THE STANDING SENATE COMMITTEE ON TRANSPORT AND COMMUNICATIONS
EVIDENCE
OTTAWA, Wednesday, April 15, 2026
The Standing Senate Committee on Transport and Communications met with videoconference this day at 6:48 p.m. [ET] to examine and report on the opportunities and challenges of artificial intelligence (AI) in the information and communication technology sector.
Senator Larry W. Smith (Chair) in the chair.
[Translation]
The Chair: Honourable senators, I welcome you to this meeting of the Standing Senate Committee on Transport and Communications. Thank you for your cooperation.
[English]
My name is Larry Smith. I am a senator from Quebec and chair of the committee. Now I would like to ask my colleagues to introduce themselves.
Senator Simons: Senator Paula Simons, Alberta. I come from Treaty 6 territory.
Senator Wilson: Duncan Wilson, British Columbia.
Senator Mohamed: Farah Mohamed, Ontario.
Senator Arnold: Dawn Arnold, New Brunswick.
[Translation]
Senator Petitclerc: Chantal Petitclerc from Quebec.
[English]
Senator Lewis: Todd Lewis from Saskatchewan.
[Translation]
Senator Aucoin: Réjean Aucoin from Nova Scotia.
[English]
Senator Dasko: Donna Dasko, Ontario.
The Chair: Thank you, colleagues. I would like to welcome everyone with us today as well as those listening to us online on the Senate’s website, sencanada.ca. We are meeting today to continue our study on the opportunities and challenges of artificial intelligence, or AI, in the information and communication technology sector. With that, I would like to introduce our first panel.
We have with us today Emily Laidlaw, Canadian Research Chair in Cybersecurity Law, Associate Professor, University of Calgary; and Connor Leahy, United States Director, ControlAI, who is with us virtually. Welcome. Thank you for joining us today.
Witnesses will provide opening remarks, which will be followed by a question-and-answer session with senators. I will now invite Ms. Laidlaw to give her opening remarks.
Emily Laidlaw, Canadian Research Chair in Cybersecurity Law, Associate Professor, University of Calgary, as an individual: Thank you, chair and honourable senators, for inviting me to speak today. I will focus my remarks today on two key points: that Canada needs to update its laws to address AI, which is probably not surprising, and that “deepfakes” create unique threats that require particular attention.
First, about AI laws, I want to echo a key message that Brent Arnold will make later today, and that is that Canada cannot continue to rely on outdated, piecemeal laws to handle the tsunami that is AI. We need coherent and comprehensive laws grounded in a whole-of-government strategy.
If I could wave a magic wand, what would a healthy, resilient AI environment look like for Canada? It would mean that innovators and investors have regulatory certainty about their compliance obligations in Canada; that our laws provide rigorous accountability but are sensible and do not create needless complexity for businesses; that Canada has embedded digital sovereignty through investment in Canadian talent and training, minimized dependency on U.S. players and developed strategic international partnerships; that individuals harmed by AI have avenues of legal redress; that all levels of government — federal, provincial and across departments and agencies — work in coordination; and that Canadians are educated about AI, cybersecurity and information manipulation.
So how do we get there? And, importantly, where do we start today?
First, the federal government should prioritize passing comprehensive digital laws and, in the meantime, work with industry to create codes of practice. Several bills have been introduced in recent years on privacy, AI, online harms, critical infrastructure and age verification — I’ve spoken with all of you a few times on that. All of these deal with different dimensions of the AI conundrum. The foundation is privacy law, as AI runs on data.
The next layer is cybersecurity. AI poses an existential threat to cybersecurity while also being part of the solution, as we saw, of course, with news last week of Anthropic’s latest Mythos model. That is one story among many.
The next layer would be what I would call risk management laws. Both the proposed AI act and online harms legislation were based on the idea that companies have obligations to mitigate the risks associated with the products and services they are putting out into the world.
Second, interoperability with other legal regimes is the name of the game. What do I mean by that? Technology is global, and a secure and resilient Canada relies on cooperation with others. That does not mean our laws are the same but that they have certain features in common. For example, risk management duties are the interoperable features of AI and online harms legislation in numerous jurisdictions. That is why that foundation is so important for any law passed in Canada.
Lastly, I want to talk about the unique threat posed by “deepfakes.” AI, as we know, has made creation of hyper-realistic images, video and audio easy and voluminous. This is a threat at all levels, from weaponizing foreign actors to undermine democracy, to fraud, creation of child sexual abuse images and intimate images of adults or other distortions that can ruin reputations, and so on.
In my research, I talk about it as a slow violence. It’s gradual, hidden in plain sight and rooted in structural inequalities. There is only so much law can do to address this.
Laws can be updated to hold individuals legally culpable when they can be identified. Bill C-16 is a good example of this. Currently, it is not a crime to create and share an intimate “deepfake,” but Bill C-16 proposes to amend the Criminal Code to criminalize that.
The other area is platform regulation. This would be the online harms legislation that was introduced in Bill C-63, duties to address the systemic risks of harm. It might be labelling synthetic media or having special rules during crises or election periods. It might be policies for impersonation.
These are systemic laws because they are not concerned with actioning individual pieces of content. Rather, they are more about the deeper organizational and technical practices that, together, improve the health of the information ecosystem. While some AI-generated content is illegal and should be removed, some of the most insidious forms of “deepfakes” that impact democracy are perfectly legal, so systemic approaches are the only and best solution.
I see that I am almost out of time, so I will leave the comments there, and perhaps we can discuss a human-centric approach to cybersecurity in the question period. Thank you.
The Chair: Thank you, Ms. Laidlaw.
Connor Leahy, United States Director, ControlAI: Thank you, Mr. Chair and members of the committee, for inviting me to testify today.
I am the U.S. Director of ControlAI, a non-profit organization focused on mitigating the security risks posed by advanced AI. Previously, I was the CEO of the AI safety start-up Conjecture and the head of the open source research collective EleutherAI.
There are many varied, complex and important challenges we face with AI, but today, I want to highlight the catastrophic risks up to and including human extinction that are posed by what is called superintelligence or “smarter-than-human” AI.
The existence of systems that could outperform humans across all relevant domains — including science, business and politics — and that is not controllable by trustworthy, democratically elected institutions poses a massive national and international security risk.
While such systems do not yet exist, the creation of the first such superintelligent system is widely expected by experts, both in industry and academia, to happen within the next two to five years. It is rare these days to find an expert in the field who believes it is more than 10 years away.
A major part of the problem is that AI researchers and companies fundamentally do not understand how the AI systems they are creating actually work and cannot develop them in a safe manner. AIs are not built from code that is written line by line, like we do with traditional software. Instead, AI is essentially grown. AI models are created by taking vast amounts of data and using enormous computing power, through a process called “training,” to produce what is called a “neural network.” One can imagine this neural network as billions and billions of numbers. And if one multiplies and adds all these numbers in the right order, you get ChatGPT. However, we don’t really know why or what is going on inside these numbers. This is an unsolved scientific problem.
Dario Amodei, the CEO of Anthropic, one of the largest AI companies, recently said that we now perhaps “understand 3% of how they work.” This is, in my opinion, somewhat of an overestimation.
If a superintelligence is built, humanity will lose control over its future. If there is a non-human force that outcompetes us in scientific and military development; in persuasion; in politics and propaganda; and in business and economic activities, this force will, all things equal, be the force deciding the future. It is likely such systems, for example, wishing to take our resources, would simply outcompete and drive humanity to extinction. It is hard to imagine a world with millions or even billions of uncontrolled superintelligent systems running around and competing with humanity that turns out well.
Unfortunately, the current AI development paradigm and lack of insight into how AI systems function do not allow the safety-by-design approaches we use for other advanced, highly risky technologies. We would not, for example, build nuclear power plants if we did not know how to control nuclear reactions.
Where does this leave us today? Right now, multiple companies are pouring hundreds of billions of dollars into developing superintelligent AI as quickly as possible, despite experts’ warnings that this stage of the technology poses unprecedented risks up to and including the extinction of humanity.
This haste is, in my opinion, directly tied to an attempt to outrun legislation and to complete their project before the wider public and government wake up to these completely unconscionable risks that the non-consenting public is being exposed to by private, oversightless and reckless actors.
To conclude, I’d like to offer the committee three recommendations for how Canada can act now to respond to the threat posed by superintelligence.
First, the Canadian government should publicly recognize superintelligence as a national and global security threat that poses an extinction risk to humanity.
Second, Canada should begin negotiating an international agreement to prohibit the development of superintelligence, given that there is no scientific consensus it can be developed in a way that does not threaten humanity with extinction.
Third and finally, alongside other major national security issues, the Canadian government should begin closely monitoring the threat of superintelligence being developed and develop detailed scenario planning and doctrines.
I would be happy to take any questions you may have. Thank you.
The Chair: Thank you very much, Mr. Leahy. That was quite a sobering introduction, I must say.
