gary Marcus / Marcus on AI
Statement
“This is a lot like the pause on OpenAI that I have been recommending here for weeks.”
Paragraph following embedded Florida injunction report, before link to Pause OpenAI, now. · View original source ↗
Source context. Marcus personally endorses an injunction and urges other jurisdictions to follow Florida. He also welcomes a potentially useful Nvidia safety platform while reserving judgment. The report of Florida legal action is contextual, not treated here as independently verified legal fact.
Scope
An injunction against OpenAI/ChatGPT development amid security and control concerns.
Conditions
Company-specific intervention until proper regulation and control; this short passage does not restate the full restart mechanism or rescind his broader September 9 stance.
Stated rationale
Marcus personally endorses an injunction and urges other jurisdictions to follow Florida. He also welcomes a potentially useful Nvidia safety platform while reserving judgment. The report of Florida legal action is contextual, not treated here as independently verified legal fact.
Source details
primary · Original authored essay visibly dated September 28, 2026. Newest located personal statement relevant to pace as of October 4. Positive reception of safety tooling does not affirm unrestricted frontier capability growth.
Discovered 4 Oct 2026.
Published 28 Sep 2026.
Accessed 4 Oct 2026.
Editorial assessment. Current candidate rests on Marcus own September 9 written recommendation; September 28 reinforces an OpenAI-specific pause without narrowing or explicitly withdrawing the broader statement. September 3 opposition to a permanent unilateral ban is retained. Searches across his Substack September–October, pause/ban, Anthropic, Sanders-Casar and newest posts found no later reversal. Preserve the explicit building-or-at-least-deploying hedge. Historical 2023 letter signature alone is not the current basis. No public change or human approval.