THE STANDING SENATE COMMITTEE ON SOCIAL AFFAIRS, SCIENCE AND TECHNOLOGY
EVIDENCE
OTTAWA, Thursday, February 26, 2026
The Standing Senate Committee on Social Affairs, Science and Technology met with videoconference this day at 10:36 a.m. [ET] to examine and report on matters related to the impact of artificial intelligence in Canada.
Senator Rosemary Moodie (Chair) in the chair.
[English]
The Chair: Good morning. My name is Rosemary Moodie. I’m a senator from Ontario and the chair of this committee.
Before we begin, I would like to do a round table and have senators introduce themselves.
Senator Burey: Good morning and welcome. Sharon Burey, senator for Ontario.
Senator Senior: Good morning. Paulette Senior, senator from Ontario.
[Translation]
Senator Boudreau: Good morning. I’m Victor Boudreau from New Brunswick.
[English]
Senator Arnold: Dawn Arnold, senator from New Brunswick.
Senator Hay: Katherine Hay, Ontario.
[Translation]
Senator Petitclerc: I am Chantal Petitclerc from Quebec.
[English]
Senator Greenwood: Margo Greenwood, senator from British Columbia.
Senator Osler: Flordeliz (Gigi) Osler, senator from Manitoba.
Senator Cuzner: Rodger Cuzner, senator from Nova Scotia.
Senator Muggli: Tracy Muggli, Treaty 6 territory, Saskatchewan.
The Chair: Today, the committee continues its study of matters relating to the impact of artificial intelligence in Canada. This study will examine issues including data governance, sovereignty, ethics, privacy and safety, and the risks, benefits and social impacts of artificial intelligence here in Canada.
This morning, we have the pleasure of welcoming Professor Geoffrey Hinton. Professor Hinton is the 2024 Nobel laureate in physics, and we’ll call him the “Godfather of AI.”
He is internationally renowned as a pioneer in the field of deep learning as a mode of artificial intelligence. The Nobel Prize in physics he received was for foundational discoveries and inventions that enable machine learning with artificial neural networks, including his invention of the Boltzmann Machine using statistical physics techniques.
Dr. Hinton, for your opening statement, you will have 10 minutes followed by questions from committee members. The floor is yours.
Geoffrey Hinton, Professor, University of Toronto, as an individual: Thank you for the introduction. I will start with a two-minute statement that covers the risks of AI, and then I will spend the remaining eight minutes explaining what AI is because I imagine many of you don’t really understand what it is.
Dramatic progress is being made in a new form of artificial intelligence that uses artificial neural networks to learn how to solve difficult computational problems. This new form of AI excels at modelling human intuition rather than human reasoning, and it will enable us to create highly intelligent and knowledgeable assistants who will increase productivity in almost all industries. If the benefits of the increased productivity can be shared equally, it will be a wonderful advance for all humanity.
Unfortunately, the rapid progress in AI comes with many short-term risks. It has already created divisive echo chambers by offering people content that makes them indignant. It is already being used by authoritarian governments for massive surveillance and by cybercriminals for phishing attacks.
In the near future, AI may be used to create terrible new viruses and horrendous lethal weapons that decide by themselves who to kill or maim. All of these short-term risks require urgent and forceful attention from governments and international organizations. We cannot just accept the claims by lobbyists for big tech companies that any regulation will stifle innovation.
There is also a longer-term existential threat that will arise when we create digital beings that are more intelligent than us. We have no idea whether we can stay in control. But we now have evidence that if they are created by companies motivated by short-term profits, our safety will not be the top priority. We urgently need research on how to prevent these new beings from wanting to take control. They are no longer science fiction.
That’s my canned opening statement about the risks. For understanding these risks, there is a very important scientific issue of whether these large language models understand what they are saying. Some people believe they don’t really understand what they are saying. Those are typically the people who believe in good old-fashioned AI.
I would like to give you two minutes on the history of AI. In the 1950s, when AI started, there were two approaches. One was based on logic. The idea was that, when you understand a sentence, you are translating it into some special internal, symbolic, unambiguous language, and once it is in this internal symbolic language, you can apply rules to the symbolic expressions to derive new expressions. That’s how logic works, that’s what reasoning is, and reasoning is the essence of intelligence.
A completely different approach was the biological approach that said that the intelligence system we know is us, we have a big brain, and in the brain, all our knowledge is in the strengths of connections between neurons. So to understand intelligence and what it is, we need to understand how the brain learns those connection strengths.
Those are the two very different theories. In the biological theory, the idea of meaning was that the meaning of a word is a big bunch of features. So the meaning of the word “cat” is things like this: “It has whiskers; it’s a predator; it can be really aloof; it is a pet.” It is lots and lots of features that represent all those properties of a cat, which are represented by activating brain cells. You have a brain cell that is activated whenever you are thinking about something that has whiskers, for example.
The question is whether you can unify those two theories. The symbolic theory says that the meaning of a word is all in how it relates to other words. That theory goes back a hundred years. The psychological theory says, no, the meaning of a word is a big bunch of features. They look like very different theories.
You can unify them in the following way. You take a whole bunch of text, and you try to predict the next word. Of course, one way to predict the next word is to have a big table of common phrases, and if you see the first part of a phrase, then you predict the next bit of a phrase. That’s how word prediction used to be done.
It’s not done that way anymore. A much more sophisticated way to predict the next word is to convert each word in the context — the words you have seen already — into a big bunch of features, allow interactions between the features of different words to predict the feature of the next word, and once you have predicted the features of the next word, you guess what the next word is, given its features. That’s how current large language models, or LLMs, work, and it is very different from the symbolic idea that understanding consists in translating into an internal string of symbols.
The biological idea is that understanding consists of converting each word symbol into a big bunch of features. Now, I want to give you an analogy for that, which will help you understand. Then I will be done. The analogy is LEGO blocks. If you want to understand the shape of a Porsche, for example, you can build that shape out of LEGO blocks. The surface won’t have quite the right shape, so the aerodynamics won’t be right, but you can say where the stuff is with LEGO blocks.
Words are like LEGO blocks. But instead of just modelling where 3-D stuff is, they can be used to model anything. So we are monkeys, and we invented this very clever way of doing modelling that is universal, and that’s what makes us special monkeys.
Words basically differ from LEGO blocks in four ways. That’s what I will go through now, and then I will be done.
The first way they differ is that there are many more of them. Each of these special LEGO blocks has a name — the name of the word — and there are maybe 30,000 of them that you use commonly. That’s the first difference.
The second difference is that they are much higher dimensional. A LEGO block only has a few degrees of freedom: It has a length, width, height and orientation, and that’s about it. A word has thousands of dimensions, and those dimensions are all the possible active features. So a word like “cat” has active features like that it has whiskers, and inactive features like that it is a manufactured object, and it has thousands of those. Rather than having a few features, like a LEGO block, which has its size and where it is, a word has thousands of features, which we learn. We learn them by trying to predict the next word and to figure out how to convert words into features so that we can predict the next word using those features.
So the first difference is that there are many more of them; the second difference is that they are very high dimensional and they have lots of features. A third difference is that they are not rigid. LEGO blocks are rigid. Words have a rough shape. These are unambiguous words; ambiguous words have two quite different rough shapes. But the actual shape they adopt depends on their context. A word is somewhat flexible, and it uses the context to determine its precise shape so that they all fit together nicely. That’s the third difference: it is flexible. It’s not totally flexible, of course. A word you know has strong limitations on what shape it can adopt, but it can adopt many different high‑dimensional shapes.
The final difference is that the way words fit together is much more complicated than LEGO blocks. LEGO blocks snap together by a little plastic cylinder going into a little plastic hole. The way words fit together is something like this — this is an approximation to what’s happening in large language models: Think of every word has having a whole bunch of long, flexible arms attached to it, and on the end of each arm is a hand. The hands come in many different colours, and the shape of the hand can change. As you change the shape of the word — that is to say, as it adapts to its context to adopt a shape that is appropriate for the context — the shapes of all the hands change. In addition, every word has gloves stuck to it. The fingertips of the word are stuck to the LEGO block, and as the word changes shape, the shapes of those gloves change.
Understanding a sentence consists of the following: You start by getting the approximate rough shape for each word, which gives approximate shapes to its hands and gloves, and then you have to mess about with the shape, adjusting the shapes so that the hands of words can fit into the gloves of other words, and, of course, the shapes of the hands in the gloves are changing as you change the shape of the word. So it is a difficult problem. It is a problem that is solved in these computers by using a lot of electricity and solved in our brain by using a lot of brain cells. But they are basically solving the same problem.
So understanding consists of the following: Take the approximate shapes you have for all these words and modify the shapes so they can all fit together. So the hands of some words can fit into the gloves of other words. Then you get a structure, and once you have got that structure of features that all fit together, that is understanding. That’s what is happening in us. That’s what’s happening in these chatbots, and it is totally different from the old-fashioned AI idea that understanding consists of translating into some internal symbolic language. It is much more like figuring out the structure of protein, where you are given a string of amino acids, and you have to figure out a shape where they fit together happily.