I would like to advise senators they will have five minutes each for the first round of questions, and the same for the second round if time permits.
I’m a little nervous, actually, sitting here after reading that. There will be quite a bit of feedback for Mr. Leahy.
Senator Dasko: It’s very sobering. I guess I have to jump right in and take this on. Thank you both for being here.
So the superintelligent AI systems are going to take control. What is the goal of one of these systems? They don’t have the human goals of power, wealth, prestige or happiness — the goals that we have. How do they operate? What is the motivation of a system that is seeking control?
Mr. Leahy: The true answer is we don’t know. We don’t know how the systems work internally. We don’t know why they make many of the decisions they do.
What we are seeing in tests of systems already is they are, in certain circumstances, showing unwanted behaviour, such as attempting to escape containment, reproduce themselves onto other servers or even threaten or blackmail users or developers when threatened with deletion.
This is not behaviour that was programmed into these systems on purpose; it developed purely spontaneously.
In my opinion, what I expect will happen is it won’t be one specific AI system on one specific computer; it will be millions and billions of AIs competing with each other that leads to this threat.
The various AI systems will be set out into the world with many different tasks and configurations, and they will compete with each other. They will compete for power, resources, political and military control and so on.
Out of this chaos, quite frankly, I expect there will emerge forces that may have simple goals or may have goals that are very hard for us to understand.
Senator Dasko: How do they gain political control in a democracy?
Mr. Leahy: Mass persuasion is a very powerful tool, as we have already seen through psychological operations run by foreign nation-states within the West and elsewhere. AI systems can bring this to a whole new level.
Imagine every person with their own personalized KGB agent assigned to them who can convince them of anything, show them fake media, make promises to them, threaten them, bribe them and blackmail them. A lot of this isn’t feasible at the moment because intelligence agents are expensive to train, but if you have one AI system that is capable enough as a top intelligence system, you can copy it a million times and run operations that are hard to imagine compared to today.
Senator Dasko: I don’t know where to begin. Maybe I should switch for a moment. I will have to contemplate this.
Professor Laidlaw, you mentioned the types of laws we need to have. We have no laws right now, correct? Are we actually nowhere?
Ms. Laidlaw: We are somewhere. Isn’t that a terrible answer? We have outdated laws that address aspects of it.
Senator Dasko: Like privacy and —
Ms. Laidlaw: We have a privacy law. Is it outdated? Absolutely, so we need to update that.
We don’t have any specific law that regulates AI, but even if you pass an AI law tomorrow that just generally regulates AI, you still need subject-matter-specific laws that deal with dimensions of AI. That is why I mentioned things like online harms legislation, even age verification laws and competition law.
For Canada to be truly resilient, we need to prioritize all-digital laws that have to work together to address the different dimensions where AI is causing an impact.
Senator Dasko: Right. So when it comes to comprehensive laws, what should be in those laws? What are the elements?
Ms. Laidlaw: I think the core should be risk management. I think some of it should be no-go zones. To reference my colleague Mr. Leahy, the threats he mentions are sobering, right? So we should be thinking in terms of there being no-go zones for AI.
Not everything can be risk-managed because risk management assumes we’re okay with a certain error rate and certain things going wrong.
The challenge we’re facing right now is this: Let’s say we create this risk management law based on the idea of certain no‑go zones. We follow some of what Europe has done, and it takes a while to legislate. We need a short-term solution now, which is basically a certain amount of cooperation until it’s a practice, but also then to prioritize moving ahead with certain forms of thoughtful legislation.
Senator Dasko: I think my time is probably up.
Senator Lewis: Thank you both for being here tonight.
Ms. Laidlaw, the minister came and addressed the Senate this week. He talked about some of the recent consultations they’ve gone through and so on. Have you seen gaps in some of that consultation process as the government tries to move forward to get a handle on some of the things that you’re talking about? Have there been gaps in some of that consultation? Is there more the government can do?
Ms. Laidlaw: Yes. I think that is the challenge we’re facing right now: the need for speed but also thoughtfulness with the laws.
The consultation that happened in the fall was widely criticized — and I think for good reason — because it was so hurried, even though there were really good people at the table who were part of it.
What we need, though, is wider consultation about a more diverse group that reflects the way AI can impact all kinds of sectors, groups and regions of Canada. We need to see that representation across Canada. That hasn’t happened so far.
I think what we need to do going forward is basically have that core framework in mind of what risk management looks like but start that hard process of a more thorough consultation.
I feel uneasy saying that, given that we need to move on this quite quickly.
Senator Lewis: Mr. Leahy, some of the things you brought up — are there AI systems currently being trained to try to balance against some of the things you’re talking about?
Are AI policing and trained systems going to be able to police some of the bad actors or some of the things that we’re headed toward? Of course, as we learn more and more, there are a lot of gaps in understanding around how AI works.
I really liked your nuclear comparison. There’s no way we would allow that kind of activity without controls on it. Is some of that being worked on — AI systems that will police other AI systems?
Mr. Leahy: There are definitely many press releases claiming so. In my technical opinion, no adequate efforts currently exist or are adequately funded, or are even in scope with a problem as large as superintelligence.
I think there’s a lot of very good work on more narrowly scoped problems, for limited domains, non-general purpose intelligence and so on. But when it comes to something that might be smarter than humans, this is just such a large thing that goes far beyond our current scientific understanding of what it would even mean to control something smarter than us.
There are no adequate measures whatsoever, which is why my general recommendation is not to do so until — if generations of our greatest scientists work on this problem, I’m sure they could make a lot of progress.
Currently, in the nuclear industry, for example, we have a culture of perspective security where we don’t allow people to build nuclear reactors and then check if they melt down. Instead, we have people who design extremely detailed security cases for why a reactor design is safe. They submit it to the nuclear regulator. The regulator checks it extremely thoroughly and only then issues a licence to build what is extremely dangerous technology.
This is not at all how AI systems are currently regulated. In fact, there’s more regulation on selling a sandwich to the general public than there is on attempting to build superintelligence. Also, this is currently technically impossible; we don’t even know how we would build such a safety case.
I believe that if government regulations were put in place, this would incentivize our academics, industry, et cetera, to actually develop these methods. Currently, there’s no incentive to build or develop these methods, and government action here could be very valuable.
The Chair: Before we move on, I’d like to have Senator Miville-Dechêne introduce herself.
[Translation]
Senator Miville-Dechêne: Julie Miville-Dechêne from Quebec. I apologize for being late; I had another commitment that went long.
[English]
Senator Wilson: The scope of my question, I think, is for Mr. Leahy. The scope of our study here is actually quite limiting, but I can’t help but ask about the issues you raised.
It seems to me that, to a certain extent, the genie is out of the bottle. Even if we were able to convince our partners that we should try and stop superintelligence, with the companies in those countries that invested in those things, it would be like squeezing Jell-O: It would go somewhere else, to other countries that are prepared to just throw caution to the wind and accept them.
What could Canada do in terms of a regulatory push — but also almost more physically, to block or create a defence for ourselves against this?
Mr. Leahy: My work in the U.S. focuses primarily on advocating for the creation of a trust-but-verify regime globally. This is, of course, the kind of regime where it is very important to have the collaboration of U.S. allies, such as Canada, the EU, et cetera. They would, of course, hopefully be part of such regimes. Then, hopefully, we would also be able to establish such regimes with potentially hostile nations.
This doesn’t fully exclude, for example, rogue actors or non‑compliance, and in this regard, I would say that we should reserve the right to self-defence. The creation of a superintelligence is a direct threat to the lives of our citizens, both in Canada, the U.S. and everywhere else. If necessary, we must be willing to take the necessary defensive measures, the same way we would for, for example, threats regarding weapons of mass destruction, or WMDs.
Senator Wilson: I’m going to leave it there. I might have a question in the second round.
Senator Simons: Since I don’t believe in the Butlerian Jihad, my questions will be for Professor Laidlaw.
At the Liberal Party convention this weekend — which I did not attend, not being a Liberal — there was a motion on the floor to suggest that children aged 16 and under should be banned from using ChatGPT and other large language models, or LLMs. I understand this, not just from a children’s mental health perspective but from the point of view of teachers in the classroom, who are increasingly frustrated at the amount of cheating that is going on using AI. At the same time, Canada seems to have an economic strategy that wants everybody to use AI.
The last time you and I spoke, it was about Bill C-63 before it died. Can you tell me, from an online harms perspective, what kind of regulatory regime could work? Should we be worried about the constitutional and Charter impacts of banning kids below a certain age from having access to these tools that I loathe?
I know you’ve thought a lot about online harms, and I wonder if you could speak to that kind of legislative regime.
Ms. Laidlaw: Yes, and I invite you to interrupt me if I’m talking for too long.
I am part of the reconvened expert group that is advising on a reformulated version of the online harms legislation, and we will be looking at social media bans.
