Pause development of frontier systems that could cause large-scale harm until independent safety proof; prohibit superintelligence until scientific safety consensus and strong public buy-in, while accelerating LawZero’s Scientist AI safety and capability research. [1][2][3][4][5][6][7]
Editorial correction. Reassess verified recommendations using Methodology v2 best-fit headlines, preserving the earlier interpretation and narrower evidence.
Source [1] · Original statement
Professor Yoshua Bengio’s Speech at the UN Security Council
Attributed to: Yoshua Bengio · Yoshua Bengio
Published
“Developers must demonstrate to independent experts that a system is safe to train and safe to deploy.”
Paragraph beginning “We must agree collectively that no one, no company and no country” · View original source ↗
Source context. Authored published UN Security Council speech. He says no company or country should develop potentially large-scale harmful AI without safety proof; this is a development gate, not merely risk concern.
Scope
AI systems that could cause large-scale harm, with training as well as deployment gated by safety.
Conditions
Independent experts; licensing, incident reporting, liability and insurance; support safe controllable AI.
Stated rationale
Interpret only the verified passage in its stated scope; do not infer pace from identity, optimism, risk concern or deployment policy.
Source details
primary · Own site publication September23; separate delivery date not independently established, use publication date.
Discovered 3 Oct 2026.
Published 23 Sep 2026.
Accessed 3 Oct 2026.
Source [2] · Original statement
Slowing down development of AI systems passing the Turing test
Attributed to: Yoshua Bengio · Yoshua Bengio
Published
“…it seems obvious to me that we must accept a certain slowdown in technological progress to ensure greater safety and protect the collective well-being.”
Answer to “Should we stop development of AI systems altogether?” · View original source ↗
Source context. Own essay explicitly discusses signing the six-month letter, then denies calling for stopping all AI systems. Historical pause support must not be erased.
Scope
Systems beyond GPT4 covered by the open letter, not AI at large.
Conditions
Protect collective well-being; fund beneficial AI and establish governance.
Stated rationale
Interpret only the verified passage in its stated scope; do not infer pace from identity, optimism, risk concern or deployment policy.
Source details
primary · Own dated essay; publication used because no separate statement event date established.
Discovered 3 Oct 2026.
Published 5 Apr 2023.
Accessed 3 Oct 2026.
Source [3] · Original statement
Professor Yoshua Bengio’s Speech at the UN Security Council
Attributed to: Yoshua Bengio · Yoshua Bengio
Published
“…offer no convincing technical solutions”
Opening third paragraph, after current leading developers acknowledge catastrophic risks · View original source ↗
Source context. Authored published UN Security Council speech. He says no company or country should develop potentially large-scale harmful AI without safety proof; this is a development gate, not merely risk concern.
Scope
Current most powerful systems developed by leading AI companies, in the context of a prerequisite for safe training.
Conditions
Contextual observation in the September 23 published speech; not a separate prohibition on all AI.
Stated rationale
The same speech identifies presently dangerous agents and inadequate safety solutions before requiring independent proof of training and deployment safety.
Source details
primary · Own site publication September23; separate delivery date not independently established, use publication date.
Discovered 3 Oct 2026.
Published 23 Sep 2026.
Accessed 3 Oct 2026.
Source [4] · Original statement
Statement on Superintelligence
Attributed to: Yoshua Bengio · Future of Life Institute
“We call for a prohibition on the development of superintelligence, not lifted before there is”
Statement heading, immediately before the two numbered reopening conditions · View original source ↗
Source context. Original statement reopened with both reopening conditions intact. FLI’s separate original press release names Bengio an initial signatory; exact personal signing day is unknown.
Scope
Superintelligence that significantly outperforms all humans on essentially all cognitive tasks; not AI research at large.
Conditions
Reopen only after broad scientific consensus on safe, controllable development and strong public buy-in.
Stated rationale
Original statement reopened with both reopening conditions intact. FLI’s separate original press release names Bengio an initial signatory; exact personal signing day is unknown.