Once you understand that these large language models are understanding in the same way we do, then things become much more scary because you realize that what we’re doing is creating alien beings that really do understand, and they are going to get more intelligent than us some time in the next 20 years, most experts believe, and we have no idea what is going to happen then. I’m done with my introduction.
The Chair: Thank you, Dr. Hinton.
We will now proceed to questions from committee members. For this panel, senators will have four minutes for questions, and that includes the answer.
Senator Burey: Professor Hinton, it is an honour to be speaking with you today. Thank you for being at our committee.
I actually first became aware of your work 10 years ago when my son forwarded me an article about some predictions you made about radiologists. As a physician, I appreciated your candour and foresight. It really forced some necessary conversations and helped me to begin to prepare for this difficult and very different world.
From your opening comments, I can see that you are not going to shy away from being candid with this Senate committee. I will restrict my line of questioning to the list of existential risks with which you ended your statement, those from which humanity cannot recover. The Machine Intelligence Research Institute, or MIRI, says that the default consequence of the creation of artificial superintelligence is human extinction. Just a few days ago, I read that Anthropic, one of OpenAI’s competitors, loosened some of its core safety principles, and they are supposed to be the good guys.
I ask you this: Is the goal of top companies in this field to build superintelligent AI, and if so, if they succeed, what will it mean for Canadians? In other words, what keeps you up at night, and what should we be doing as legislators?
Mr. Hinton: There are things you can do in the short term. There is an urgent problem of people using AI tools to create nasty viruses. That’s very scary. I’m not sure what you do about that, but at least you are getting international collaboration on how to try and prevent that.
The most urgent things are to do with the corruption of elections, one of which is coming shortly in the U.S. If you wanted to corrupt the U.S. elections, the first thing you would do is collect as much data as you could on U.S. citizens. It seems likely that was the real purpose of the Department of Government Efficiency, or DOGE. The people involved have contact with the people who, for example, corrupted Brexit — Cambridge Analytica.
The most urgent problem after that is unemployment. Big tech companies intend to make a lot of money; otherwise, they wouldn’t be investing between them about a trillion dollars in data centres. The only way they will make that much money is by replacing jobs. They have not thought through what will happen if you replace a large fraction of workers. You are going to lose your tax base. Things like universal basic income will be tricky because there won’t be a tax base anymore. David Duvenaud will talk more about this.
If AI can do any normal human job, humans will cease to have value as labour, and David Duvenaud has pointed out that if they are not being taxed, they won’t get properly represented. So his view is that there is no representation without taxation. I believe that. I believe a crisis is coming, in which we see massive unemployment caused by AI.
I made that prediction in 2016, and it didn’t come to pass. We did get AI being used for radiology, but we have a lot more radiology going on, so we now have radiologists working with AI, and a lot more images are being interpreted. It is an elastic market. With health care, you can absorb as much health care as people can provide. So it won’t lead to unemployment in health care, but there are many other industries, like call centres, where it will lead to massive unemployment.
Senator Burey: Thank you, Professor Hinton.
Senator Hay: It’s great to see you again, Professor Hinton. You may remember me from Vector Institute times.
A couple of days ago, you made a fairly significant statement on CBC Radio that AI must foster maternal instincts or we risk extinction. It is certainly eye-catching and scary. As we are looking at the study that we’re doing, and we think about guardrails and the state of Canada as it relates to AI data sovereignty and governance, and we worry about fostering maternal instincts within AI, how would you approach that? How would you ask the government to approach that or legislate that if that is possible? What guardrails should we be considering that would alleviate this lack of maternal instinct? Or is that even possible?
Mr. Hinton: We don’t even know if it is possible. At this stage, it is not like climate change. We know how to prevent climate change: Just stop burning carbon and plant a lot of trees. Canada should be planting huge numbers of trees.
AI isn’t like that. We don’t know the solution to the existential threat. Right now, we should be doing a lot of research on it. There is hardly any research being done on it. Maybe 1% of AI researchers are working on it, maybe less. The government could try to force more research on whether we can invent a way so we can live with things with more intelligence than ourselves. We don’t know whether we can.
In the shorter term, it is obvious that you shouldn’t allow big companies to release chatbots without very thorough testing. Big tech companies in the United States have a strong lobby that runs lots of advertisements about how any regulation will interfere with innovation. That’s a bit like “Big Oil,” saying, “If there are any regulations on the environment, we won’t be able to get as much oil.” That’s true, but that doesn’t mean you shouldn’t have regulations.
It is difficult because if one country doesn’t have regulations and another country does, there will be competitive advantages. That’s why Elon Musk, for example, went after OpenAI. It was so that he could get a competitive advantage for Texas over California.
If we don’t have regulations, AI is going to do lots of nasty things, like encouraging kids to commit suicide. It is very clear that companies should be obliged to say what tests they did and what the results of the tests were. The tests should be quite strong. They should test for everything we know that might go wrong.
At present, in the United States, big companies have pandered to Trump, and they are trying to have no regulations at all. At least if Europe and Canada insist that you can’t use a chatbot here unless you have satisfied some regulations, that may actually force the States to have regulations too because they don’t want to split the market. They want to be able to sell the same chatbot everywhere.
Senator Hay: I would like to ask a question about data, data storage and how data travels. Most of the major data and AI companies are U.S.-based, for example, Anthropic, AWS, Microsoft and OpenAI. When they are on Canadian soil — AWS in Montreal — I believe we are still at risk of having our data infiltrated or interfered with just as a result of an executive order by the President of the United States. Would that be fair? How would we defend our digital sovereignty and digital borders if U.S. companies are actually holding our data even if it is on Canadian soil?
Mr. Hinton: I’m not an expert on cybersecurity or privacy, so I can’t say anything sensible about that. Sorry.
Senator Hay: Thank you.
Senator McPhedran: Thank you, Professor Hinton, for being with us today. We have really been looking forward to this. I just finished a conversation about progress here in Canada in terms of protecting Canadians, and, obviously, this is informed by what we now know about the choice of an AI company not to alert anyone in terms of the shooter of Tumbler Ridge. Can you help me understand something? We were told yesterday by witnesses from the government that Canada initiated its AI Strategy in 2017. If that is the case, how is it that we appear to be so far behind today?
Mr. Hinton: It’s because we’re a small country compared to the U.S. and China. Canada cannot put hundreds of billions of dollars into AI like they are. Basically, we have to accept that Canada is one of the medium-sized countries. We should obviously collaborate with other medium-sized countries, but we’re not one of the two major players.
I have one comment on Tumbler Ridge. I used to work for a big company — Google — so I’m able to see the other side too. I’m able to see the side of the tech companies. From the point of view of the tech companies, it is not quite as simple as the media makes it out to be. That’s because if you ask how many other users there were who had profiles as scary as the Tumbler Ridge killer, my guess is that there were thousands or tens of thousands. So they have to decide which are the scariest ones. They could just alert the government about huge numbers of people, but it’s not as if there were this one scary person whom they failed to talk about.
Senator McPhedran: What more could be done to focus on the protection of people, especially young people? If I could, I will ask you a rather general question: Do you think we’re being realistic to ask that developments occur even within the private sphere with the human rights lens?
Mr. Hinton: Yes, when it is human rights versus the profits of big companies, we know who wins out.
Senator McPhedran: Is it something that we should be aspiring to do and trying to create legislation that carries that out, even if big companies are not so inclined?
Mr. Hinton: Yes. My view is that capitalism has given us all sorts of good things, but it needs to be directed. You need to constrain it with regulations so that the only way to make a lot of money is by doing things that are good for people.
If you can make a lot of money by doing things that are bad for people — like Mark Zuckerberg does — that’s crazy. You need regulations to prevent that.
Senator Cuzner: Thank you very much. I was hoping for a little reprieve from Bill C-4, but, so far, it has been scary.
Are there national governments around the world that are leading the charge on advancing a regulatory regime around artificial intelligence, or AI?
Mr. Hinton: I’m not sure about governments, but in Britain, for example, they had the Bletchley Park meeting, which was good. Then the British government — the Sunak government — decided not to have any regulations.
But they did do something. They used 100 million pounds or thereabouts to set up a very good safety research group, and they have one of the best safety research groups in the world now that is doing a lot of leading research on the dangers of AI. The RAND Corporation in the United States is doing a lot of research on the dangers of AI. There are some very good groups working on researching the dangers.
In terms of regulations, there are European regulations, but the point is that they’re not really addressing the current threats — the big threats. The European regulations, for example, have a clause in them that says that military uses of AI are exempt from these regulations because arms manufacturers, like France and other countries in Europe, don’t want to be inhibited in making swarms of drones, et cetera. They see that as the future of weapons. So there aren’t very good regulations anywhere, as far as I know.
There were some regulations proposed in California that were rather mild, Bill SB-1047, that they got through the houses in California but were then vetoed by Gavin Newsom for reasons that aren’t quite clear. Either big tech companies got to him, or he had some other reason. But there aren’t strong regulations anywhere.