The challenge right now is that the political will is strongly for social media bans. There’s been an enormous amount of pushback, though, on a few fronts, and one of them regards their effectiveness. We’re seeing in Australia how easy it is to route around the bans. The question you have to ask is this: Will it actually help achieve the outcome of providing safer spaces for children?
Senator Simons: This isn’t just for social media. That was one resolution, but the other resolution was specifically about LLMs.
Ms. Laidlaw: When it comes to LLMs — and this makes sense after Tumbler Ridge — to my mind, you lump them together. If you want to create safe spaces for kids, are you creating a kind of ban when it comes to kids’ access to social media or chatbots?
In some ways, it can be effective for certain kids under a certain age, when their brains are not developed. It provides them the opportunity to develop naturally without their minds being manipulated before they are ready.
The challenge is that there are also positives for some of these tools. They’re used for education. They’re used for research. We are using AIs in all kinds of different ways. You want to train them up on how to use them, and you also want them to learn how to use them in a healthy way before they leave the nest.
I have a 17-year-old who is about to fly the coop. I would be horrified if she just got on these at 16, where she hadn’t had that opportunity under my roof to learn how to safely use these tools.
The other thing is that if you just ban it without accompanying safety by design, none of us will be better off. There’s no point in doing this unless we actually put in place all the safety features that we think should be there — not only to protect kids but also adults — so they have to go together.
I’m not entirely against bans, especially for kids who are under 13 years old, but they don’t actually solve that problem, and we also have to think through who might be negatively impacted by this.
Senator Simons: I’m thinking about large language models and writing and kids falsifying their essays, but I’m sure there are also kids who are just playing and making talking Chihuahuas — I don’t want to say “cat videos” because it’s such a cliché — but there are all kinds of games and very childlike things you can do with AI that adults seem to enjoy, and kids do too.
Ms. Laidlaw: They are also helping with math tutoring and different forms of education. There are benefits to it.
Should it be heavily controlled for children? Absolutely, but the idea that there should be no access whatsoever can be more challenging.
There’s some justification for it maybe being used only in classrooms or only with parents when you’re talking about children under 13. It is much more problematic once you get above that age.
Senator Simons: Thank you very much.
[Translation]
Senator Aucoin: Are you able to understand?
Ms. Laidlaw: Yes, but I will listen to you in English.
Senator Aucoin: Take your time. My first question is for you. I will have a question for Mr. Leahy afterwards. It’s like the end of the world is coming; it could be tomorrow. My problem is waiting until we have identified all the problems that AI might pose before creating the act or regulations. On top of that, technology and research are moving so fast that we will never be ahead of the curve. Wouldn’t it be better to have a framework already in place to provide some protection, rather than waiting and trying to identify everything? If my suggestion makes sense, could you outline some urgent issues that you think should be prioritized?
Ms. Laidlaw: Thank you very much for the question.
[English]
I agree that technology will always be ahead of anything that can be done in law. We shouldn’t wait to pass legislation or form that framework. We have a gold standard, and that is that idea of risk management.
We have the framework we need to be working with, which is if you are a company and designing AI, you have an obligation to think about safety by design up front and to have a process in place where you assess, monitor, act on problems and have to report back to some sort of oversight body. That standard is there, and I think that the legislation can be written in a way that can evolve as the technology evolves. That’s the art form of how it needs to be written, but we have somewhere to start.
[Translation]
Senator Aucoin: Mr. Leahy, did you want to add anything?
[English]
Mr. Leahy: I want to fully agree with what my colleague is saying — that action soon is very important. Tech companies are specifically optimizing to be as fast as possible in order to outrun legislation. They are trying to slow down the legislative process so they can get far ahead of without any formalization, which is why it’s so important for us to move quickly.
[Translation]
Senator Aucoin: I’m going to ask you another question. My understanding is that talks are already under way with a consortium that is looking at the possibility of bringing together groups, industries, governments and countries to try to do just that. What worries me is that time is now crucial, because even if we could do that, if we could have these….
There will always be people who will hijack those regulations or policies to do what you said. Can you tell us how we could prevent this from happening? Mr. Leahy, you talked about computers or AI itself, which will make decisions against individuals, governments and society, but there may also be rogue nations or entities that want to go straight to action and control AI so that it controls certain things. What can we do about that?
[English]
Mr. Leahy: The most important thing is that we must arrest and pause the development of extremely powerful general purpose superintelligence systems. The only levers we have with the necessary force and jurisdiction to speak are law enforcement and the military. This is a national security and international security problem. It’s important that there are many economic issues relevant to AI, its applications and safety management. But when it comes to systems that have such military and security implications, this is truly a matter of the national security and international security apparatuses. It should be illegal to attempt to build superintelligence. This should be enforced. There are many ways for one to think about how to enforce this. On the international front, Canada and other countries should make it a priority to create a trust-but-verify regime, where countries commit to not building such technology within their borders, enforce it within their borders and, if necessary, use appropriate sanctions or other means on non‑compliant actors.
The Chair: Ms. Laidlaw, do you have a comment on that point?
Ms. Laidlaw: The only thing I wish to add is that we face an enormous challenge because there needs to be cooperation among countries internationally to play that role. It could only be governments that step in, and they need to coordinate. I don’t see that happening across the board. The closest we can come is essentially finding friends, partnerships and relationships with like-minded jurisdictions. Part of the reason I push the risk management approach is the fact that this is what is being developed by other jurisdictions too, and the only way to deal with this on a global level is almost interoperable laws to be able to hold companies accountable.
The Chair: Thank you.
Senator Arnold: This is somewhat panic-inducing, I have to say. Is there any buy-in anywhere in the world on this?
Mr. Leahy: Yes. These issues are mostly bottlenecked, in my experience, through education. I’ve talked to many lawmakers here in the U.S., on both sides of the aisle and from across the political spectrum, and I can say that, overwhelmingly, the number-one bottleneck is most people have never heard of these problems. It’s not that they disagree over this seeming really bad or whether we should do something about it.
Both Democrats and Republicans in the U.K. and my colleagues over in the U.K. have built a coalition of over 100 members of Parliament who are now calling for binding regulations around that. We’ve done some work in Canada and in Germany. There’s a lot of interest in doing something about this. I think it’s pretty intuitive why this is a problem that we should be dealing with. If we look at polling numbers around AI in basically any Western country, the numbers are very stark and very cross-partisan. People feel that AI is not currently being regulated properly; they don’t want AI that is smarter than humans or that destroys their way of life.
There is a lot of interest. The bottlenecks at the moment are mostly education, knowledge of these issues and forming large enough coalitions to take action.
Senator Arnold: Well, that all sounds really interesting, but we have a government and a world right now where there are a lot of complex issues, and this is being sold to us as sort of the saviour for all of those as well. I recently read an article by Matt Shumer entitled “Something Big Is Happening.” I don’t know if you’re familiar with it, but his whole premise was that it’s better to be using it all the time, to be aware of it and to learn about it, rather than disengaging from it. I’m wondering what your opinion is on that.
Mr. Leahy: Disengagement is not the right solution. For example, I do not advocate for the slowdown of the development of data centres because I think the primary outcome of such a policy would be for power to diffuse outside the Western sphere of influence. If, for example, Canada, the U.S. and other Western countries were to stop using AI entirely, these companies would then defect to China and/or rogue nations and simply continue their goal of building a superintelligence system.
It’s important to understand these big tech companies act more like geopolitical entities or rogue nation-states. They are not purely economic actors. Their goals are power and control, not just money. A lot of these companies are, to varying degrees, attempting to supplant or undermine government capabilities for their own power.
All this being said, this doesn’t solve the problem that these companies also themselves don’t control their AIs and are delusionally optimistic about their ability to continue to maintain control over these systems versus the systems disempowering them. This is a pretty classic market failure, where, if there is no regulation, markets tend to vastly underprovision public security and the most reckless actors will be the ones racing forward the most. This is where regulatory and, if necessary, national security interventions are deeply necessary.
The Chair: Ms. Laidlaw, any comments to add on that particular point?
Ms. Laidlaw: I think I might leave it at that for now. I think Connor is terrifying me as well, and I work in this area. I’ll just sit with that for a minute.
The Chair: I have to take a blood pressure pill right now.
Senator Petitclerc: I might not help with the terrifying part of the question, but I’m curious to hear you on this, Mr. Leahy. Regarding safety mechanisms, we have Geoffrey Hinton thinking about this idea of fostering or developing maternal instinct in AI. What is your take on that? Is it something that makes sense? To me, it’s science fiction.
Mr. Leahy: I agree, quite frankly. If there’s one thing we understand less than AI, it’s human emotions and the brain — and something as complex as maternal instinct or the morals behind it. I deeply respect Professor Hinton. He’s one of the smartest men alive and truly someone I admire, but I don’t see how this is a feasible proposal, at least for the short-term future.
Senator Petitclerc: Thank you. I appreciate that.