Source details
primary · Original source reopened on October 6, 2026; attribution and surrounding conditions checked directly.
Discovered 6 Oct 2026.
Accessed 6 Oct 2026.
Source [5] · Original statement
FLI original release identifies Bengio as an initial signatory
Attributed to: Yoshua Bengio · Future of Life Institute
Published
“Initial signatories include AI pioneers Yoshua Bengio and Geoffrey Hinton”
Source context. Host page displays March 27, 2026 publication. Other FLI listing and external reporting have different launch dates, so this source is used only for directly verified publication and endorsement, never a claimed exact signing day.
Scope
Bengio’s endorsement of the conditional superintelligence prohibition.
Conditions
The accompanying statement permits lifting the prohibition only after scientific safety consensus and public buy-in; publication date is not his signing date.
Stated rationale
Host page displays March 27, 2026 publication. Other FLI listing and external reporting have different launch dates, so this source is used only for directly verified publication and endorsement, never a claimed exact signing day.
Source details
primary · Original source reopened on October 6, 2026; attribution and surrounding conditions checked directly.
Discovered 6 Oct 2026.
Published 27 Mar 2026.
Accessed 6 Oct 2026.
Source [6] · Original statement
What if automating AI R&D triggers an intelligence explosion?
Attributed to: Yoshua Bengio · Alan Chan and coauthors, including Yoshua Bengio
Published
“…policymakers should develop ways to pace and constrain scale-ups of automated AI R&D.”
Page 6, Steering and constraining an intelligence explosion, first policy paragraph · View original source ↗
Source context. Original paper lists Bengio as a coauthor; arXiv v1 submission September 28. The surrounding section distinguishes preparing options and considering intervention choices from imposing every option now.
Scope
Automated AI research and development scale-ups.
Conditions
Menu includes safety/stakeholder requirements, possible capability-growth limits, compliance tools and options for specific-workload pauses.
Stated rationale
Original paper lists Bengio as a coauthor; arXiv v1 submission September 28. The surrounding section distinguishes preparing options and considering intervention choices from imposing every option now.
Source details
primary · Original source reopened on October 6, 2026; attribution and surrounding conditions checked directly.
Discovered 6 Oct 2026.
Published 28 Sep 2026.
Accessed 6 Oct 2026.
Source [7] · Original interview
Yoshua Bengio thinks he knows how to build safe superintelligence
Attributed to: Yoshua Bengio · 80,000 Hours
Statement
“…what we need is to accelerate that effort, so that’s a lot of engineering.”
Why companies can’t stop racing [00:36:35], answer following “what can LawZero do in coming months?” · View original source ↗
Source context. Original interviewer transcript, recorded April16 and published May7. Bengio asks for faster engineering of LawZero’s Scientist AI; surrounding discussion covers an initial guardrail and later capable agents, so the program includes capability development rather than only audits.
Scope
LawZero’s safety-oriented Scientist AI research and capability-development program.
Conditions
Demonstrate the proposed safety architecture; this is not an endorsement of racing untrusted frontier agents or lifting the conditional superintelligence prohibition.
Stated rationale
The proposed approach aims to retain capability while improving honesty and controllability.
Source details
primary · Original interviewer transcript reopened October6; exact quote, recording date and publication date checked separately.
Discovered 6 Oct 2026.
Published 7 May 2026.
Accessed 6 Oct 2026.
Editorial assessment. Full-context source reopening resolves the earlier Unclear proposal. The September 23 speech describes leading systems as posing catastrophic risks without convincing technical solutions, then makes independent proof a prerequisite for training as well as deployment. The omitted original conditional-superintelligence statement independently establishes an endorsed PAUSE, with both reopening conditions preserved. The September 28 coauthored R&D paper offers governance options rather than an operative general halt or repudiation. The historical six-month GPT-4 pause remains historical; beneficial safe research is outside the harmful-development gate. Independent review found the omitted April16 Scientist AI interview. Its affirmative faster research program is preserved as a company-scoped ACCELERATE exception; the current dangerous-development and conditional ASI gates remain the defensible PAUSE best fit, without implying all AI research should halt.