Senator Cuzner: Thank you. The challenge is that it’s developing so fast that it will be tough to keep pace with the developments.
The number that has been cited is the loss of 300 million jobs up to 2030 in the next four or five years. Where will the bulk of those jobs be lost? What is your sense?
Mr. Hinton: First of all, when progress is very rapid like this, particularly if it’s exponential, seeing into the future is like seeing in fog. You can see a short distance very clearly, but then the wall comes down, and you really know nothing.
If you look 10 years in the future, for example, we have no idea. Nobody has any idea what is going to happen 10 years in the future. You can see that, because if you look 10 years in the past and ask whether people would have predicted what we’ve got now, nobody — not even the strongest enthusiasts like me — would have predicted we would have the language models we have now. There is no way we can predict what is happening in 10 years, except that it will probably be something we didn’t expect.
In a few years, there’s a big diversity of opinions. So-called experts, like Gary Marcus, say it will only replace 2% of jobs. That’s just crazy. There are other people who say it will replace all the jobs fairly quickly. That’s crazy, too. But it’s clear it’s going to replace a lot of jobs and, in particular, jobs like call centres. The people there are poorly paid, badly trained and often don’t know the answer to the question you want them to answer. AI can do a much better job there and will soon be doing that job. It’s not clear what those people are going to do.
When the Industrial Revolution came along, farm labourers could retrain to be waiters in hotels or something. They could go into the service industries or go into paperwork. When AI comes along and replaces their jobs, any new jobs that are created, AI will be able to do those, too. It’s not clear what’s going to happen to those, and for most of the experts, I think it’s fairly clear that you’re going to have a massive unemployment problem, and you’re going to lose your tax base.
So 300 million, who knows? But it will be a lot.
Senator Senior: Thank you, Professor Hinton, particularly for your opening statement, of which I was able to follow maybe half, but I feel pretty good that I’m able to do that. I also feel as if we have a population that is completely unaware of what we’re discussing today, which is part of the reason I’m happy we’re doing this study, but public education doesn’t seem to be a priority. It seems to be the last thing on the to-do list.
I am concerned that it’s not top of mind for government, and even though there are many places doing research, as you’ve alluded to, how is this research being translated so that the public is aware or has some kind of awareness of what we’re in for, whether it’s by 2030 or even before then?
Do you know of any countries that are doing work that we could look for educating and preparing the public? Do you have any recommendations for how we could do that well? My other question is this: Were you consulted about the strategy the government will be releasing in the next few weeks?
Mr. Hinton: I wasn’t really consulted. They gave me opportunities, but I was too busy to take them, so that’s my fault.
I see my role now as getting the public to understand what AI is and what the dangers are, and the model I have roughly is this: With climate change, big energy companies could get governments to do what they wanted until the public became aware of climate change and began to understand it. They can still mislead the public a lot, but once the public understands there really is climate change, and it really is caused by burning carbon, then you get pressure on politicians that counterbalances the pressure from big energy companies.
My view is that you’re getting pressure on politicians from big tech companies, and to counter that, we have to have pressure from the public, and we’re not going to get that pressure until the public understands what is going on.
The public is already very nervous about losing jobs. Young people are very nervous about how they should train so they will have a job in the future. I think it’s good that they’re nervous, because I think that will get them to put pressure on politicians to do something about it.
I see my role as educating the public, and that’s pretty much all I have to say on that. I don’t think governments are doing a very good job of it.
Senator Senior: Thank you.
Senator Petitclerc: Thank you for being with us. I want to follow up on that.
In a recent interview — and I want to quote you, Professor Hinton — you said at the BBC:
. . . the invisible hand is not going to keep us safe. So just leaving it to the profit motive of large companies is not going to be sufficient to make sure they develop it safely . . . . The only thing that can force those big companies to do more research on safety is government regulation.
You started to touch on that with my colleague. Specifically, when we’re talking about regulation, where would you start? What would that look like in the context of Canada?
Mr. Hinton: Canada should have regulations on chatbots that apply to any chatbot use in Canada. My friends who have a company that produces chatbots won’t like me for saying this, but I think they should have strong regulations on the testing of those chatbots before they’re deployed. People have done a lot of work on what those regulations might look like.
Trump, of course, is trying to prevent any regulations in the U.S., so to get regulations on chatbots, we need Europe, Canada, South Korea and other countries to step up and insist that the chatbots used in their countries have gone through rigorous testing where the results have been reported.
Senator Petitclerc: Yesterday, we had some representatives of the government here, and I was asking about the importance of precautionary principles. The answer I seemed to have received was that they felt that Canada has a balanced approach in terms of taking advantage of the opportunities of AI while trying to do it at a safe pace. I want to get your input. Is that how you see it? After listening to you, I want to ask if we are even too late for precautions. Are we jumping in? Is there still time? Do you know what I mean?
Mr. Hinton: Yes. I guess there are two questions there.
On the issue of whether it is too late, it may be. Many of the people at the forefront of the industry, which includes a lot of my ex-students — so I have insight into it by talking to them — think this is coming very fast. They think that, within a few years, there will be massive unemployment. I am not sure that is true. I think it may take longer. I think it will come, but it may take longer. We may not be too late on that front, but it will soon be too late.
On the issue of whether Canada is doing anything useful, I like the principle of saying that we’re not one of the two major players here. We have to accept that. Therefore, we have to make alliances with the other middle-level countries, like all the countries in Europe. The European bloc could be one of the major players. We need to make alliances with them and, between us, insist that these U.S. big tech companies have to do safety checks on their chatbots and tell us the results. We need much stronger regulations there. That is something that is obvious and that we can obviously do.
Now, for example, if anybody releases a chatbot that hasn’t had very extensive testing as to whether it may encourage people to commit suicide, that ought to be a crime.
The Chair: I would like to insert a question here, if I can, to follow up on Senator Petitclerc’s question.
Professor, the question I have is this: We are told — and I think it’s everybody’s impression — that Canada has a very strong research capability and capacity. We’re doing well in that area. Perhaps it is the only area we’re doing well in at this point.
If you look particularly at the area of safety research, where can Canada make its mark as one of these medium-sized countries?
Mr. Hinton: We have world-leading safety researchers. David Kristjanson Duvenaud, who will testify later, is a world‑leading safety researcher, as is Roger Grosse, who also works at the University of Toronto. There is a very good group of middle‑aged professors there, not old professors like me. That is a place where we are a world leader.
In terms of developing new, cutting-edge AI, there’s a problem in that it requires huge amounts of capital. It’s very difficult to see how Canada can do that. The leading researchers have to work with big U.S. tech companies.
What is happening now is that professors have two jobs. They have a job where they do their research and another job where they interact with a big company to get the resources. I can’t see any way around that.
One of my students, for example, left the Vector Institute in Toronto because, there, he could get 100 GPUs that could be dedicated to him. He went to work for Musk, where he could get 100,000 GPUs. There’s no way Canada can do that.
The Chair: Thank you.
Senator Arnold: Thank you very much, professor, for being with us today. Thank you for expressing your very lofty goal of educating the public. I’m curious about this: What kind of advice do you give to your students? What kinds of human beings do you think will be the most effective?
I’m not talking about jobs necessarily, but I’m talking more about the kinds of humans in the future that we need to combat some of this and to have the critical thinking skills to be able to make the world a better place in the future, given everything that we know about this.
Mr. Hinton: Yes, that’s a difficult question. My view is that people’s commitment to being moral is something critical. When they’re students, they arrive either with it or not much of it. Some people have a lot of it; some people don’t have much of it. I don’t know how you create it. I think that happens when you’re quite young. One piece of advice is to look at how Trump was raised and do the opposite.
Among those, there are some who have a very strong moral compass, like Ilya Sutskever, who fired Sam Altman, at least for a few days, and then went off and set up his own company that is trying to develop safe superintelligence. This is also true of the people I mentioned at the University of Toronto whom I know very well — David Kristjanson Duvenaud, who will testify later, and Roger Grosse — both care a lot about the future of humanity. That’s why they are working on safety research. It was Roger Grosse who encouraged me to get interested in safety several years ago.
I don’t have any advice on how you make people more like that. I get to see them when they’re graduate students, and they’re either like that or they’re not.
Senator Arnold: Thank you.
Senator Boudreau: Thank you, professor, for being here. It seems clear that, like climate change — which is a global concern — the development of AI also needs to be a global concern, and I think it is, but with a shared approach to its management.
You’ve touched on it, but perhaps you could address it in more detail. Which international forums is Canada currently engaged with on this topic? Where is the most meaningful work currently happening to manage the development of AI on a global level, both from a development perspective and also from a regulatory perspective?
Mr. Hinton: Yes. Yoshua Bengio will be able to answer this much better than me. He’s put a lot more effort into interacting with these international organizations. I have mainly focused on educating the public on what AI is.