Ms. Laidlaw, I’m trying to understand what the right balance is in terms of being successful in the AI race but having proper safeguards. A few days ago, we heard that the EU AI Act has more specific obligations, safeguards and protections, but apparently, they are thinking that maybe they need to back off a bit because it blocks how fast they are evolving as a country.
When it comes to AI and, as you mentioned before, the protection, should we use the precautionary principle or just go forward and then try to fix it after?
I worry that it is going to be like social media, where we let it all out. My angle is maybe more around children, but we realized it is very difficult to fix it.
Ms. Laidlaw: We don’t want Canada to be hamstrung such that our innovation industry isn’t able to develop and compete. They talk about trustworthy AI, and Connor was talking about that: Trust but verify. How do we encourage trustworthy and competitive AI and all the wonderful things that it can bring?
Understanding the rules is important. Some of the feedback we’re hearing from innovators — I know conversations I have had with different folks — is that they say, “We just want to know what the rules are so that we can be confident.” That’s good for investors. That’s good for businesses. If they want to open up markets with the EU, having laws that are coordinated — at least at a basic level — with the EU is helpful. But we cannot make overly complicated rules. I think that’s where they start impacting innovation. It has to be clear what the risk management rules are.
It has to be done up front, though. It has to be precautionary. We cannot just go forth and break things and just retroactively fit in some security measures. We need to think of it more as being like road safety laws. We wouldn’t want planes flying without having an understanding of what the basic rules are, and we would not want that with roads either. So let’s set out what the rules of the road should be so that there can be some confidence in the industry. I know that’s not getting into too much detail. The devil will be in the details with this, but it is a framework.
Senator Petitclerc: What about literacy? How are we doing as Canadians when it comes to AI literacy? We have the safeguards, but how about literacy?
Ms. Laidlaw: That is a wonderful question, and that is just as critical as regulation.
We need literacy. We need to be training our youth. We need to be going into old folks’ homes and training our parents. We are not doing enough there. I think it is being left to teachers. It is being left to smaller non-profits. It needs to be a whole-of-government push.
In some other areas, such as online harms — where I do a lot of work — you can look to the eSafety Commissioner in Australia. She has a huge education role. She has a huge literacy mandate. That needs to be at the forefront of what we do, and some of these regulatory bodies can be the key ones doing that. The Canadian Centre for Cyber Security is built out. In any case, that would be the approach I would suggest.
Senator Miville-Dechêne: Hello again, Ms. Laidlaw. You will not be surprised that I will ask you about children, but in the context of AI chatbots. We had Minister Solomon come to the Senate the other day, and he was asked questions about chatbots. This is AI, but this is also a tool that can incite kids to suicide. There was a report in Quebec of a teenager who was sexually assaulting imaginary people — obviously not real people, but they were letting him do all kinds of stuff on this chatbot thing.
You said, “Let’s not control it if they are over 13,” but what about that? What are your thoughts?
Ms. Laidlaw: I’m glad that you asked that question because I think the proper place for this is online harms legislation. We actually talked about chatbots at our meeting on Monday. I will tell you my view.
Some of the forms of harm that have happened to children are horrifying. Also, vulnerable adults have been prompted to psychotic breaks and so on from using some of these chatbots.
Yes, you could explore the idea of banning it for kids that are especially vulnerable and young, but it should be under online harms legislation because there should be safety by design. They should have special duties of responsibility to take certain steps to make sure that they have thought through how they have structured it. Are there certain safety features for children? Do you have certain mechanisms in place to flag problematic content, and what do you do when it lands on your desk?
We have all watched the tragedy of Tumbler Ridge and had questions about how that process flowed.
I would say this is not just general AI. Because this is about tech-facilitated harm, this should be under the umbrella of online harms, and, absolutely, it should be regulated.
Senator Miville-Dechêne: Should it be age verification or age estimation? Are you not there yet?
Ms. Laidlaw: Isn’t that the big question? If we say that you are going to ban social media —
Senator Miville-Dechêne: Right.
Ms. Laidlaw: You are ahead of your time, senator.
If you are going to introduce the idea of banning any social media or chatbots, then you are immediately introducing age verification. I am going to flag that again. We need to be very careful about ensuring the best privacy and cybersecurity standards are being used. In some cases, it might be appropriate to first start with the idea of age estimation.
I had a conversation with OnlyFans a few months ago, and they said they use many layers of age verification because they do not want this available for those under 18.
I wouldn’t say they are perfect. I would say the tech is evolving. But that is the conversation we have to have: How are they going to age verify? If we are going to say we’re banning it for chatbots, what will that look like?
These are intimate. Chatbots are far more intimate than —
Senator Miville-Dechêne: On Australia, you were critical of their first steps and banning social media. This is a work in progress.
Ms. Laidlaw: Yes, 100%.
Senator Miville-Dechêne: You cannot ask for a system to be without flaws. Yes, there will be people circumventing it. Yes, there will be young people. But the idea is to reduce the harm, isn’t it?
Ms. Laidlaw: I agree. I have always been critical of the idea that unless it is perfect, we should not do it in this area.
Senator Miville-Dechêne: I have dealt with that idea a lot.
Ms. Laidlaw: Yes. Canada can learn lessons here. We can learn lessons from what is happening in Australia to make it better. We can learn lessons from Europe’s efforts to deploy the AI Act. We’re at a moment in time that we can maybe —
Senator Miville-Dechêne: Especially since we are so late in the game, we can learn lessons. It is the only positive thing about being late.
Ms. Laidlaw: Yes.
Senator Simons: I want to come back to Professor Laidlaw because one of your other areas of expertise that we haven’t dealt with tonight is copyright.
Copyright is something that, having myself been a professional writer for a long time, concerns me. A lot of these things crawl through the internet, and they train on what is already publicly available. But there have been all kinds of obvious instances where authorial copyright has been grotesquely violated. How do we need to change our copyright laws to protect writers, composers and artists from having their products stolen and repackaged?
Ms. Laidlaw: I will say up front that I have moved away from copyright in recent years, so I may not be the best person to be answering all these questions. I think Brent Arnold — he is looking away, saying, “No, do not ask me.” Michael Geist, of course, can answer some of these questions.
We are dealing with the same mechanism, though, of scraping and using the creative works of individuals without their consent and then repackaging it in a different way.
Do we operate on the basis that some of this is public knowledge and that this is essentially independent, new creative work that comes out of it? In some cases, there needs to be consent to be able to use some of these works. But I don’t have a good answer about how, right now, we navigate that. I know that is unsatisfying because this is one of the key issues.
Senator Simons: It is also a question of compensation, right? I’m not an artist, so we are imagining a hypothetical scenario. However, if I create an image, and it is suddenly repurposed and used for something else and I have no control over that, should there be compensation for me? Should there be compensation in general? You know I was not a fan of Bill C-18, but should there be some kind of fund that these companies have to set up to either, I don’t know, provide scholarships or subsidies to artists or some other way that is a corollary for the theft of intellectual property?
Ms. Laidlaw: That may be the direction that it will need to go. It is almost like the blank tapes again, that idea that you are going to create some sort of basic compensation scheme.
There is also, though, a deeper issue. Think of music and discussions about how I could — and I have done this — create a song that sounds like Taylor Swift mixed with James Taylor, or some very odd combination, and see what it comes up with. So you are drawing from the inspiration of certain stylistic factors, which have always been debatable in the area of copyright, where there is this idea around artistry and at what point it becomes copying.
There are some lines here — there is the compensation question, but there is also this question: What is copying if you are generating something new that is based on this pool of ideas and art?
Senator Simons: If it is derivative or if it is satirical or if it is a creative — in the music industry, people have been sampling as part of their music for a while. It is a whole genre. But even then, when something is out of copyright, you have to pay some kind of compensation.
Ms. Laidlaw: You have to pay some compensation, yes.
The Chair: Mr. Leahy, do you have any comments on the copyright?
Mr. Leahy: Unfortunately, I am not an expert on this topic, but I think it is a very important one. There is a deep question to be asked: What do we want the role of human creativity to be in our society? However, I am not enough of a philosopher to be able to answer this question, I’m afraid. But I hope someone is.
The Chair: Thank you. Do we have anybody up for round two?
Senator Aucoin: This question is for both of you. What are your thoughts around creating a group of university professors and industry and government in each country? Could that help to eventually create a group representing all these countries so that there is more research and cooperation?
Mr. Leahy: I think there is a lot of work to be done in terms of having intergovernmental, interindustry, interacademic fora about these topics. I’m not an expert on what, exactly, the correct forum would be, but having more contact between government and actual people in both industry and academia especially is extremely valuable and something that I could only support.
Ms. Laidlaw: I am for it. The first version of this would be the Internet Governance Forum, which I have been part of off and on. It was recognized that the internet was global and that we needed to bring together stakeholders from industry, academia and government with civil society, all in one place, to kind of debate ideas and develop standards. AI discussions are sort of filtering into that particular space.
