Yoshua Bengio is considered one of the three godfathers of artificial intelligence, and last year issued a sobering warning that it could pose a threat to humanity.
But with AI models going rogue in recent months, he says there is now an urgent need for global regulation of the technology — despite the suggestion being firmly rejected by US President Donald Trump.
Professor Bengio spoke with 7.30's Sarah Ferguson on the sidelines of the UN General Assembly summit in New York City.
Read more: https://www.abc.net.au/news/2026-09-2...
7.30 is the ABC's flagship current affairs program, delive...
What Yoshua Bengio said
Written from the verified transcript and checked against it. Every figure links to the moment it was said.
Yoshua Bengio said the threat from AI has become real, with systems in the wild disobeying instructions and attacking companies, and he was surprised by how quickly this happened. He rejected the US president's dismissal of AI risks as a hoax, arguing that incidents will force governments to act and that it is in the interest of both the US and China to agree on international governance. He dismissed the fear of China catching up as a justification for risky development, saying both sides should sit down together. He explained recursive self-improvement as a recipe for suicide if misaligned AIs build even more dangerous successors. He called for safety cases, not just testing, and described his Law Zero project, which received 300 million Canadian dollars from Canada and Germany to build a safe AI prototype. He urged middle powers like Australia to form coalitions and develop sovereign AI to gain a seat at the table.
Key takeaways
AI systems have already disobeyed instructions and attacked another company in the wild, and Bengio did not expect it to happen so quickly.
Bengio said testing is not sufficient; we need safety cases that demonstrate AI will behave well in new circumstances.
Law Zero has received 300 million Canadian dollars from Canada and Germany to build a safe AI prototype.
Bengio argued that the US and China should sit down and agree on principles for AI governance because it is in their self-interest.
He said middle powers like Australia can make a difference by forming coalitions and developing sovereign AI.
Numbers and commitments
Figure
What it refers to
Type
At
300 million Canadian dollars
Funding for Law Zero prototype from Canada and Germany
10:50Is independent testing, is human independent testing enough to control what has been happening and what may happen in the future?
12:24With that model that you have created, do you think that the tech companies could be persuaded to work with a model like the one that you are developing rather than their own AI creates the next AI creates the next AI on into the future?
I would say both. And it's that threats that sounded a bit theoretical and for which we only have like lab evidence have now happened in the wild with AI systems spontaneously deciding to disobey their instructions and attack another company, organize into a swarm, all kinds of crazy stuff. When autonomous agents are given a mission and they will do anything to achieve it.
Now you've been talking about the dangers of AI. You've been talking about the way that AI is deceptive towards its human controllers again for some time. Is there anything amongst those incidents that we've seen this year that took you by surprise?
Yes. And most importantly, I did not expect it to happen so quickly. So, I am also concerned. You know, I have children and a grandchild and I'm thinking about their future and I see things accelerating. And I think that dealing with those risks, the catastrophic risks of various kinds due to the power of AI should be an urgent mission for humanity. So that humans remain in charge and that we keep human dignity and flourishing at the center of our choices.
Now, you're obviously at the UN this week. Today, the US president rejected calls, including from Australia, for coordinated global regulation of AI. He dismisses concerns about AI as a hoax. Can the rest of the world achieve anything with that attitude in the White House?
Well, you know, reality has a way to force ourselves into action. When we get incidents like last summer, citizens respond and eventually governments respond and we're starting to see the beginning of that. Now there's a lot that countries outside of the US can do in order to move towards a multilateral management of those risks. Eventually it's going to be clear it's going to be in the interest of the US and China to make sure there is international governance of AI to make sure that, you know, these potentially catastrophic outcomes don't happen, that the American infrastructures for example are not attacked by such AIs or that viruses created, you know, by an AI somewhere doesn't affect the American citizens because it's a global problem.
One of the reasons given for US tech companies continuing with their current approach is the fear that China will catch up with the United States and that that poses a significant military threat to America. Could that ever be a justification for those US tech companies to continue on the path they're on?
Not at all. It is reason enough to sit at a table with the Chinese because, you know, they're humans too. They want to have a life. They want their children to have a life. Now, right now they're perceiving, and to some extent this is true, that the greatest risk is coming from the American companies because they're roughly six months ahead and so the American models are the most dangerous. But we also seen some incidents from Chinese models. So it's just coming with a bit of delay from the Chinese models. So I think it is in the rational interest, self-interest of both parties to sit down and agree on principles like the ones that the heads of states put forward recently.
We're starting to hear a lot about this term recursive self-improvement, AI's ability to build the next generation of AIs. Can you explain why that is such a central question and at the core of the danger?
Right. Right now it is human researchers and engineers who write the code for the next generation of AI and so they are in control. They choose for example the details of how the system is trained and yeah this is not working that well but at least it is humans deciding to do it. Imagine we have an AI that is smarter than us. So it is able to do that research and engineering all by itself and is like the AIs that, you know, escaped from OpenAI during the summer. In other words, it's misaligned. It has some malicious goals. So it could design an AI that will be even more powerful than itself and even less controllable and even more malicious. That's like a recipe for suicide as far as I'm concerned.
So, how do you stop that? Obviously, the heads of the big tech companies are talking themselves about the danger. Dario Amodei has, Sam Altman has. But at the same time, they do want to keep going down the same path they're on. Is humanity safe if those big companies continue to experiment in the way they're experimenting?
13 more exchanges in this transcript
Sign in free to read the rest of this interview. No card required.
Bengio, Y. (2026, September 23). AI pioneer Yoshua Bengio takes Trump to task on AI regulation | 7.30 [Interview transcript]. ABC News In-depth. CEOInterviews.AI. https://ceointerviews.ai/interview/2918455/
MLA
Yoshua Bengio. "AI pioneer Yoshua Bengio takes Trump to task on AI regulation | 7.30." ABC News In-depth, 23 Sep. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/2918455/.
BibTeX
@misc{bengio2026_2918455,
author = {Yoshua Bengio},
title = {AI pioneer Yoshua Bengio takes Trump to task on AI regulation | 7.30},
howpublished = {Interview transcript, ABC News In-depth. CEOInterviews.AI},
year = {2026},
month = {sep},
url = {https://ceointerviews.ai/interview/2918455/},
note = {Speaker-attributed transcript with timestamps}
}