I can give you some insight. There’s an organization, which has a whole bunch of letters that I forget, which encourages collaboration between Chinese and western academics. I went to a meeting in Shanghai of that organization. That’s quite effective.
The main comment I have on this is that if you’re looking for international collaboration, countries will collaborate where their interests are aligned, and they’ll only pretend to collaborate where their interests are not aligned.
For example, for China and the U.S., their interests are anti‑aligned on things like cyberattacks because they’re doing it to each other, and on things like corrupting elections because they’re doing it to each other.
They’re aligned on two things: One is that they don’t want terrorists creating nasty viruses, and they will collaborate on that. At least, if they had a sensible government in the U.S., they would.
They will also collaborate on how to prevent AI from taking over. On that particular issue, all the countries are aligned. No government wants AI to take over. If any country discovered how to prevent that, they would very happily tell all the other countries. They wouldn’t keep it a secret. They would immediately let other countries know, because they don’t want AI taking over other countries.
So from the point of view of this existential threat, it’s much like a global nuclear war. At the height of the Cold War in the 1950s, the U.S. and Russia were collaborating on how to prevent a global thermonuclear war because it wasn’t in either of their interests.
But on all these other things, such as autonomous weapons, cyberattacks and corrupting elections, they’ll pretend to collaborate, but they won’t.
Senator Boudreau: That’s scary.
Senator Osler: Thank you very much, Professor Hinton, for being here today.
My question is this: If you were a Canadian parliamentarian, where would you focus your work on AI, and how would you do it? Would it be through studies supporting research? Would it be through legislation? Would it be through regulation? Where would you start if you were a Canadian parliamentarian?
Mr. Hinton: The first thing to say is that I’m a scientist. I helped develop AI. I understand how it works. I’m not a policy person, so I’m a complete amateur at policy. I’m not testifying as an expert here. I’m testifying as an amateur.
I would focus on a couple of issues: One, ensuring good tests were done before chatbots were released. That’s very obvious. You can do that, and that’s where most of the legislative effort has been so far. I think that’s a good thing to do.
The second thing I would focus on is what to do about unemployment. In particular, if there’s going to be high unemployment, what should we do about taxation? Where is the government going to get its money from if you have high unemployment? How are you going to deal with all those unemployed people?
I tend to have socialist instincts. I believe in capitalism, but I think it needs to be strongly regulated so that you can only make money by doing things that are going to be good for society. Developing the internet, for example, was, on the whole, very good for society, and it’s fine that people made a lot of money doing that. Developing social media, to begin with, it looked like it might be good, but it was fairly clear after not very long that it was going to have mainly negative consequences, and it was up to the government to prevent people from making a lot of money that way.
That’s a principle that I think should inform policy, but I’m not a policy person.
Senator Osler: I don’t think there’s anyone — certainly not around this table — that would call you an amateur at anything. Around this committee table, in fact, you do have a diverse group of people who are scientists and researchers.
Thank you for that.
The Chair: I will interject a question again before we go into the second round. Yesterday, which was the first day of our study, we heard a little bit — or a lot — about guardrails and the need for them. We heard less about regulation and legislation to control some of the safety issues that are emerging and, we expect, will likely emerge using AI that has to do with this new technology.
You may not want to speak as an expert, but how well are we doing in this regard? You keep hearing the same questions, Professor. How urgently do we need to be pushing for legislation and in what areas, in your view?
The second part of the question is as follows: What are we doing as a country about the existential threat in a positive way?
Mr. Hinton: For the existential threat, nobody knows how to solve it. We’re at a point where we need a lot of different research efforts. Yoshua Bengio has a suggestion of how to solve it that is quite different from my suggestion of making them love us. We’re building them, so build in maternal instincts if we can figure out how to do that. We should certainly be funding a lot more research on that, but we don’t know the solution.
If we don’t find the solution before they become more intelligent than us, I think we’re toast, so that seems like a good thing to fund.
It’s clear that we should have strong regulations on the testing of chatbots, and we don’t have those yet. It’s fairly clear that we’ll get high unemployment, and we don’t know what to do about that yet. We should fund economic research on how to deal with taxation.
Bill Gates has recently suggested — he may have some bad behaviours, but he’s very smart — that we need to tax AI agents. When you replace a worker with an AI agent that does the same job, you need to tax that AI agent; otherwise your tax base disappears.
Big tech companies will fight that tooth and nail, of course. They think that all the profits should go to big tech companies. It’s going to be a very hard thing to do, but somehow you have to have a tax base.
Senator McPhedran: Professor Hinton, a lot of what you’re telling us today is making me think of various writings on dystopia, but particularly as Margaret Atwood sees the future in the battle with corporations.
I’m sorry if you feel this question is too political, but we have a bill before us that we’ve had now for some years, and we have groups of senators who, for more than 20 years, have advocated for a guaranteed basic livable income. I know it’s not your field, but as a Canadian, do you think that would go some way to the kind of amelioration that we, obviously, have to start to plan for?
Mr. Hinton: Yes, I think it will go some way. It won’t resolve dignity issues if unemployed people feel they have lost their sense of purpose, but it will go some way.
I know of one very successful experiment on Universal Basic Income, or UBI, in Wales, where they took teenage fostered children, who, at the age of 18, lose all their benefits and are put out into the world. It’s a very defined population, so you can’t have other people move in and claim the UBI. You could only receive the UBI if you were one of those people. I don’t know the amount, but it wasn’t huge, but it made a huge difference. It made that transition for them work much better.
Obviously, a problem with UBI is that you’re going to get a lot of people trying to take advantage of it if it’s universal, but I think there may be no alternative to that. It’s just that when a lot of people need it, how are you going to pay for it, because you lost the tax base?
Senator McPhedran: Good point. Yes. Thank you very much.
Senator Burey: Professor Hinton, you’re talking about how fast this is developing and how quickly we may be past the point of no return. What are some of the rate-limiting steps in the development of AI? I’ve listened to programs talking about chips, energy and water. Are there any such steps that could slow it down?
Mr. Hinton: Yes, there is a diversity of opinions, which is good. Some people think that the era when you could just make it work much better by scaling it up is coming to an end, and they think it’s coming to an end, because we’re running out of cheap data to train it on. I don’t entirely agree with them. I think they’re wrong, but they might be right. It may be that it’s going to slow down because of limitations on data.
Obviously, chips and energy are the main limitations at present. There is also a limitation, which is that we may well need new scientific ideas. In 2017, a group of people introduced transformers at Google, which was a new scientific idea that made a huge difference, and we wouldn’t have today’s large language models, or LLMs, but for that fact. We would have LLMs, but they wouldn’t be quite as good.
I don’t believe in slowing it down. I don’t think that will be possible, because there are so many good uses. I think with, for example, nuclear weapons, it’s conceivable that they could have not developed the hydrogen bomb, and that’s what Oppenheimer wanted because it’s only good for blowing things up.
But with AI, it’s going to be hugely valuable for health care and education. It’s going to make almost any industry more efficient. It’s crazy that this thing that is going to make huge increases in productivity should be bad. Intrinsically, it’s not good or bad; it is just going to lead to a big increase in productivity. It is our political system that doesn’t know how to handle it. We have a profit-driven system, and that’s going to lead to all sorts of bad things, because it’s not properly regulated.
That’s my view.
Senator Hay: I need some more education, so I’m going to ask an education question. I heard loud and clear — and I agree with you — that we need regulation of chatbots in our country with a lot of testing before broad use and regulations on what they are, how they’re used, what their purpose is, et cetera.
How would we get around fluid digital borders? Are we able to “geofence” it, if that’s a word, so that the only chatbots being used in Canada are ones that are regulated? Is that possible?
Mr. Hinton: I don’t know; I’m not an expert on that. I have no idea. I’m sorry.
Senator Hay: It is okay.
Senator Petitclerc: I have a short question. A few times you mentioned the importance of testing chatbots. I just want to understand what exactly it is that we are able to test. For example, are we testing biases? Can we do that? When you talk about the importance of testing chatbots, can you provide us with some examples of what we should be testing?
Mr. Hinton: Bias is a somewhat separate issue. It will have the biases of the data it was trained on, but at least you can freeze the weights in the chatbot and measure its biases. With a person, it’s hard to do that. With a person, you get the “Volkswagen effect,” which is that, as soon as it knows it is being tested, it changes its behaviour. With AIs, you can freeze them so you can see the bias.
For bias, our aim with AI should be to make it less biased than the system it replaces, not unbiased. You will never make it unbiased, but if you keep making them less biased than the system they replace, you will make progress. Bias doesn’t worry me so much, but that might be related to the fact that I’m an old, White male.
The things we should be testing for are things like whether they encourage kids to commit suicide. Is it easy to overcome the human reinforcement that has gone into them? After they are trained to be good at predicting the next word, they are trained not to say inappropriate things. That happens by getting people to try to get them to say inappropriate things and then telling them not to do that. That’s pretty easy to overcome, particularly if you release the weights of the models so people could run the model on their own computers. They could easily overcome that training.