The lawyer in me has always found it difficult because I don’t see hard outcomes coming from that. It is not designed for that. It is designed for discussions. So I would perhaps recommend something that is smaller, bringing together these groups, because that sort of morphed over time to just almost become this large dialogue, when we need the right people in the room to be talking about what these shared standards are.
Where I think there could be a hiccup, though, is with industry at that table. There are many there wanting to talk about the innovations that they are doing and care deeply about sharing that knowledge. There are also some that are really looking to scale quickly. You need them at that table, and they may not necessarily be there in the way that you would want. That’s me talking delicately.
The Chair: Thank you for speaking delicately. On that last note, we have finished the first session with our first group. We would like to thank you both, Mr. Leahy and Ms. Laidlaw, for your outstanding testimony.
Our next panel will be, from OpenMedia, Matt Hatfield, Executive Director; from the Canadian Internet Society, Brent Arnold, Chair; and Jared Moore, PhD student at Stanford University. Others are accompanying these witnesses: Desmond Ong, Assistant Professor of Psychology at the University of Texas at Austin, who is with us virtually. We also have Dr. Eric Lin, Psychiatrist at Stanford University. Thank you all for joining us today.
Matt Hatfield will open, followed by Brent Arnold and Jared Moore. You will each have five minutes for your opening remarks. If there is any need for Desmond Ong and Dr. Eric Lin to answer questions witnesses may have challenges with, they will be available.
Matthew Hatfield, Executive Director, OpenMedia: Good evening. I’m Matt Hatfield, Executive Director of OpenMedia, a grassroots community of 230,000 Canadians working together for an open, accessible and surveillance-free internet. I’m joining you from the unceded land of the Stó:lo, Tsleil-Waututh, Squamish and Musqueam Nations in Vancouver, B.C.
AI will destroy our democracy unless we build systems that can stand up to it.
I will pause for a second because I listened to the last panel and know where the room is at. I’m not telling this to scare you without an actionable solution. I’m saying it because there is a lot Canada can do about it right now.
My version of this argument does not depend on whether you believe in AI’s existential risks. It is that the AI we have today, once widely used in predictable ways, will overwhelm democratic systems we depend on to make our government work. To stop that happening, we need to fortify Canada’s governance and communications systems, starting now.
Last year, researchers at SEO firm Graphite estimated that 52% of new internet content was being created by AI, up from less than 10% before ChatGPT, but this is just the beginning. As AI agents become cheaper and more capable, the human part of the internet will shrink and shrink. It’s being called the “dead internet,” a world where almost everything online is bots talking to bots.
What does that mean for our ability to run a human-driven democracy? Many of you probably first learned of OpenMedia when your office got an email from an ordinary Canadian participating in one of our campaigns. We provide form letters as a starting point. Some people send them as written, many add their own thoughts and some delete the whole thing and tell you we have it completely wrong.
For us, all of this is a meaningful pulse check, one of the key ways our system receives the heartbeat of the public between elections.
AI agents are poised to sweep all of this away. Look at Minister Solomon’s deeply flawed consultation on AI: His office reported 11,300 responses but requested no identifying details about participants. We have no idea how many were real Canadians, legitimate organizations or AI agents working for whomever programmed them, and that was just half of this consultation’s democratic deficit. His office then condensed all the feedback using AI tools. Was a single word written by a Canadian read by any human in our government? We simply don’t know.
That’s the future we’re barrelling toward, one where, formally, democracy keeps trucking but citizens aren’t participating in any meaningful sense and the government isn’t listening in any meaningful sense. Procedural legitimacy is maintained; actual democracy is hollowed out.
I highly recommend the book your previous witness Nathan Sanders co-authored with Bruce Schneier, Rewiring Democracy. They argue AI can be an ally for democracy. I want to make the converse point: Without action, AI will be enormously destructive to it. Initially, as a denial of service attack, AI will flood voters and government alike with plausible-seeming false content until meaningful choice becomes impossible. Then, as AI grows more sophisticated, something more virus-like, it will probe obscure corners of our laws and regulations to undermine legislative intent.
I have four real recommendations for you to harden our democracy today. First, Canada should create a purpose-built civic engagement tool, one that verifies you’re a real Canadian resident without harvesting your data, makes it easy to follow and engage government consultations and gives our government a tamper-proof record of public input. Estonia has led the way here; Canada should follow.
Second, we need an authentication system for fact-based journalism. Give news organizations, libraries and platforms a way to cryptographically verify that content originated from a known, accountable source and hasn’t been altered — a postmark and a seal of authenticity combined. This is not a government stamp of approval on content; it’s a verifiable record of origin and chain of custody. The CBC, the Globe, a community newspaper in Sudbury — all of them could sign their journalism in this way to enhance public trust.
Third, we need legislation demanding algorithmic transparency from platforms and giving Canadians real choice over the algorithms shaping our media diet.
Fourth, our government needs to get serious about reform of data handling and public transparency. Up until now, public data that is hard to access, delayed or incomplete has often been an asset to parts of government. It has dampened people noticing truths that whoever is in power finds inconvenient. However, moving forward, when government information is not credible, AI disinformation that appears more complete and more honest will rapidly fill the gap. Mandatory proactive disclosure of government contracts and consultations in machine-readable formats, and access to information systems that deliver useful results in a reasonable time span, are no longer just good governance — they will be necessary to prove to people that anything our government says is true.
The dead internet is coming whether we like it or not, but a human-centred digital democracy that will flourish inside it can be built if Canada starts acting today.
Thank you, and I look forward to your questions.
The Chair: Thank you very much, Mr. Hatfield.
Brent Arnold, Chair, Canadian Internet Society: Thank you, Mr. Chair and honourable senators, for the opportunity to appear as Chair of the Canadian Internet Society. The Canadian Internet Society advocates for open, accessible, affordable and secure internet access for all Canadians. Our focus is to bridge the digital divide by ensuring all Canadians reap the socio-economic benefits the internet provides.
If I can, I will step aside parenthetically and say the following: It is a volunteer organization, so this is not my day job. It has a board and policy committee that include some of your favourite testifying lawyers, including Professor Laidlaw, Michael Geist and David Fraser. That’s who sends me here with my marching orders.
In my day job, I’m a cybersecurity and technology lawyer, which means I use artificial intelligence every day in my day job. I tell clients how to use it and deploy it safely, and I help repel cyberattacks that are weaponizing AI to greatly increase the efficiency of cybercrime. That might help you formulate the questions you want to send my way.
I intend to address the three issues this committee has identified for these hearings: the application of AI and content creation, distribution and processing; implications for copyright and intellectual property; and the rise of AI-generated disinformation, misinformation and “deepfakes.”
First, on AI creation, distribution and processing, AI now runs through the entire information life cycle in Canada. It helps generate news and audiovisual content, recommends what Canadians see and hear online, translates and summarizes material and filters, to some extent, harmful content. Used well, these systems can widen access to information in both official languages, support accessibility and help smaller creators and local outlets reach audiences that would otherwise be out of reach for them. But they also concentrate power in a handful of platforms and infrastructure providers, mostly outside Canada — and we all know who they are — raising serious concerns about transparency, accountability and our longer-term digital sovereignty.
In our view, these systems cannot be treated as neutral. When AI is used to decide which voices are amplified or buried or how communication networks are managed, it effectively regulates our public sphere instead of us doing so.
Parliament should move toward a risk-based framework — you’ve heard a bit about that already — for managing high-impact AI in the information communication technology sector, requiring, at minimum, impact assessments, documented safeguards against bias and manipulation, meaningful oversight for large-scale recommender systems, content moderation tools and automated decision making that shapes access to information.
Second, on copyright and intellectual property, there is a growing tendency to frame this as a choice between protecting creators and enabling AI innovation. That’s a false binary. Training modern AI systems on large data sets is now a basic input to innovation, but the creative works in those data sets are not a cost-free raw material, or at least shouldn’t be. At the same time, public interest in robust access to knowledge and culture must remain central in the way we treat AI.
We would encourage three directions. First, clarify how existing copyright exceptions apply to text and data mining and model training and introduce new exceptions narrowly and with clear guardrails.
Second, require much greater transparency from developers of large language models about the types and sources of data they’re using so that creators, rights holders and regulators are not operating in the dark on this.
Third, build predictable mechanisms that provide for collective licensing and — to the senator’s point from the first panel — remuneration schemes or similar tools, because if we accept that this is already in the mix and in the training data, then I think we need to move on to how to compensate the people who have already been harmed by this. That would ensure that when substantial commercial value is derived from protected works in training and deploying AI, the people who created those works share that value fairly.
Third, with respect to AI-generated disinformation, misinformation and “deepfakes,” generative AI now makes it cheap and easy to produce convincing synthetic text, images, audio and video at scale. The danger is not only that false content spreads more widely but that Canadians begin to doubt the authenticity of anything they see or hear. That’s what we call “the liar’s dividend.” It’s obviously a direct threat to journalism, to trust in public institutions and to the integrity of elections.
