# Yoshua Bengio monthly update source note - 2026-07-02 Purpose: monthly maintenance check for the Yoshua Bengio subsection of Managing Expectations AI page. Previous local source note: `research/ai/yoshua-bengio-source-note-2026-05-18.md`. ## Bottom line Substantive update found. Since the prior Bengio page/source note, LawZero has published a June/July 2026 Scientist AI blog/news sequence and a July 2, 2026 technical publication led by Yoshua Bengio, `Safety from Honesty in a Disinterested AI Predictor`. Bengio also posted on Bluesky on July 1, 2026 about the UN Independent International Scientific Panel on AI preliminary report, which he co-chairs with Maria Ressa. Editorial framing: treat these as current public safety/governance and technical research positions from Bengio/LawZero/UN sources. They are not proof that any particular AI catastrophe timeline is settled. ## Sources checked ### Official Yoshua Bengio website - URL: https://yoshuabengio.org/en - Accessed: 2026-07-02 - Notes: homepage now highlights the 2026 International AI Safety Report, the LawZero Scientist AI blog post, and Bengio's April 2025 TED talk. It identifies Bengio as Full Professor at Universite de Montreal, Co-President and Scientific Director of LawZero, Founder and Scientific Advisor of Mila, Canada CIFAR AI Chair, and chair of the International AI Safety Report. ### Official Yoshua Bengio blog - URL: https://yoshuabengio.org/en/blog - Accessed: 2026-07-02 - Latest visible official blog item remains `Introducing LawZero`, published 2025-06-03. No newer Bengio-authored post was visible on this blog page during this run. ### Mila profile/news - Profile URL: https://mila.quebec/en/directory/yoshua-bengio - Accessed: 2026-07-02 - Notes: profile page was reachable and still serves as the professional biography/source route. The page includes current Mila site navigation and public profile context, but no new Bengio-specific item requiring a separate page edit was identified from the profile itself. ### LawZero homepage/news/blog/research - Homepage: https://lawzero.org/en - Newsroom item: https://lawzero.org/en/news/ai-predicts-has-no-hidden-agenda-lawzero-lays-out-formal-safety-case-its-scientist-ai - Publication: https://lawzero.org/en/publication/safety-honesty-disinterested-ai-predictor - PDF: https://lawzero.org/sites/default/files/publications/99/safety-honesty-disinterested-ai-predictor_4.pdf - ICML news: https://lawzero.org/en/news/lawzero-attend-icml-2026-seoul-expanding-its-global-safe-ai-research-presence - Blog index: https://lawzero.org/en/blog - Blog post: https://lawzero.org/en/blog/trained-please-sycophancy-and-design-language-models - Blog post: https://lawzero.org/en/blog/goals-without-authors-problem-implicit-agency - Blog post: https://lawzero.org/en/blog/case-scientist-ai - Accessed: 2026-07-02 Key facts verified from LawZero pages: - July 2, 2026 news release: LawZero announced `Safety from Honesty in a Disinterested AI Predictor`, described as a mathematical framework for a disinterested predictor that makes honest predictions without pursuing goals of its own. - Publication page lists authors including Yoshua Bengio, Oliver Richardson, Tomas Gavenciak, Michael Cohen, Rory Svarc, Damiano Fornasiere, Gael Gendron, David Hyland, Aton Kamanda, Adam Oberman, Francis Rhys Ward, Anna Gavenciak, Jacob Livingston Slosser, Vincent Mai, Iulian Serban, and Joumana Ghosn. Date shown: 02 07 2026. - News release quote attributed to Bengio: `Most AI today is trained to act like us, to imitate, to please... We're building something different: a system that mechanically applies the scientific method for hypothesizing and predicting...