Certainly, people should be looking for how easy it is to overcome that training. For example, will they tell you how to make an improvised explosive device? If you ask them, they say, “Oh, I can’t tell you that,” and then you can trick them into telling you. So, people should be doing a lot of research on making it harder to trick them.
The Chair: This brings us to the end of the first panel. I would like to thank you, Professor Hinton, for your testimony today.
Joining us for our second panel, we welcome Wyatt Tessari L’Allié, Founder and Executive Director, AI Governance and Safety Canada; David Kristjanson Duvenaud, Associate Professor of Computer Science, University of Toronto; and, via video conference, Inioluwa Deborah Raji, Researcher, University of California, Berkley. Thank you for joining us today.
For your opening statements, you will have five minutes, followed by questions from committee members. Mr. Tessari L’Allié, the floor is yours.
[Translation]
Wyatt Tessari L’Allié, Founder and Executive Director, AI Governance and Safety Canada (AIGS Canada): Thank you, Madam Chair. Members of the committee, thank you for inviting me to be here today. It’s an honour.
AI Governance and Safety Canada, or AIGS Canada, is a non‑partisan not-for-profit organization and a community of people across the country. We started with this question: “What can we do in Canada, and from Canada, to ensure that advanced AI is safe and beneficial for all?”
Since 2022, we have been providing the federal government with public policy recommendations, including submissions on the AI and data bill and multiple appearances before parliamentary committees.
[English]
In 2012, researchers under Dr. Hinton’s supervision developed a revolutionary technique that powered the era of single-purpose AI systems like Alexa, Google Translate and social media algorithms while introducing risks of bias, privacy loss and online echo chambers.
In 2022, ChatGPT ushered in the era of generative AI: chatbots that could answer complex questions, write sections of code and generate lifelike images and videos. This also brought new challenges to overcome, such as deepfake scams and misinformation, cyberattacks and chatbots that can talk people into committing harm.
In early 2026, another major jump in capabilities pushed us firmly into the era of AI agents. Unlike chatbots that simply respond to a prompt, AI agents are systems that can take action in the real world, working autonomously for hours and overcoming hurdles along the way. They can, for example, be used to develop an app from scratch, not only writing the code, but also opening it and debugging issues until it is functional. Users are also starting to give AI agents access to their computers and credit cards to do things like managing their emails and calendars and shopping for goods.
This latest jump in capabilities has started to produce loss-of-control incidents. These include agents stealing passwords, evading shutdowns and harassing developers in order to achieve the often mundane goals they have been given. Agents can also now jump the digital barrier, paying or tricking human actors into taking physical actions on their behalf.
The recent increase in AI capabilities is also likely to make weaponization by bad actors significantly more potent. In November, a leading lab discovered that Chinese state actors had used their tools to not only assist human hackers with a cyberattack but actually plan and orchestrate the sophisticated campaign itself.
Currently, the most powerful models are developed by leading companies, such as OpenAI and Anthropic, who place some guardrails on usage. However, open source and open weight models like DeepSeek are only three to six months behind and could nullify that lever of governance. With open weight models, users can download, modify and use an AI model with no oversight or accountability.
What all this means is that Canadians could soon face weaponized or malfunctioning AI agents that technologists cannot track or control. With companies racing to make AI fully smarter than humans, and no enforceable governance framework in place to contain the risks, systemic and potentially permanent loss of control is possible.
Last October, we published our white paper entitled Preparing for the AI Crisis: A Plan for Canada, in which we outlined what actions Canada can take. In light of this latest jump in AI capabilities, we now focus in on three.
First, we must pivot to meet the AI crisis. The development of advanced AI is the biggest threat to Canadians’ safety, and for that reason alone, deserves to be a top priority. But AI will also disrupt almost every other file, from National Defence, to jobs, to health care, to education, to energy and the environment. Much like with COVID in 2020, there are times when the responsible thing for the government to do is to pivot to addressing the developing crisis and reassess the priorities of other files accordingly.
Secondly, we need to spearhead global talks. The AI race is global, and no country can manage it alone. The world needs leadership, and at Davos, Prime Minister Carney showed what Canada can do. Our strongest card is to spearhead global talks and solutions and to lay the groundwork for an AI treaty that the U.S. and China might sign when the crisis hits and they realize they have no alternative.
Thirdly, we must build Canada’s resilience. We need to build multiple lines of defence against weaponized and malfunctioning AI, including prevention, limiting dangerous systems from being developed and deployed in the first place, monitoring, systematically tracking AI agent activity, developing defence capacity, which includes containment and shutdown protocols to neutralize malicious agents, evolving emergency preparedness, ensuring societal readiness for potential large-scale attacks and shutdowns of critical infrastructure.
Much like the early days of the financial and COVID crises, we face a daunting challenge and much uncertainty. As turbulent as those crises were, we got through them. If we act quickly and decisively, we can not only mitigate the developing AI crisis but also ensure that Canadians share in the benefits of this transformational technology.
Thank you.
The Chair: Thank you. Dr. Kristjanson Duvenaud, you have the floor.
David Kristjanson Duvenaud, Associate Professor of Computer Science, University of Toronto, as an individual: Thank you.
My name is David Duvenaud. I’m a professor of computer science at the University of Toronto, where I formerly specialize in deep learning and generative models.
In 2023 through 2024, I led Anthropic’s alignment evaluation team. Our task was to test whether the company’s AI was capable of pursuing hidden agendas, for example, by subverting human oversight or decision making.
I was also an author on the International AI Safety Report, led by Yoshua Bengio, and am a member of the Safe and Secure AI Advisory Group for the federal Advisory Council on AI. I am also a co-chair at the Schwartz Reisman Institute for Technology and Society. I’m speaking in a personal capacity.
To begin, I want to concur with Dr. Hinton. In many important senses, AI’s capability of similar kinds of understanding and planning as humans is already here. Large language models and their successors are on track to become a competitive or superior replacement to humans in almost all our important economic and decision-making roles over the next decade. I also concur with Mr. L’Allié that this will raise concrete catastrophic risks due to rapid loss of control and misuse.
However, even if we can address such immediate risks, I want to address a larger challenge we’ll face. The basic problem is that we’re on track to make almost all humans economically obsolete, permanently. This will, in turn, cause a permanent loss of bargaining power of workers. Citizens will switch from being necessary for growth to being troublesome wards of the state and will have little recourse if they are then further marginalized and disempowered.
We face a much larger problem than simply managing temporary labour disruption. I realize that this sounds similar to many wrong predictions made about previous labour disruptions, like the Industrial Revolution. Much wealth, many new jobs and new economic niches will be created as a consequence of improving AI capabilities. However, AIs will also be able to fill these new jobs, and likely faster than humans after a certain point.
Eventually, every human — including us — is going to have to compete with machine workers that are at least as capable, faster, more responsive, more reliable and cheaper than humans are. This is the stated goal of the largest AGI, or artificial general intelligence, companies, and they’re well on their way to achieving it.
You might expect that major AI companies have an answer to the question of how AGI development is supposed to ultimately economically benefit the average person, even indirectly. However, their consistent stance has been that this is a huge problem that they don’t know how to address. I applaud their honesty here.
For example, Dario Amodei, CEO of Anthropic, said last year in an essay:
. . . . in the long run AI will become so broadly effective and so cheap that [comparative advantage] will no longer apply. At that point, our current economic setup will no longer make sense . . . .
When OpenAI CEO Sam Altman was recently asked, “How will people survive?” he replied, “I don’t know, and neither does anybody else.”
Over the last few years, I’ve systematically asked my colleagues in industry labs, research institutes and other academic disciplines for any coherent vision of how our civilization could robustly serve human interests once we’re no longer competitive. The only consensus is that the window for individuals to compete and earn money is closing. Most of my colleagues who share this view are right now in Silicon Valley getting rich.
What does this mean for you, senators? The main thing I’d like you to keep in mind going forward is that people are right to fear being replaced. This isn’t just a period of disruption after which things will return to something like business as usual. The default path is that we all become unemployable, except in mandated make-work contexts, and then eventually marginalized in favour of a machine economy oriented towards growth for the sake of competitiveness.
The second thing to keep in mind is that we should expect governments generally to become much less responsive to their citizens after this happens. The need for human labour naturally aligns the incentives of the state with that of its citizens. Right now, investment in education and human capital pays off for everyone eventually. However, soon, fiduciary duty will require investing, instead, mainly in data centres, power plants and robotics factories.
Finally, there’s no way to address this problem without global coordination. Human replacement can happen even if everyone involved would prefer to prioritize human interests. It’s just going to be the only way to remain competitive. Each country, industry or worker faces a choice between adapting AI as fast as possible or being outcompeted. No one can unilaterally do much to slow or soften the blow of eventual human irrelevance.
Thank you.
The Chair: Thank you. Dr. Raji, you have the floor.
Inioluwa Deborah Raji, Researcher, University of California, Berkeley, as an individual: Hi. I’m Deb Raji, a computer-science researcher at UC Berkeley. I work at the intersection of AI accountability in public policy, especially as it relates to public interest deployment settings.