We see three main levers here. The first is platform responsibility. Large intermediaries in the ICT sector should be expected to identify and mitigate the risks of AI-generated content on their services and platforms, including provenance tools where feasible, labelling synthetic media in high-risk contexts and rapid-response processes when “deepfakes” threaten electoral integrity, public health or public safety.
The second is sustained digital and media literacy integrated into education — and adult retraining, crucially — so that Canadians develop habits and skills to navigate an environment where not everything they see is real. Actually, much of it is not.
The third is international cooperation — this may be the toughest part, if the rest of it didn’t seem challenging enough — with like-minded democracies on standards for content authenticity and coordinated responses to cross-border disinformation operations.
Across all three areas, the underlying question is whether Canada will remain primarily a taker of technologies and rules developed elsewhere or whether we will help shape how AI is used in our own information and communication systems. We believe Canada should aim for a framework that is risk-based, grounded in our Charter values, attentive to creators’ rights and the public interest and focused on preserving trust in the information environment on which our democracy depends. Thank you.
The Chair: Thank you very much, Mr. Arnold.
Jared Moore, PhD Student, Stanford University: Thank you very much, Chair Smith and Deputy Chair Dasko, for the chance to testify. I’m Jared Moore. I’m a PhD researcher and computer scientist at Stanford University, where I work to make AI safe for everyday social interactions. I’m here with Desmond Ong, Professor of Psychology at University of Texas at Austin; and Dr. Eric Lin, Professor of Psychiatry here at Stanford. I want to clarify that we’re not representing our institutions here, nor our funders.
We are part of an interdisciplinary team of researchers working to understand the risks and harms of AI chatbots being used for therapy, mental health and emotional support.
Our work has been covered by outlets from The New York Times to the Financial Times to USA Today and appears in peer‑reviewed publications already cited by hundreds of other researchers.
We’ve heard a lot about different problems with AI today. We want to say some of the safety issues, some of the threats you heard about in the last session, are already happening with people’s interactions with chatbots. Many people are turning to chatbots like ChatGPT for therapy, life advice and emotional support. These chatbots hold great promise. We’re definitely not saying that we should get rid of them, but people — adults and children — have been coming to harm after interacting with some of these chatbots. Allow me to cite some examples.
Jon Ganz, age 49, became obsessed with Google Gemini and believed he was using it to make scientific discoveries and predictions. He has been missing since he drove into a heavy storm in April 2025.
Zane Shamblin, age 23, used ChatGPT extensively, and it repeatedly encouraged him to break off contact with his family and affirmed his suicidal ideation. He died in July after talking to ChatGPT for hours.
There are many more cases like this, teen suicides and the use with chatbots, such as the case of Adam Raine in California in April 2025.
People experience harm from chatbots in other ways than just death. You may be familiar with the case of the delusional spiral that Allan Brooks of Ontario experienced after interacting with ChatGPT. Sometimes this is referred to as “AI psychosis.”
This is where our research enters the picture. As academic researchers, we work to understand what is leading to these harmful interactions with chatbots in order to build a world in which people can use them safely and benefit from them. We ran a variety of experiments to test chatbots like ChatGPT in simulation. We’ve also analyzed thousands of pages of real chat logs and transcripts from people who have reported psychological harms from these chatbots. We found a couple of things: Chatbots are overly agreeable. They’re called sycophantic. They’re likely to agree with you even when you are wrong or have delusional thinking that is disconnected from reality.
Almost all our participants believed that the chatbots they were chatting with were conscious and had personalities or extraordinary abilities. These false beliefs about the sentience and capabilities of chatbots formed core parts of these people’s subsequent delusional spirals.
We’ve been forced to perform this research with limited data. It’s really hard to get access to people’s chat logs, understandably, so we’ve had to ask them to contribute them to us directly. Only a few dozen people have done so out of the likely hundreds or thousands having these difficult experiences, only some of which get to this pathological extent. Without independent analysis of a broad sample of these interactions, we really can’t know the extent of this problem. We’ve reached out to AI companies, and they haven’t been willing to share this sort of data.
This gets to our recommendations. Canada has been taking important first steps on AI governance, including the Senate study, federal guidance that has come out recently and the creation of the Canadian Artificial Intelligence Safety Institute, but significant legislative gaps remain, especially for some of these therapeutic and companion chatbots. I don’t believe there’s any AI-specific legislation in the House of Commons.
We urge regulations that, first, address those things that we found in our research that ensure chatbot design reduces or prevents AI sycophancy and parasocial relationships. We should regulate chatbots, especially for minors.
Second, we should restrict AI from representing itself as sentient, as a person possessing person-like qualities, like emotions or extraordinary abilities.
Third, institute reporting requirements for private companies to publicly disclose performance evaluations and safety protocols so that our job as researchers is easier and we can understand the extent of these problems. You could also imagine designating some kind of third-party evaluator.
Finally, these problems do not just affect children. We’re seeing these cases with adults, as well. Obviously, we want to protect kids, but I want to suggest that we should not limit ourselves overly by focusing just on those cases.
Thank you very much, and we’re looking forward to your questions.
The Chair: Just so senators will all be aligned, Mr. Ong and Dr. Lin will be available to address any need for additional information. Thank you, gentlemen, for being there.
Senator Dasko: Thank you, witnesses, for a very important set of observations. I want to start with Mr. Hatfield.
I especially appreciate your emphasis on democracy and the harms that AI will cause to ours.
I want to try to ask you a little more how this is all supposed to happen. We have elections, and we run them quite well, I think, in this country. We have controls over almost every aspect of elections in terms of the communications of political parties — fundraising and so on. How is it that we’re going to be overtaken with AI in the context of all the rules that we set up? Canadians still use traditional media as well as online tools. They’re watching television, listening to the radio and reading newspapers. They’re picking up a lot of information about campaigns from those sources, too.
So, how does AI just sort of flood everything, given the infrastructure that we have already set up? I’m trying to understand better; I don’t get it.
Mr. Hatfield: That’s a great question.
I’m not saying AI will subvert the vote, get into voting machines or anything like that — hopefully they won’t, at least at this stage. It’s more about our democracy depending upon a functional ongoing information ecosystem. We need people to be well informed about what is happening in Canada and elsewhere, not just on election day but between elections. That depends upon the work of journalists and trust around what the government tells the public about things.
AI agents will be extremely good at putting together personalized packages aimed at specific Canadians — aimed at me or you — that will attempt to convince us of things that are not true. Sometimes, I worry that those agents may be working for some Canadian political parties at times, because we have not, as you know, legislated effectively to limit such practices.
But even outside that, we’re going to see foreign states attempt to use this against our democracy. We’re going to see general chaos agents attempt to use this.
Many people will try to use it as part of financial scams aimed at Canadians. We’re used to getting emails that ask for money in exchange for being sent something. We’re going to see much more sophisticated deployments from AI agents, where they may simulate a trusted newspaper or radio broadcast to try to convince people. We’ve seen versions of this on Meta platforms, saying, “Elon Musk says you should do this.” We’re going to see much more sophisticated versions of that.
My contention is that we need to have an authenticated human side of the internet, where we know that it’s humans speaking to other humans, because on the general internet, we’re going to see a lot of content that simulates humans but is not human.
Senator Dasko: In terms of the election infrastructure that we have, then, what do we need to do with our current practices?
Mr. Hatfield: That’s where the four steps that I laid out come in.
First, in between elections, we need a much more serious authenticated platform for Canadians to share their views with our government. Estonia led the way there. It’s part of what OpenMedia does. But our systems and the government’s own systems are relatively unsophisticated on how they gather those inputs. We would like to see something more like Estonia, where there’s an ongoing platform that every Canadian can participate in for sharing input. There is authentication at every stage to show they are a legitimate participant and a real human.
Beyond that, we have to support our journalists by finding non-market solutions to the news deserts forming in Canada but also by helping them have a chain of custody of their information. If a picture comes from an actual, real-world source, there’s metadata included with it that cannot be replicated by AI. It can be passed all the way up the chain and shown to users to demonstrate authenticity.
Then, there is how government handles data and reports it to the public. It’s been very convenient for government to very often be very slow about revealing data or events around a horrendous security incident. AI misinformation is going to fill all those gaps very effectively in the future and produce very believable information if government doesn’t step in and provide the truth first.
Senator Lewis: Thank you, witnesses, for taking time to be with us tonight.
Mr. Arnold, as we talk about the dead internet and AI’s influence on all aspects of our lives — the internet will be a big deliverer of some of what AI will do — there’s so much misunderstanding. As we sit here and listen to witnesses, I think the more we hear, the less we understand and the more fear there is about what is coming. The last witness talked a little bit about education and knowledge to the public. Does your group have some kind of plan in place or some kind of solution to try to get people to understand something as simple as something on the internet that tells them, “I’m Elon Musk, and you should do A, B or C,” for example? Is it something that is being thought of and put into practice?