Assessment reviewed
Superseded editorial interpretation ·
PACE
Require independent proof that potentially large-scale harmful AI is safe to train and deploy; support a controlled development path, not stopping all AI research. [1][2]
Initial statement. Research draft; independent adversarial review and human editorial approval remain separate.
Source [1] · Original statement
Professor Yoshua Bengio’s Speech at the UN Security Council
Attributed to: Yoshua Bengio · Yoshua Bengio
Published
“Developers must demonstrate to independent experts that a system is safe to train and safe to deploy.”
Paragraph beginning “We must agree collectively that no one, no company and no country” · View original source ↗
Source context. Authored published UN Security Council speech. He says no company or country should develop potentially large-scale harmful AI without safety proof; this is a development gate, not merely risk concern.
Scope
AI systems that could cause large-scale harm, with training as well as deployment gated by safety.
Conditions
Independent experts; licensing, incident reporting, liability and insurance; support safe controllable AI.
Stated rationale
Interpret only the verified passage in its stated scope; do not infer pace from identity, optimism, risk concern or deployment policy.
Source details
primary · Own site publication September23; separate delivery date not independently established, use publication date.
Discovered 3 Oct 2026.
Published 23 Sep 2026.
Accessed 3 Oct 2026.
Source [2] · Original statement
Slowing down development of AI systems passing the Turing test
Attributed to: Yoshua Bengio · Yoshua Bengio
Published
“…it seems obvious to me that we must accept a certain slowdown in technological progress to ensure greater safety and protect the collective well-being.”
Answer to “Should we stop development of AI systems altogether?” · View original source ↗
Source context. Own essay explicitly discusses signing the six-month letter, then denies calling for stopping all AI systems. Historical pause support must not be erased.
Scope
Systems beyond GPT4 covered by the open letter, not AI at large.
Conditions
Protect collective well-being; fund beneficial AI and establish governance.
Stated rationale
Interpret only the verified passage in its stated scope; do not infer pace from identity, optimism, risk concern or deployment policy.
Source details
primary · Own dated essay; publication used because no separate statement event date established.
Discovered 3 Oct 2026.
Published 5 Apr 2023.
Accessed 3 Oct 2026.
Editorial assessment. April2023 own essay supports a slowdown of systems beyond GPT4 and the six-month open letter; explicitly denies stopping AI altogether. September2026 safety licensing does not demonstrate repudiation of all historical pause support. LawZero builds safety-oriented AI, not an unrestricted pace endorsement. Unresolved: None. No human publication approval.
Assessment reviewed
Earlier position ·
PAUSE
Supported the six-month letter for systems more powerful than GPT-4; explicitly excluded AI research at large and promoted beneficial applications. [1]
Scope distinction. Supported the six-month letter for systems more powerful than GPT-4; explicitly excluded AI research at large and promoted beneficial applications.
Source [1] · Original statement
Slowing down development of AI systems passing the Turing test
Attributed to: Yoshua Bengio · Yoshua Bengio
Published
“…it seems obvious to me that we must accept a certain slowdown in technological progress to ensure greater safety and protect the collective well-being.”
Answer to “Should we stop development of AI systems altogether?” · View original source ↗
Source context. Own essay explicitly discusses signing the six-month letter, then denies calling for stopping all AI systems. Historical pause support must not be erased.
Scope
Systems beyond GPT4 covered by the open letter, not AI at large.
Conditions
Protect collective well-being; fund beneficial AI and establish governance.
Stated rationale
Interpret only the verified passage in its stated scope; do not infer pace from identity, optimism, risk concern or deployment policy.
Source details
primary · Own dated essay; publication used because no separate statement event date established.
Discovered 3 Oct 2026.
Published 5 Apr 2023.
Accessed 3 Oct 2026.
Editorial assessment. Historical evidence retained separately. Current assessment is a later scoped statement, not an invented express repudiation.