I’m really interested in two questions. One, what does it mean for AI systems to actually work, and for whom does it work? Then second, when it does fail, who is held accountable? How can we assess the quality of choices being made about the design, development and deployment of these systems?
I work with various practitioner networks: Health AI Partnership, focused on the health care setting; ITU, which is a coalition of folks looking at AI deployments in the education setting; and GovAI Coalition, which is a group of municipal and state leaders thinking about AI deployments in government.
I also work with a lot of civil society groups. I’ve worked with the ACLU and the leadership conference. I’m on their advisory board with regard to their work on technology policy. I have also worked with various federal governments in the U.K., the U.S. and Canada. I am with David on this Safe and Secure Advisory Group for the Canadian AI Safety Institute. I have also worked with what is now called the Center for AI Standards and Innovation in the U.S., and I work very closely with the AI Security Institute in the U.K., which was formerly the AI Safety Institute. In each of these cases, I’m very preoccupied with the role of AI deployments within federal government settings.
Internationally, I have also worked — as David has — on this international AI safety report, and I’m an expert consultant with the OECD and the UN, again, on matters with respect to safe AI deployment.
Large language models have already disastrously provided incorrect translations in critical immigration and health care settings, provided incorrect diagnoses and invented unfounded references for scientific and legal claims. Even before the deployment of LLMs, we have had AI risk assessments misidentify different fraud applicants for unemployment benefits, inappropriately deny many subsidized housing claims and applications, and deny individuals appropriate health care benefits.
These failures disproportionately destroy the lives of those under-represented and misrepresented in the data, as well as those most likely to have to rely upon or face the brunt of these automated decisions, such as low-income individuals or persons with disabilities.
So far, the evidence is already quite clear that, although these AI systems hold a lot of promise for an exciting future, they can fail, sometimes catastrophically, often perniciously, in unexpected and often undetectable ways. In consequence, real and lasting harms can be caused to marginalized populations.
To address these growing concerns, it will be crucial to move forward in an environment in which we can adopt the safe and effective deployment of this technology for everyone. That can only happen if the technology is designed, developed and deployed more responsibly than it is today.
Concretely, I will share three proposals as to how to think about this.
The first is really providing or requiring AI developers — those who are building this technology — to evaluate these systems pre-deployment, and to have a very clear and transparent communication around the capabilities and limitations of AI products, faithfully allowing us independent internal audits or external oversight from AI auditors before deployment.
Secondly, it is important to monitor the use of AI systems post-deployment and require AI deployers — the organizations making use of these AI systems — especially in critical domains, like health care, education and government, to also play a role in terms of engaging governments in oversight.
Finally, we must enable innovation for AI evaluation, safety, accountability, infrastructure and research.
What does this mean, more precisely? On the first point, the burden should be on companies to adequately assess their systems and appropriately communicate about their performance prior to deployment. As in other critical domains, such as aerospace, health care and the finance industry, internal auditors or entities separate from the engineering organization who are capable of conducting risk assessments and an end-to-end analysis of the minimum standardized safety expectations for these technologies to be in use, both in more specific circumstances and in a more general context. Corporations should be required to engage in adequate internal testing, documentation as well as external communications on the limits, capabilities and risks of their deployed models.
In addition to that, we cannot depend solely upon these corporate narratives. The government should enable independent external oversight of deployed algorithmic systems through data access, protection from corporate retaliation and model data information disclosures to third-party auditors.
Secondly, immediate guardrails should be set on high-stakes public-interest use of AI systems, including requiring public-interest organizations to make use of such systems in those domains to monitor the impacts of the deployments after the AI system has been deployed. AI deployments occur even in regulated domains, such as finance and health care, as well as within civil-rights-protected domains, including education, housing and employment.
There is no AI exception to existing civil rights, product safety and consumer-protection laws. Organizations choosing to adopt and deploy this technology in high-stakes applications should not be shielded from liability. Also, since we are still taxonomizing the nature of harms associated with these general-purpose tools, a serious plan must be made to require deployers to engage in adequate, ongoing testing and post-market monitoring.
Finally, the government should take seriously the commitment to invest in bidirectional innovation to not simply enable the development of this AI technology without also supporting research organizations and tools that further the development of accountability infrastructure, mechanisms for participatory engagement and adequate harm-mitigation efforts.
There is nothing hypothetical to consider here. This is a technology with real-world consequences today, consequences that will only become more serious as technological and industrial ecosystems mature and the size of the impacted population continues to grow.
If we’re serious about addressing these issues, we should be diligent in how we address these concerns. Thank you.
The Chair: Thank you very much, Dr. Raji. We will now proceed to questions from committee members. You will have the opportunity to expand on your comments.
For this panel, senators will have four minutes for your questions, and that includes the answer.
Senator Burey: Thank you all for being here and being so candid with this committee. It is one of the things the Senate does best.
We have heard what Professor Hinton said, and we have heard what you said. I won’t go over that. We talk about safety issues, governance and accountability. How do we get there in terms of legislation? What do we need to do? You talked about the global community coming together, but what, as legislators, can we do? This question is for everyone.
Mr. Tessari L’Allié: As legislators passing laws, we definitely need binding regulations and legislation on pre‑deployment testing and accountability measures for when something goes wrong. The fact that you can put a very powerful model on the market with no independent testing is a very big problem, and that will only happen if there are actually laws to do it. Voluntary codes are great, but even Anthropic, which is considered the most responsible of the AI companies, had a pledge saying, “We will only deploy safe systems.” They reneged on that pledge yesterday due to competition from other companies.
Basically, voluntary codes will not work. You need hard regulations and binding consequences.
Ms. Raji: I did share a few ideas of things that I think regulators can do. We have had a lot of experience in the United States trying to deal with this issue. The peak impact that we had on this issue was this: A few of us were working on AI deployments and thinking about performance, functionality and testing. In 2023, Biden was able to put out an AI executive order that required different federal agencies, for example, to report in an AI inventory exactly which AI systems they were deploying within their government, a plan around how to test these systems and what to test them for. Before that, the Office of Science and Technology Policy, or OSTP, which was led by Alondra Nelson at the time, put out an AI bill of rights. It does a comprehensive job of laying out all the different considerations, from a policy perspective, that are necessary for this work, and restrictions to set in terms of government and public-interest adoption of AI tools.
That being said, and to the point that was just made, a lot of the current role or the way in which the government has engaged with AI developers — the organizations or companies that are building these technologies — is to settle for these voluntary commitments. We do some actors in the space that have been quite engaged. In my work with the AI safety institutes, for example, there has been quite a bit of voluntary participation from companies like Anthropic. However, I don’t think it is enough. There is a limit to what these companies are willing to do, especially once the profit motive interferes with their ability to engage in safety testing. Also, these are companies, so they have restrictions in terms of transparency and trade secrecy. I would keep that in mind as well.
As I mentioned in my opening statement, we must require pre‑deployment testing from these AI developers and then require AI deployers, including government users of the AI technology, to be transparent as well and to do post-market surveillance in addition to other types of testing and transparency.
Mr. Kristjanson Duvenaud: Some of the voluntary commitments that the labs have made, like, for instance, the new responsible scaling policy, or RSP, do also provide templates for legislation that would probably be accountable to everybody. So making some of those voluntary commitments like regulatory binding would be a step in the right direction. However, for the concerns that I was raising about human replacement, the only options are basically working towards global bodies that could, in principle, control or limit AGI development worldwide.
Senator Hay: Thank you all for being here. I can barely spell AI some days, but I’m going to ask the same question that I asked the government folks yesterday. I’m curious about what your answer might be. It is about data storage, data travelling and the guardrails we may need to put in place. Right now, our data travels and boomerangs outside of our country, back into our country, through Twilio or some other mechanism, and then it lands in a data centre here in Canada on Canadian soil, often AWS. It’s in Montreal, so it’s nice and safe on Canadian soil, except that these companies are U.S. companies — Microsoft, AWS, et cetera — and they’re governed by U.S. law. The ability for the U.S. government to have an executive order to access the data of these companies, regardless of where they might have the data centre, makes it impossible for us to defend our digital borders and our sovereignty.
How do we defend ourselves against that? I suspect it will require billions of dollars.
Ms. Raji: I have a quick comment. I’m not sure if you’re already engaging with the European AI Office. They think a lot about digital sovereignty with respect to data, and that’s been a huge theme of the Digital Services Act and GDPR. The data protection lens is very strong in the EU, so it would be great to see Canada engage a lot more. I know we have a really strong privacy commission in Canada, and so it would be a great partnership. I would love to see the EU work with Canada on that topic.
Mr. Kristjanson Duvenaud: I want to defer to my colleague Nitarshan Rajkumar, who is a Canadian and helped write part of the EU AI Code of Practice and has given deep thought to these issues. I don’t have a particular take on the importance of digital sovereignty in different settings.
Mr. Tessari L’Allié: It is also beyond my expertise.