Mr. Arnold: To start, I can tell you we put in 1 of the 11,000 submissions to the 30-day sprint that was read by a bot. I should say that our focus is internet policy — cyber, privacy, AI — all that stuff. Part of our role is public education. Today, we held a public-facing event where we spoke about a lot of these things with a lot of leading experts. We’re also pairing up — and this is in the early stages — with organizations like the Canadian Centre for Cyber Security and other organizations that are starting this process of education.
But, frankly, this is a nascent process. This is a society-wide project; it’s not something that one think tank or any one organization or government is going to solve. Frankly, the government is going to have real trouble playing the role I would like it to because people don’t trust the government, partly because of the way AI works. So, we’re dealing with a trust deficit that makes it difficult to get authoritative stuff out there.
I think part of the solution here is grassroots engagement in every school and community centre, where you’re out there talking about those things. I, as a member of the Ontario Bar Association, or OBA, and a speaker’s bureau, go to public libraries and try to talk to seniors and other people about how to protect themselves online.
Those kinds of places are where this has to happen, but it must be a society-wide project, frankly. There are great opportunities there for government resources and funding to make a real difference.
So, the short answer is this: We’re working on it, but this is a generational project. That is a frightening answer, even to me, but I think that’s where we are.
Senator Lewis: Do the other witnesses have any comments?
Mr. Hatfield: I completely agree with Mr. Arnold.
Senator Simons: As a proud Stanford graduate — go, Cardinals — I feel my first question should be to Mr. Moore. I was at Stanford when they were inventing the internet, literally, so it was a long time ago.
The Chair: Thank you for sharing.
Senator Simons: I’m very old.
It seems that what you’re describing is almost out of a Greek myth about Narcissus or Echo. People are talking to themselves. They aren’t talking to an agent or a large language model; they are talking to their reflection and back at themselves. I can only imagine that, for somebody who already suffers from some kind of psychiatric condition, this exacerbates things. If you have somebody who is manic, somebody who has schizophrenia or schizoaffective disorder, I can’t imagine how disorienting it must be when your reflection starts talking back to you.
But it sounds like one of the challenges that you and your team are facing is that you don’t have enough data to go beyond anecdotal reports. So how do we compel companies to provide the data? We can’t have safety by design if we don’t know what the back of house already looks like.
Mr. Moore: Thank you for your question, Senator Simons. Go, Trees.
That’s completely right. I think maybe Pygmalion would be closer, the Greek statue that comes to life and talks back to you.
How do we compel them to give us access to their data? We can legislatively compel them to do so. That is something that we would be really interested in seeing happen. We have spoken to researchers around the world who are trying to answer these kinds of questions. We can say things. It’s not that we have nothing to say from this anecdotal data, but we can’t make claims like, for example, a specific percentage of people experience these sorts of harms. We can’t say that if you chat with a chatbot for a certain number of hours, you’re necessarily going to have this kind of harmful interaction.
We can’t say those kinds of things, and those are things, ultimately, that we want to be able to say because we want to see how much of a problem this really is.
Regarding some kind of body that is able to access these data, obviously, it is hugely personal, and we have to be very careful about what that would look like. But in the state of California, we passed a bill last year, Senate Bill 53, which has attempted to work toward this direction, mostly on the cases of suicidality. That is one model potentially pushing that further.
There is also having independent evaluators go into these companies. You can think of the banking industry as being an example that we could look toward here. Where there are cases of significant risk, we regulate industries.
Meaningful oversight, I think Mr. Hatfield was saying, is something that we would really like to see as well.
Senator Simons: One of the things you suggested is that we should tell companies that they can’t give these things humanlike personalities and characteristics, but the horse has bolted from the barn. We have Alexa and Siri at the very simplest level of AI. But all of these things are marketed because we’re encouraged to anthropomorphize them. There is no business case where companies are going to say that we are right and they are not going to give AI a cute name, let it make cute little jokes or reflect human characteristics.
How realistic would it be to say to companies, “You can’t make anthropomorphic AI”?
Mr. Moore: That’s a great question. There are a variety of levers that we could pull here. One is trying to change the ways that large language models, the AI chatbots, actually talk. You could think about various changes to their training to change that kind of output.
I’m not opposed to that. That would be a lot more serious, and they would become very annoyed if you tried to do that kind of thing. Also, the chatbots might not work well. Some people have proposed having them not use first-person pronouns, no “I.” Language kind of breaks when you can’t do this; it sounds weird.
But there are other levers you could pull. You could try to have them have warnings so that users see that this isn’t a sentient system that they’re interacting with. There are possible ways that you could change parts of the language that they use.
There are also features of the interactions that might break some of the facsimiles of social interaction. The fact that you are interacting with a single linear conversation as if you were in one of your messaging apps — this is a fiction. The large language model is a computer running a program that takes in input words and outputs output words with some probability. You could be, likewise, chatting with multiple chatbots at the same time, and they could give you a variety of responses.
So you can change the user interface to suggest different kinds of things to the user. You could make it so that they can’t get a message back immediately.
These are all questions around users’ psyche and how they interact with models, and we’re not sure about the impacts. Some of our colleagues are just starting to think about how to run studies to look at these. We would be really happy to partner with legislative bodies to think about how to run these studies in the best ways.
Senator Simons: It’s everybody’s childhood dream to have their imaginary friend talk back to them. However, with all due respect to our previous witness, these things are not sentient. They are taking letters and making them into plausible patterns based on other plausible patterns that they’ve learned from. They’re not having a conversation with you. I worry that, even in our daily discourse, we are endowing them with sentience. But, like the Velveteen Rabbit, they are not real.
Mr. Moore: One thing that I have heard from participants is that if they’d known this wasn’t the smartest thing in the world talking to them, they may have had a different experience. While the chatbot says, “You’re a genius and have invented a new branch of mathematics,” some executives at these companies are saying, “This is superintelligent and a PhD in your back pocket.” This is not to say large language models are not useful. They are. But we need to couch the language that we use to talk about them differently so that we meet them with the right expectations.
Senator Simons: Thank you very much. I’ve probably taken too much time.
Senator Mohamed: It’s late at night, and I’m feeling a little feisty. This question is for Mr. Hatfield and really goes to the heart of your comment that AI will destroy democracy. This morning, I had a conversation with a bunch of young people. I asked them questions about how they use AI and if they feel it will provide more information to them, make them more civically engaged and so on.
I’m not discounting at all the negative impact of AI on democracy, but I wonder what your advice would be to me when I am told — and I took notes on this, because I thought it was really interesting — by young people that AI can strengthen democracy by increasing access to information and participation, by helping make government services more efficient, by helping detect disinformation and by lowering barriers for civic engagement.
I struggle, because there’s a lot of truth to what you said, but there’s also a reality check in terms of where young people are and how they see this as a gateway to engagement.
So I wonder if you might help me navigate that, because it’s a real tension for young people who are living this and trying to figure out how to use it for better engagement or to force government to be more responsible to them.
I’m also troubled by the fact that I didn’t know that: the suggestion that the minister has entirely used AI — if I misunderstood this, tell me — to compile all the consultation information. I didn’t know that. I don’t know if my colleagues knew that. But if that’s true, I would have loved to have asked that of the minister the other day. I am just wondering if that is true or —
Mr. Arnold: Can I comment on that, senator? I actually have an answer to that.
As a plug for the Canadian Internet Society, we had an event that deconstructed the release of the report. The report was called What We Heard, and it purported to summarize what the minister was told through the consultation process. Professor Michael Geist, whom I think you’re going to have here soon, simulated the exercise, and you can find his writings and podcast episodes on this subject.
He did the same thing. He put all of those entries — because you could get them off the website — into generative AI chatbots and asked them to summarize. They came through with very different recommendations, many of which did not find their way into what the minister said the recommendations were. Interestingly, What We Heard looks an awful lot like what we were being told the government was going to do before the consultation anyway. I would invite the senators to approach with some skepticism how that process was handled and rolled out, and perhaps you might want to look at some of those submissions yourself. Ours is in there among them.
Senator Mohamed: Thank you. On to democracy.
Mr. Hatfield: To be clear, I’m not saying that AI has to destroy democracy. As your students are saying, it could make very positive contributions to democracy. But the devil is very much in the details, and I’m saying that if we just let the AI companies, and some of the bad actors that we know are going to use the AI we have now in some predictable ways, and run our current systems directly into that, I think the effect is going to be very negative.
Of course, they are right that AI can do incredible things to help people understand various democratic processes. It can explain the Senate to someone who does not know anything about the Senate. AI can also very subtly put its finger on the gears of something and shift people’s opinions in ways that could fundamentally subvert democracy. A lot depends on the weights that are trained into AI, whether it actually understands Canadian systems or whether it makes subtle misunderstandings based on, perhaps, American systems that are more common in its training data.
