# Yoshua Bengio monthly update source note - 2026-09-02

Purpose: monthly maintenance check for the Yoshua Bengio subsection of the Managing Expectations AI page.

Previous local update note: `research/ai/yoshua-bengio-update-2026-08-02.md`.

## Bottom line

Substantive update found. Since the Aug. 2, 2026 check, the main new material is not a new Bengio-authored official-site blog post, but it is enough to keep the public AI subsection current:

- Bengio's public Bluesky feed posted a Sept. 1, 2026 link to a Guardian long-read on AI deception, summarizing his view that misaligned behaviours can emerge from reinforcement-learning-style training and that LawZero is rethinking how AI systems are trained.
- Bengio's Bluesky feed also posted Aug. 18, 2026 about a new SciencePetr interview/video on AI risks and current training techniques.
- Bengio amplified Bill Gates's late-August AI-transition essay, agreeing that frontier-AI risks include cyber, bio and loss-of-control concerns and saying society needs stronger preparation.
- LawZero's public research page now surfaces the technical publication `Language Models Recognize Dropout and Gaussian Noise Applied to Their Activations`, co-authored by Yoshua Bengio and Oliver Richardson as supervisors. It is a technical safety/interpretability lead about models detecting and localizing activation perturbations, with possible implications for evaluation/training-awareness work.
- Google Scholar's recent-publication list now surfaces additional August 2026 AI-safety/governance leads, including `How to Verify Consistency of Probabilistic Claims`, `The 2026 Singapore Consensus on Global AI Safety Research Priorities`, and `Open technical problems in open-weight AI model risk management`.

Editorial framing: these are current safety/governance signals, public warnings and technical-research leads. They should be framed as Bengio continuing to argue for safer training, honesty/coherence, governance and preparation - not as settled proof of any exact AI catastrophe timeline.

## Sources checked

### Official Yoshua Bengio website

- URL: https://yoshuabengio.org/en
- Accessed: 2026-09-02
- Result: reachable. The homepage still identifies Bengio as Full Professor of Computer Science at Universite de Montreal, Co-President and Scientific Director of LawZero, Founder and Scientific Advisor of Mila, Canada CIFAR AI Chair, 2018 A.M. Turing Award recipient, and chair of the International AI Safety Report.
- Visible homepage features remained the 2026 International AI Safety Report, the Scientist AI blog post/link, the April 2025 TED talk, and the June 2025 `Introducing LawZero` blog item.

### Official Yoshua Bengio blog

- URL: https://yoshuabengio.org/en/blog
- Accessed: 2026-09-02
- Result: reachable. Latest visible official blog item remained `Introducing LawZero`, published 03 June 2025. No new Bengio-authored official-site blog post was found.

### Mila profile/news

- Profile URL: https://mila.quebec/en/directory/yoshua-bengio
- News example rechecked: https://mila.quebec/en/news/milas-founder-yoshua-bengio-co-chairs-first-un-scientific-panel-report-on-ai
- Accessed: 2026-09-02
- Result: profile reachable and still useful as a professional biography/source route.
- Mila's July 7 story says Bengio co-chaired the first Preliminary Report of the UN Independent International Scientific Panel on AI with Maria Ressa. It describes the preliminary report as a standing evidence mechanism and says one central finding is that current safeguards cannot keep pace with AI capability growth.
- Newer Mila news surfaced in August, but the visible items checked were not specifically new Bengio subsection updates.

### LawZero homepage/newsroom/research

- Homepage: https://lawzero.org/en
- Newsroom: https://lawzero.org/en/newsroom
- Research: https://lawzero.org/en/research
- Publication: https://lawzero.org/en/publication/language-models-recognize-dropout-and-gaussian-noise-applied-their-activations
- Accessed: 2026-09-02

Key facts verified:

- LawZero's research page says the Scientist AI approach was conceived by Yoshua Bengio and is meant as a safety-centered path toward advanced AI.
- Newsroom still lists `LawZero advances safe-by-design AI with support from NVIDIA` and the July 2 Scientist AI safety-case release as the latest press releases; no newer press-release item was found.
- The research page and publication page now surface `Language Models Recognize Dropout and Gaussian Noise Applied to Their Activations`, by Damiano Fornasiere, Mirko Bronzi, Spencer Kitts, Alessandro Palmas, Yoshua Bengio and Oliver Richardson. The page says tested Llama, Olmo and Qwen models could detect/localize dropout or Gaussian-noise perturbations, and discusses a possible data-agnostic `training awareness` signal and AI-safety implications.

### International AI Safety Report / UN AI Panel

- International AI Safety Report homepage: https://internationalaisafetyreport.org/
- Publications page: https://internationalaisafetyreport.org/publications
- UN preliminary report page: https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report
- Accessed: 2026-09-02
- Result: reachable. The International AI Safety Report site still lists the Feb. 3, 2026 report and summaries as the latest main publications.
- The UN preliminary report page says the report is an independent scientific assessment across seven domains and states its central warning: current safeguards cannot keep pace with AI capability growth. It says the panel will continue its evidence base through consultations, scientific engagement and thematic briefs, with the next annual report informing the second Global Dialogue in May 2027.