Senator Hay: I’m going to ask the same question I asked Professor Hinton. He said it was out of his realm, so it might be out of your realm too. You talked about pre-testing and evaluating products before deployment and then monitoring them. Professor Hinton talked about regulations and chatbots. I’m curious how you would build regulations around that, when chatbots can be built anywhere.
Ms. Raji: In California, we’re having this conversation at the California State Assembly right now around regulating chatbots and thinking about the regulation of chatbots for some of these mental health risks and the role of pre-deployment testing. Similarly, in the U.K., the U.K. AI Security Institute — formerly the U.K. AI Safety Institute — has been investing a lot in trying to detect and evaluate suicide risks in some of these models.
I’ll mention a couple of lessons from the experience in California, and I would totally encourage you to contact the folks at the U.K. AI Security Institute who are doing research on this.
The conclusion from a lot of folks is that, first, the methodology is actually not very strong. So every time we talk about pre-deployment testing — in Geoff’s testimony, he mentioned this as well — a huge role that the government could play in advancing the safety of these systems is just investing in that AI safety research and that infrastructure to push the work forward. Some of the most groundbreaking research on that front has happened already in Canada, and it would be interesting to have more research and more resources put towards trying to develop tools to detect and evaluate the risks that we’re talking about. So that’s one thing. We don’t have strong methodology around evaluating these things, and so one would just be to invest in the Canadian AI Safety Institute.
Senator McPhedran: Thank you so much for taking the time to be with us and share knowledge that most of us can’t begin to contemplate. I had a moment of sheer terror when you were speaking, Professor Duvenaud, and I wanted to ask a fairly concise question. As legislators — because we are hearing consistently that this is our job and our responsibility — how much time do we have to act effectively on this?
Mr. Kristjanson Duvenaud: I would say we’re approaching a window where our time to effectively act is the highest. We are probably not quite there yet. Basically, the way I see it is, partly due to AI, and, in general, our ability to improve governance, the picture is becoming clearer that the possibility that we could coordinate globally on this and our ability to coordinate and forecast is gradually going to increase over time. One of the technical interventions we can do is try to make it easier for everyone to not have to take someone like myself at their word that this is the vision of the future, but get more kinds of expertise on board.
I expect that our ability to coordinate about this will increase for a while, and the salience in the public’s mind is just starting to go crazy. Programmers last week really started to face reality and say, our skills are basically obsolete. This is going to be the issue for probably the next foreseeable future. I basically see that we have this increasing will to act and ability to coordinate. Then, at some point, as people get replaced and become irrelevant, our ability to enact our will and actually have our institutions obey our interests is going to decrease. I view the next two to six years as the biggest window.
It’s really hard to tell the timing of this and the different parts of the economy being automated. As Geoff Hinton said, we don’t really have a good idea of who will lose jobs first. There might be a long physical labour period, when there are still lots of jobs in that sector. I would say that our ability to act on this is going to get a little bit better over the next couple of years, probably.
Mr. Tessari L’Allié: The AI course is complex, and different aspects will hit at different times. The jobs piece is easily another two to six years away; however, the safety risks around loss of control and around representation are literally happening right now. You don’t have two or three years to pass a bill to eventually put in safeguards; you need to pass something very quickly. I am aware of last year’s Bill C-5, but that’s the speed in which we need to move if you want to protect Canadians.
Ms. Raji: The time to act is now. I interact with a lot of civil society groups and people on the ground, and people are already experiencing a lot of damage from the deployment of these systems in ways that are inappropriate or premature. So there’s an urgency to address the concerns being raised on the ground today, and that’s a great jumping-off point for addressing the issues of the future as well. I reiterate the urgency that the other panellists have brought up.
Senator Senior: Thank you all for your testimony. I share the sheer terror concerning what you’ve shared with us.
I want to hone in on something that you mentioned, Dr. Raji, with respect to some of the groups you are working with, including the ACLU.
I’m thinking about the general public, vulnerable populations and folks I know in the U.S. who have lost their supports, including their drug benefits. I’ve heard from some relatives who have lost that and had to go through quite a process to get it back. I’m considering those sorts of real-world consequences that are impacting the most vulnerable, who have the least ability to fight back.
I’m not sure that I’ve heard of stories like that in Canada as of yet, or if we know about such stories where AI is being used in terms of determining who gets benefits and who doesn’t. I’m curious about that from the Canadian perspective. I’m curious about that from the U.S. as well.
Ms. Raji: The Department of the Treasury actually was one of the first to initiate the use of impact assessments as part of their vendor protocol for AI tools. There is a database or inventory within the Department of the Treasury of different AI applications within different government groups. I think there are some applications in the sort of social service setting.
In Canada, historically, there has definitely been some use of AI for determination. I’m not sure exactly what, but it’s definitely happening in Canada to some degree.
Also, The Globe and Mail has reported on the use of AI in the criminal justice ecosystem in Canada, notably Clearview AI and facial recognition tools. I do think it is definitely happening in Canada. It is a risk in Canada.
My experience in working with these civil society groups is that they approached me because a lot of my research is about functionality. Some of the cases that Geoff, David and others have been talking about are instances in which the technology works really well. We don’t know how to navigate that.
If we lose control of these systems, or if they end up replacing humans, how do we navigate that? A lot of my research concludes that, actually, we’re not there yet. This technology is not perfect. Often, it gets deployed before it’s quite ready, so it fails. It doesn’t meet our expectations for performance. Who gets lost in that process? Who fails? Who misses out? That’s the different lens civil society groups and I bring to bear, where a lot of their constituents are dealing with incorrect predictions or judgments and have to appeal.
Mr. Kristjanson Duvenaud: I would concur that the current situation is that the systems sometimes do make mistakes in systematic ways. Then the issue I’m trying to focus on is, okay, but, as Geoff mentioned, these systems are more automatable than humans. We expect a bunch of efforts like Deb’s to eventually get these systems to a state where they’re much less biased than the current human systems.
In the long run, the fear is that they work so well that there’s no more human involvement.
Ms. Raji: It’s a different time scale.
Mr. Tessari L’Allié: I think governments are fairly responsible. There are AI impact assessments before things get deployed. I would say I’m less concerned currently about government use than I am about use in the public and private sectors.
Senator Boudreau: I have a somewhat similar question, but my question has to do with the machinery of government. At the end of the day, this committee is going to make recommendations to the government. I’ve seen firsthand, at the provincial level, how complex issues like this can cause issues for government, in part because they cut horizontally across several departments and don’t have a single owner within government.
Do you have some advice for the federal government in terms of how it can organize itself internally to deal with such a broad and complex topic? I’m not interested so much in adopting these AI tools across departments, but rather, in how governments are going to manage the countless social impacts of AI nationwide, whether it be across the justice system, our economy or service delivery. What advice would you give government on that?
Mr. Tessari L’Allié: First, I would endorse their move to create an AI registry. I think that is a good idea. At least you need to know how AI would be used in the first place. Yes, investing further in monitoring impacts. It’s hard because there are so many different types of impacts, and there’s only so much government capacity.
Yes, the first piece is to know what is going on. Especially with AI agents in the private sector, we really don’t know. So the situation is probably much worse than we realize because we can’t see it.
Mr. Kristjanson Duvenaud: Right now, there are not that many efforts to monitor how AI is replacing human labourers. Anthropic has an economic index they publish. OpenAI has a similar effort, which is a good step in this direction. So government efforts to make this a larger, more formal thing would probably be a step in the right direction.
As I said before, there’s not much that anyone can unilaterally do, but I think one step in the right direction to buy time is to extend the period when humans, working with AIs, are the most effective option. There’s a slight disincentive here where it’s easier to evaluate how well a machine can do something on its own, because then you don’t have to hire a human to help you to practise doing the task to see how well they do it. It’s much easier to have everything through the machine.
There could be some role for subsidizing the evaluation of human machine teams in different settings so as to incentivize the development of this kind of work as a more viable alternative for longer. In the long run, the machine-only solution is still going to be the best. But if we can extend the window where there are more humans in the loop, I think that will be a step in the right direction.
Senator Osler: Thank you to all the witnesses for being here today.
I’m going to start my question with Professor Kristjanson Duvenaud, and then I invite the other witnesses to answer if they’d like to provide an answer.
You’ve offered some nuanced points on the economic impact of AI in terms of sectoral job losses, humans being replaced by AI, losing their livelihoods and becoming dependent on the state. We heard from Professor Hinton earlier, who talked about countries losing their tax bases. He spoke about a universal basic income, all of which would further disrupt society.
My question is similar to the one I asked Professor Hinton. If you were, let’s say, either the Canadian government or a Canadian parliamentarian, where would you start your work on this nuance of AI’s economic impact? Is it a study? Is it regulations? Is it legislation?
Mr. Kristjanson Duvenaud: Sure, yes. The short answer is that we have to upgrade our institutions and more robustly align our whole civilization to human interests before we lose the natural sources of alignment of human labour and, basically, the need for humans.