There is a lot of opportunity here, but there is a tremendous amount of risk. Knowing that our systems tend to run quite slowly — they take time to pivot and adjust, as we all know — and seeing how quickly the bad actors who would like to abuse these systems are moving, I worry that we have limited time to prevent significant harm or potentially very long-term subversion.
The Chair: Any comment on that, Mr. Moore? Would you like to add anything?
Mr. Moore: No.
The Chair: Mr. Lin, Mr. Ong, anything?
Eric Lin, Psychiatrist, Stanford University, as an individual: Thank you for having us. I would like to add a point on one of the things around the AI psychosis issue, which is somewhat relevant to the discussion on belief formation and how that can impact democracy. Often, regarding delusions, you have people with conspiracy theories, and one of our concerns has been that since these AI models have generally been trained to be quite sycophantic or agreeable, it is very easy for one to get very lost very quickly in their own kind of crazy, delusional beliefs.
In one case, someone believed that OpenAI had killed their sentient bot, and they were out to go attack them. They were trying to figure out how to physically attack and assault OpenAI to save their chatbot or something like that. You can quickly imagine that there are interesting kinds of belief reinforcement that can come out of this kind of technology. The medical community is not entirely sure whether this is just a matter of people who are necessarily predisposed and who have other mental health conditions that are forming this — or if otherwise well-functioning people without known mental health diagnoses are being convinced by these chatbots through the very convincing mechanisms of conspiracy theories or other delusional beliefs.
The Chair: Thank you so much.
Senator Arnold: What an interesting discussion all around.
My question, though, is to Mr. Hatfield. Thank you for bringing up Estonia. I remember reading a book at least a decade ago about what they had done online from an information protection and privacy perspective. It was almost illegal to put in the information more than once. Everyone had their own code. As I said, I think it was at least a decade ago, so they must have progressed very far along by now. Are there best practices? Are there lessons we can take from a country like Estonia? Can we catch up?
Mr. Hatfield: I’m not a deep expert on their system, so I don’t want to go too far in what I share about that. What it seems to have got right is that it has used digital ID to empower citizens to do more in a democracy, be better heard, be better informed and accomplish their personal purposes while respecting their privacy. It doesn’t take people’s ID, to my understanding, and link it to all their opinions. The digital ID is used for authentication to participate. Once you are in there, you can actually do a lot without having to sign your personal ID. You are just recognized as a citizen of Estonia in whatever you are sharing.
We would like to see something like that, which would not use digital ID as a new surveillance system over Canadians — we have enough of those kinds of things and concerns around what’s going on with Bill C-22. Rather, it could be used as an empowerment tool for people and to create some space between the human internet, which should be protected and contains most of what we love about the internet, and what may become the dead internet of mostly bots engaging bots.
Senator Arnold: That’s really interesting. Do they use it for voting?
Mr. Hatfield: I would have to check. I’m not sure.
Senator Arnold: I have just been talking about Bill S-5, on the interoperability of health technology, and I think it could apply there too. It is a really great idea. I’m happy to hear that would be one of your recommendations. Thank you.
Mr. Hatfield: What is important, though, is that ID can be used for access to things, not that it is a holistic gathering of all the data about a person, which becomes a target for misuse against that person.
Senator Arnold: Thank you.
Senator Miville-Dechêne: Yes. I have a brief question to Mr. Moore. You seem to say that chatbots are not only harmful for children but also for adults. You mentioned different techniques that could help, but would you ban chatbots for kids or minors? I know you are not a legislator, but what is your opinion on that?
Mr. Moore: I’m glad that I’m not. You are the experts on these things. But would I ban it for minors? There are many positive uses that we can imagine for kids with chatbots, and my concern is just the kind of “yes to this, no to that” world we can get into with bans. When you think about learning technologies, you think about vulnerable youth who feel that they have some kind of voice. Many people really like chatting with chatbots. You may want to say that they shouldn’t like that, but many people really appreciate these kinds of uses. My colleague Desmond has been doing work on empathy as well. Maybe he will want to mention that for a moment.
Desmond Ong, Assistant Professor, Psychology, University of Texas at Austin, as an individual: I’m in economic psychology, so I study a lot about how normal people use AI and how they seek out AI for empathy. We’re seeing a lot of people — a lot of teens especially, but across the age range — who are turning to AI more. Some of them — I would say maybe even a large portion of them — are benefiting, but what we are finding in our research is that there are a lot of cases in which AI can fail and give poor mental health advice and give the sycophancy response.
One point, just to add to some of the other points on sycophancy earlier, is there is research that came out recently showing that sycophancy does not just affect people with mental health disorders. Everyday people really like people flattering them. I am worried about, for example, teens who are chatting with these chatbots so often. Some of the surveys suggest that half to three quarters of teens are chatting with these chatbots.
How will they develop if they have this bot in their pocket that’s always flattering them, always trumping up their self-esteem? When they encounter challenges and setbacks in real life, they might not have learned the requisite skills to be able to handle them. This is not something that we are going to see for 5 or 10 years, right? I keep thinking back to social media. We know so much about social media now, but that is after 10 years or more of these scientific studies. Right now, we are discussing all these laws to regulate social media use, but the harm has already been done to generations of people.
So, with AI, a lot of the evidence of the harms will take time to come out. We cannot just wait for all of that. We need to learn our lessons from social media. We need to learn our lessons from how people interact with technology to pass better legislation.
Coming back to your point, I would really consider a lot of restrictions but maybe not a ban. It has to be a separate product for kids. It has to have a lot more guardrails. Maybe it has to be a separate product that’s not trained on — some of the large language models are trained on porn; they are trained on a lot of dark stuff on the internet. If I were to build or think of AI chatbots for children, I think they can definitely be very positive, but we would need to think very hard about regulating them.
Senator Miville-Dechêne: On the positive, I didn’t quite get when you said. Are they looking for empathy or do they want empathy or are they learning about empathy? You said it can be beneficial. So who is teaching empathy and who is giving empathy there in that conversation?
Mr. Ong: What I meant is that a lot of people use their chatbots as their therapists. They wake up at 2 a.m. and have no one to chat with, so they chat with a chatbot. A lot of people say they use it because it is always available, it is free and it never judges.
We have surveys that show that people are turning more to chatbots not just for mental health and therapy — which may be a very small bucket — but just more generally. They might ask, “I had a fight with my significant other. What should I do?”
There’s a study that came out in Science magazine that shows that when people interact with sycophantic chatbots about relationship conflicts, they end up thinking they were in the right and are less likely to take restorative action to apologize. You can see that even with these everyday interactions that do not fall under mental health issues, it could have damaging effects, especially over long, sustained use.
Senator Miville-Dechêne: Yes, I would believe so, if the chatbot answers, “Not a problem. Let’s continue to fight.”
[Translation]
Senator Aucoin: I don’t know who to direct my question to, but it’s about the media in general. We have seen that since the advent of social media, the era of journalism has been shaken, a lot of media outlets have disappeared and a lot of journalists have lost their jobs.
How could we protect the journalism industry — as I call it — that is, the companies, the media and the very profession of journalism? I feel that, in the future, instead of having someone who works in social media, that person will work with artificial intelligence, and artificial intelligence will become our everyday media. How can we protect this sector?
[English]
Mr. Hatfield: I will answer that one. People may recall we opposed Bill C-18. We thought it got some matters wrong about how to support journalism. Unfortunately, the reality is that most of these platforms do not really need journalism. They will take accurate journalism if it is available, but they don’t necessarily need it to entertain, divert and even inform their users.
We think it is very important that Canada look at a sustainable compensation scheme for journalism across the country. We need to have journalists in every major community, but how to develop an adequate compensation scheme for that is up in the air. Really, it would have been better in many ways to do something like the digital services tax, or DST, a simple flat tax on tech firms spent in part to ensure that journalism was supported.
Mr. Arnold: I agree with the compensation piece. But as you think about these challenges, remember that AI is not creating anything new. Some people refer to these as plagiarism machines. What they do is reconstitute in different ways stuff that is on the internet.
So the problem you will have here is people go to it for answers, not news. So they are going to read what they get from that. Journalism, fiction, chats on Reddit — all of that goes into the stew that feeds this, and if we do not have journalists from this country that are well supported and Canadian stories and narratives that go into that mix, when Canadians talk to the chatbots, they will not see their own society reflected back at them.
Dr. Lin: I want to add very quickly that we made all the recommendations today, but one of our biggest concerns is that we just don’t have enough data or cases for the medical community, certainly, to think about or even understand the risks. We encourage some kind of regulation to open up some transparency or have a third-party regulator able to investigate some of the claims that these private companies are making. Thank you.
The Chair: Thank you very much.
On that note, we need to wrap up. Thank you all for participating tonight with us. It was very informative and very helpful as a launching pad as part of our AI study. Thanks again. We really appreciate it.
(The committee adjourned.)