### Bluesky

- Profile: https://bsky.app/profile/yoshuabengio.bsky.social
- API checked: public Bluesky author feed via `https://public.api.bsky.app/xrpc/app.bsky.feed.getAuthorFeed?actor=yoshuabengio.bsky.social&limit=20`
- Accessed: 2026-09-02

Recent visible posts/leads since the Aug. 2 update:

- 2026-09-01: Bengio linked a Guardian long-read, `If you build something vastly smarter than you, it better be on your side: can we stop AI from deceiving us?`, and said he explained why misaligned behaviours emerge from reinforcement learning, why they will continue to pose risks as models become more capable, and how LawZero intends to rethink training.
- 2026-08-27: Bengio amplified Bill Gates's `The turbulent AI era is here. The choices we make now are critical`, saying Gates highlighted growing AI risks including cyber, bio and loss of control, and that society needs better preparation for frontier models.
- 2026-08-25: Bengio posted that he discussed AI impacts/risks and LawZero's work with the Governor General of Canada during a visit to LawZero.
- 2026-08-18: Bengio linked the SciencePetr YouTube video/interview `AI Is Already Creating Problems. Here's Why.` and said the discussion covered growing AI risks and how they often stem from current AI training techniques.
- 2026-08-13: Bengio posted a thread on motivated reasoning/internal incoherence in AI agents and said this risk motivates LawZero's Scientist AI approach where honesty and internal coherence of beliefs are central to training.
- 2026-08-05: Bengio posted a thread arguing that current frontier models trained with reinforcement learning can amplify misalignment as capability increases and that developers should be legally responsible for model damage.

### X/Twitter

- URL: https://x.com/Yoshua_Bengio
- Accessed: 2026-09-02
- Result: public profile URL was checked as a target, but reliable chronological post extraction was not available in this unattended environment without authenticated/API access. Bluesky provided reliable public-feed extraction for this run.

### LinkedIn

- URL: https://linkedin.com/in/yoshuabengio
- Accessed: 2026-09-02
- Result: public LinkedIn content was treated as partially blocked/login-walled. No reliable recent-post extraction was available without authenticated access.

### Google Scholar

- URL: https://scholar.google.com/citations?user=kukA0LcAAAAJ&hl=en&oi=ao&view_op=list_works&sortby=pubdate
- Accessed: 2026-09-02
- Result: profile reachable with a browser-like request. Recent visible 2026 publication leads included:
  - `Adsorption energies are necessary but not sufficient to identify good catalysts` - ACS Catalysis, 2026.
  - `How to Verify Consistency of Probabilistic Claims` - arXiv preprint arXiv:2608.11181, 2026.
  - `Bayesian symbolic regression with entropic reinforcement learning` - arXiv preprint arXiv:2608.09617, 2026.
  - `Insuring AI: Incentivising Safe and Secure Deployment` - Harvard Data Science Review, 2026.
  - `The Future of the AI Summit Series: One Year Later` - Oxford Martin School AI Governance Initiative policy brief, 2026.
  - `Realigning AI technology towards the Sustainable Development Goals` - Nature Machine Intelligence, 2026.
  - `The 2026 Singapore Consensus on Global AI Safety Research Priorities` - arXiv preprint arXiv:2608.14611, 2026.
  - `Open technical problems in open-weight AI model risk management` - arXiv preprint arXiv:2608.07514, 2026.
  - `Safety from Honesty in a Disinterested AI Predictor` - arXiv preprint arXiv:2606.29657, 2026.
- Caution: Scholar entries were used as publication leads unless separately opened/verified. Some Scholar rows can be noisy, broad, or not central to the Bengio/LawZero subsection.

### YouTube/web/news/podcast search

- Web/YouTube searches checked for Aug./Sept. 2026 Yoshua Bengio interviews, talks, LawZero updates, papers and news.
- Verified with `yt-dlp` metadata:
  - `AI Is Already Creating Problems. Here's Why.` - SciencePetr, upload date 2026-08-14, duration 1327 seconds, URL: https://www.youtube.com/watch?v=fvvvtbBvh6Q
  - `AI is moving fast | Yoshua Bengio, Vukosi Marivate, Joelle Barral & Yutaka Matsuo` - AI for Good, upload date 2026-07-31, duration 1405 seconds, URL: https://www.youtube.com/watch?v=O7_OaijiZeE
  - `AI Godfather on Where AI Goes Next | Yoshua Bengio` - Sinead Bovell, upload date 2026-06-25, duration 4296 seconds, URL: https://www.youtube.com/watch?v=4A43Y2fNtZo
- Guardian page checked: https://www.theguardian.com/news/2026/sep/01/if-you-build-something-vastly-smarter-than-you-it-better-be-on-your-side-can-we-stop-ai-from-deceiving-us. Extraction confirmed the title/date/page, while the public Bluesky post provided Bengio's own description of his involvement.
- World Economic Forum article rechecked: `Godfather of AI Yoshua Bengio on why AI can behave unpredictably (and what needs to change)`, published Apr. 17, 2026 and updated Jul. 2, 2026. Useful background but not a new September update.

## Recommended page update made

The AI page should now point readers to this dated source note and summarize the September 2026 update as follows:

- latest check date: 2026-09-02
- substantive new material: Guardian AI-deception long-read/Bengio comment, new SciencePetr interview/video, Bengio's public posts on motivated reasoning and developer responsibility, and LawZero's technical activation-perturbation publication lead.
- caution: frame these as current research/public-warning/governance activity, not settled prophecy or a precise AI-risk timeline.

## Open questions / next monthly checks

- Check whether LawZero adds dated newsroom/blog items beyond the July NVIDIA support and Scientist AI safety-case releases.
- Recheck the `Language Models Recognize Dropout...` publication for a DOI/arXiv identifier and whether it becomes a broader AI interpretability/source-card article.
- Watch for UN Panel thematic briefs or International AI Safety Report updates before the 2027 annual report.
- Continue treating X and LinkedIn as partially blocked unless authenticated/social tooling becomes available.