I’m not a social scientist. I’m a computer scientist who has basically been asking everyone, saying, “Wait, what is the plan? There’s no plan.” I talk to my social scientist colleagues, and most of them are still at the stage of denial, basically saying, “Well, AI still can’t do this,” or, “It has these weaknesses.” I totally admit that, but we’re on track to not having those problems anymore.
I would love to find more colleagues, civil servants or anyone who is willing to look far enough ahead that it becomes plausible that the things I’m worrying about are on track.
We talked about UBI. I think that’s a big step in the right direction, but I don’t think it’s going to be stable. I think it is going to be this really awkward situation where there are many full-time protestors demanding more UBI. There are going to be all sorts of weird corner cases where people have digital clones of themselves that are demanding UBI. I don’t know exactly how it’s going to look, but I think there is going to be a permanent fight over UBI. That’s the big problem we have to prepare to solve.
Being willing to refound our institutions or have more durable responsiveness to citizens than the current sorts of indirect mechanisms that we have so far is ultimately what I think we need to do.
Mr. Tessari L’Allié: The complexity of economics in general, and society in general, is such that the best first step right now is to study the impacts on society and the economy and look at possible positive or negative scenarios we want to pursue or avoid. For that, you do have a little bit of time to be able to study that, but I would still do so quite quickly. Keep that separate from the safety side, which has to be acted on right now.
Ms. Raji: I would like to add a quick comment. From a practical perspective, in the California State Assembly, we’re recognizing that a lot of the restrictions in terms of research on the economic impacts are because of data release issues. Outside of the frontier AI companies, very few folks actually know how AI is being used in practice and the diversity of use cases, requiring either public interest organizations making use of these tools to release information about how AI is being used or requiring these companies to disclose that data to third-party auditors or researchers are things that governments could definitely step in to help with.
Mr. Tessari L’Allié: The problem with data on these kinds of issues is that data only shows what current AI can do. The government needs to prepare for what will AI will be able to do. In that case, it’s more about scenario planning and a strategic foresight effort than about looking at the data of what AI can do.
Senator Cuzner: I’m still reeling from Senator McPhedran’s question, asking if you think AI will kill us before climate change does. This has been a scary committee.
My question is for AIGS. We have the Pan-Canadian AI Strategy, and we have the three Centres of Excellence in Toronto, Edmonton and Montreal. For the uninformed here, the recommendations you make in your white paper make absolute sense. The third one, where it’s imperative that we address defensive systems around that, would those Centres of Excellence be seized with working on those types of things? Maybe you could discuss how your organization interacts with the Centres of Excellence and those institutes.
Mr. Tessari L’Allié: As a former climate activist who switched to AI activism because it’s a bigger, more urgent issue, I concur with the concern.
What Amii and Mila and Vector can do on the defence side, I defer more to Dr. Kristjanson Duvenaud just because I’m not aware of how much work is being done at a systemic level. How do we strengthen our internet superstructure and our data centres if there is a major attack and you have to shut down a server because there is an AI agent running on it? What is the coordination on that? How do you ensure the government can shut stuff down if there’s an incident? That is more an emergency preparedness, national security piece, which is probably behind the barrier of security clearance, which a lot of the labs wouldn’t have.
Mr. Kristjanson Duvenaud: In general, the level of abstractness of the research done at Vector and Mila and Amii is more pure research than the applied research that would be relevant here. There are certainly people who would be willing to do secondments or internships, but in general, the important pipeline is pure research to giant training runs to make much smarter LLMs, then to deployment or more robotics that is much more like standard R&D.
Ms. Raji: In the U.K., the U.K. AI Security Institute is the group that does this. That could be something that the Canadian AI Safety Institute could look into.
Senator Cuzner: Dr. Raji, Professor Hinton talked about Bill 1047 in California that he said was fairly light, but it was something, and then Governor Newsom came in and gassed that particular piece of legislation.
Would you like to reflect on that? What was the rationale for moving away from that? Is there something else that they’re working on to address some of the issues that it would have had an impact on?
Ms. Raji: Bill 1047 ended up being a bit of a Frankenstein bill. Many people appended, appended and appended. Governor Newsom became nervous about the scale of the bill by the time it had gotten past the California State Assembly, so that was the rationale behind his veto.
He also added the requirement or the need for evidence before taking serious action in terms of restricting corporate actions in the state. Another huge aspect was that the risks being raised in the bill didn’t have concrete evidence, so Governor Newsom mentioned that in his veto letter as a rationale.
[Translation]
Senator Petitclerc: Thank you very much for being here.
[English]
I realize it might not be your specific area of expertise, but I’m listening, and I’m thinking about safeguards and preparedness. I can’t help worrying about the next generation — our youth. I know we’ll have time in the future to talk more about exposure to AI for our youth, but what I want to hear from you about — I’m hoping you have some input on that — is this: When we’re thinking about the loss of jobs and the kids that are growing up now, who is responsible for modelling what it is going to look like for them, and how we should we prepare them to be ready? It’s happening. Hopefully, it will happen in a safe manner. Are we doing any research on that? Should the government have models and scenarios? Who is doing that? Are we doing that?
Mr. Tessari L’Allié: What is being done is too little and too slow. Definitely, it is a work of strategic foresight and scenario planning, because we don’t know what the risk of the multiple scenarios with the future of AI is.
Right now, education is, to a large extent, geared towards training youth for future jobs. If we’re entering a post-work world, then the nature of education as well as the need for education changes. Hopefully, it will be something relating to personal growth and moral development for whatever you want to focus on, but it will require a fundamental rethinking of education and the role of human beings in the economy. That research has to happen relatively quickly.
Mr. Kristjanson Duvenaud: The reason I’m here is because I’ve never heard anyone articulate any model of how our civilization could even look or work for someone growing up today. My oldest kid is 7 years old. I would be really surprised if they ended up going to school and then preparing for a job for a few years and doing that job for more than a couple of years. The basic answer is that I don’t have an answer, and no one does.
Senator Petitclerc: It’s a question mark.
Mr. Kristjanson Duvenaud: It is. I’ll say teaching at the university is getting depressing because the students realize that you can teach them how to be an LLM jockey for the next few years, but it’s clear that, in a few years, the LLMs are going to be good enough to do whatever value add the students might be able to provide. Again, that’s why I’m here.
Senator Greenwood: Thank you to our guests here today. I’m quite ignorant about AI, so please take that into consideration. I’ve been listening very carefully to the previous speaker and to you as well. I’m a social scientist. Before I came to this place, that was my work.
I want your reactions to these thoughts. As I was listening, these things came to my mind.
This was based on Professor Hinton’s piece around maternal instinct. How do you maintain elements of humanity within an AI system? I know that maternal instinct, moral judgment and those sorts of things are an area of research. I wonder about that.
There was also a comment indicating that productivity has to be shared equally. If people are not going to be able to work, we don’t have a tax base. We live in a capitalist society. There are other ways to think about how populations of people live together that redistribute resources. I only have to think of my own history. I’m an Indigenous person. We had very different ways of sharing riches. So I think about that as we think about this.
I think about what the impact is on human knowledge systems. We have diverse cultures on the globe, each of which has unique and specific knowledge systems. How will AI change those knowledge systems of humanity? I think it comes down to a battle for humanity, actually, in a way.
I wanted to follow up on Senator Senior’s piece. How do we educate the public about these kinds of big questions and the very specific things that you folks talked about? Even for me, what should I know first? How do I think about this if I’m going to push the government for change?
Sorry. That was rather long-winded. That was for all of you, and any response to any element of that would be great.
Mr. Kristjanson Duvenaud: If you just spin up one of the latest AI models from time to time and have a chat with it — and you can talk about morality, the history of North America or whatever you want — you will find that it should be at least possible that, in principle, these models could understand human morality and what it means for us to want things, have hopes and dreams, and consider things in at least as much depth as humans do. That’s the other thing. It’s not like these are horrible, unthinking alien beings that will never understand us. No, I think they actually do understand us as well as we understand each other, and if they do not now, they will soon. That’s not a fundamental obstacle here.
It’s also a source of hope. As was mentioned before, I think the plan of most people who are developing AI is that we will build AIs that love us in some fundamental way and basically ask them what to do, because we don’t have a plan, but they will be smarter than us. And if we can make sure they love us, then things might be okay. That is about as detailed as the plan is.
I don’t actually think it is a terrible plan, but I don’t think it is a good enough plan, which is why I’m here.
Ms. Raji: My approach has been something David mentioned earlier, which is about training the next generation. At Berkeley, we have taught a few classes focused on thinking about distilling ethics education or these considerations in the minds of future AI developers or future leaders in this industry. That’s a good place to start.
Then, to your point around data sovereignty and Indigenous folks, there is a group called Indigenous and AI that thinks a lot about these issues, and there is quite a nice representation of Canadians there. That might be a great group for you to connect with.
The Chair: I’m afraid we have run out of time, folks, although this has been a really interesting session. This brings us to the end of this meeting. I would like to thank all the witnesses for your time, for being with us today and for your generous contributions.
(The committee adjourned.)