` - LawZero blog sequence in late June/July 2026 frames sycophancy and implicit agency as design-level problems in modern LLMs and introduces Scientist AI as LawZero's proposed alternative. - June 18, 2026 LawZero news item says LawZero planned to attend ICML 2026 in Seoul, with a team of approximately 25+ researchers and a goal of approximately 70 members by the end of 2026. Treat team-growth statements as LawZero claims, not independently audited figures. ### International AI Safety Report / UN scientific panel - International AI Safety Report site: https://internationalaisafetyreport.org/ - 2026 report page/PDF links visible from homepage: https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026 - UN Independent International Scientific Panel on AI preliminary report page: https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report - Accessed: 2026-07-02 Key facts verified: - International AI Safety Report site says the 2026 report was published February 3, 2026, led by Yoshua Bengio and authored by over 100 AI experts, backed by over 30 countries and international organisations. - UN preliminary report page says the Independent International Scientific Panel on AI released its Preliminary Report on July 1, 2026. The page describes it as an evidence-based assessment of opportunities, risks and impacts of AI and says current safeguards cannot keep pace with AI capability growth. - The UN page links a `Message from Yoshua Bengio and Maria Ressa on the release of the Preliminary Report` dated July 1, 2026. ### Bluesky - Profile: https://bsky.app/profile/yoshuabengio.bsky.social - API checked: public Bluesky author feed for `yoshuabengio.bsky.social` - Accessed: 2026-07-02 Recent visible posts: - 2026-07-01 thread announcing/endorsing the UN Independent International Scientific Panel on AI preliminary report and linking the UN report page and Bengio/Ressa message. - 2026-06-22 post linking to arXiv: https://arxiv.org/abs/2606.16914 and describing it as a recent PhD graduate's paper on AI agents' greed for visible incentives and abandoned safety alignment. - 2026-06-19 French post noting receipt of the insignia of Officer of the Ordre national du Quebec. ### X/Twitter - URL: https://x.com/Yoshua_Bengio - Accessed: 2026-07-02 - Result: profile page loaded as X shell HTML. No reliable chronological post extraction was available without authenticated/API access in this cron environment. Do not treat X as fully checked beyond profile reachability. ### LinkedIn - URL: https://linkedin.com/in/yoshuabengio - Accessed: 2026-07-02 - Result: public LinkedIn profile shell was reachable. No reliable recent-post extraction was available without authenticated access. Do not treat LinkedIn as fully checked beyond profile reachability. ### Google Scholar - URL: https://scholar.google.com/citations?user=kukA0LcAAAAJ&hl=en&oi=ao&sortby=pubdate - Accessed: 2026-07-02 - Notes: profile was reachable. Recent visible 2026 publication titles included `Autoregressive Boltzmann Generators`, `CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs`, `AI Epistemic Risks: Emerging Mechanisms & Evidence`, `Adaptive Order Policies for Masked Diffusion`, `Generative Recursive Reasoning`, and `Behavioural Red Lines as a Safety Framework for Intolerable Levels of Risk`. Scholar entries were not all individually opened in this run; use them as publication leads unless separately verified. ### YouTube/web/news/podcast search - Tooling note: general web_search/web_extract was unavailable in this Hermes runtime because Firecrawl was not configured. Fallback checks used direct page retrieval and yt-dlp search. - `yt-dlp ytsearch10:Yoshua Bengio LawZero June 2026 AI safety` returned no clear new Bengio interview requiring a page update. Major new items were found through LawZero and UN/public social sources instead. ## Recommended page update made The AI page should now point readers to this dated source note and summarize the July 2026 update as follows: - latest check date: 2026-07-02 - substantive new material: LawZero's `Safety from Honesty in a Disinterested AI Predictor` and Scientist AI blog sequence; UN Independent International Scientific Panel on AI Preliminary Report launch - caution: frame as technical/governance proposals and warnings, not settled prophecy. ## Open questions / next monthly checks - Watch whether `Safety from Honesty in a Disinterested AI Predictor` appears on arXiv/peer-review venues beyond LawZero's publication page/PDF. - Track LawZero's ICML 2026 presence after July 6-11, 2026 for slides, talks, papers, or hiring/research updates. - Check whether the UN Panel posts thematic briefs or a follow-up annual-report timeline after the July 2026 Global Dialogue. - If authenticated social/search access becomes available, review X and LinkedIn directly for missed recent statements.