THE STANDING SENATE COMMITTEE ON TRANSPORT AND COMMUNICATIONS
EVIDENCE
OTTAWA, Tuesday, March 24, 2026
The Standing Senate Committee on Transport and Communications met with videoconference this day at 9 a.m. [ET] to examine and report on the opportunities and challenges of artificial intelligence (AI) in the information and communication technology sector; and, in camera, to examine and report on maintenance of activities or essential services in the federally regulated rail and marine sectors in the case of labour disruptions.
Senator Larry W. Smith (Chair) in the chair.
[English]
The Chair: Good morning. Welcome to the Standing Senate Committee on Transport and Communications. My name is Larry Smith, senator from Quebec and chair of the committee. I’d like to ask my colleagues to introduce themselves.
Senator Simons: Good morning. I’m Senator Paula Simons from Treaty 6 territory, Alberta.
Senator Mohamed: Good morning. Farah Mohamed from Ontario.
[Translation]
Senator Quinn: Jim Quinn from New Brunswick.
Senator Cormier: René Cormier from New Brunswick.
[English]
Senator Arnold: Dawn Arnold, New Brunswick.
Senator Lewis: Todd Lewis, Saskatchewan.
[Translation]
Senator Miville-Dechêne: Julie Miville-Dechêne from Quebec.
[English]
Senator Dasko: Donna Dasko from Ontario.
Senator Wilson: Duncan Wilson, British Columbia.
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’re meeting 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 now introduce from the Dais at Toronto Metropolitan University, Viet Vu, Manager, Economic Research; from Young Digital Leaders of Canada, Dana Cramer, President and Chief Executive Officer; and from Epoch AI, Wing Hin Anson Ho, researcher. Thank you all for joining us today.
Witnesses will provide opening remarks of approximately five minutes, which will be followed by a question-and-answer session with senators.
Viet Vu, Manager, Economic Research, the Dais at Toronto Metropolitan University: Good morning, Mr. Chair, Madam Deputy Chair and other honourable senators. It is a privilege being here.
My name is Viet Vu, and I’m Manager of Economic Research at the Dais at Toronto Metropolitan University. The Dais is a public policy and leadership think tank where we focus on the intersection between technology and our economic and democratic systems.
We field the longest survey of online harms in Canada and have recently released a new study to examine the impact of generative AI technology on the creative sector, adding to a decade of expertise we’ve developed in AI. Much of this work would not be possible without the support of Canadian Heritage, the Future Skills Centre and, most recently, the Social Sciences and Humanities Research Council, or SSHRC, for funding a partnership grant on AI and work quality with the Institute for Work & Health.
In 2023, my colleague Angus Lockhart contributed to the Standing House of Commons Committee on Human Resources, Skills Development, Social Development and the Status of Persons with Disabilities study on the implications of artificial intelligence on the labour market. The year 2023 seems far away now, but what hasn’t changed since then is our view that AI will not cause mass unemployment in the medium run.
That said, AI can still radically change the ways that we work. In a new paper that I co-authored at the Dais, we found that creative professionals use generative AI to augment their work thoughtfully. From copywriters using chatbots to simulate conversations with an editor to illustrators relying on generative tools for minor alterations, creative workers understand both the shortcomings of this technology and where it is strong.
Yet many creative professionals still face risks from the diffusion of generative AI because workers outside the creative sector often use these tools to replace them outright. Using chatbots to “develop news, entertainment, or artistic content” ranked in the top 3% of tasks by usage. Any policy response pertaining to creative and information professionals must consider both of these usage realities.
Supporting creatives means two things. First, access for small, independent creators to understand their rights and take appropriate actions to protect their creations must be improved. Second, lawmakers should take initiative in proactively debating regulatory or legislative issues in the domains of attribution and ownership, instead of letting case law by itself determine the outcome.
Before I continue, I must also mention the fact that many in the sector are non-profit organizations. Through joint research we conducted with a consortium led by Imagine Canada, funded by Employment and Social Development Canada, we’ve found that non-profit organizations are often left out of the conversation when it comes to AI. This also has to change.
Regarding “deepfakes,” our experiment-based research found that the methods used by many social media platforms to label synthetic or “deepfake” content are wholly inadequate. Current labelling approaches produce no discernible impact on users’ exposure to labelled “deepfakes,” which we believe reflects fears of lowering engagement from more aggressive labelling.
What is perhaps more concerning is that, even when more aggressive labelling techniques are used, such as a blackout box that users have to click through to view content, they do not change users’ beliefs about the credibility of the information they see. As a result, policies around labelling synthetic content are not by themselves sufficient and should be seen as part of a comprehensive policy response to misinformation and disinformation. This can take the form of both a duty of care included in online harms legislation as well as new labelling standards that can complement legislative efforts for mandatory labelling of “deepfakes,” as seen in the EU Artificial Intelligence Act.
From my remarks, it may seem that I am against AI, so let me be clear that I believe that AI is useful and will play an important role in our economy. However, it’s still a normal technology that can be used badly. The institutions we have developed to respond to technological disruption are equipped to deliver solutions. As lawmakers, I urge you to be clear-eyed and separate the real benefits and dangers of AI from the hype originating from those with commercial interests to sell it.
The Chair: Thank you, Mr. Vu.
Dana Cramer, President and Chief Executive Officer, Young Digital Leaders of Canada: Good morning, senators. Thank you for having me here today. My name is Dana Cramer, and I am President and CEO of the Young Digital Leaders of Canada. We are a national not-for-profit working to build the next generation of Canadian digital leaders. In this work, we are recognized by the United Nations as Canada’s Youth Internet Governance Forum coordinators, and we also coordinate Canada’s Young AI Leaders Community. Additionally, I am a PhD candidate at Toronto Metropolitan University, where my expertise regards global digital governance.
My remarks today will focus on two areas: first, the issues of misinformation and disinformation; and, second, regulatory opportunities that could stop the unethical distribution of synthetic content, either copyrighted works or pieces that can cause harm.
Each year, the Young Digital Leaders of Canada convene youth to discuss pressing digital topics at the Canada Youth Internet Governance Forum. For the past two years, we have consistently seen youth heavily engaged with the topic of online harms. Through qualitative remarks, we have found there is a high degree of concern year-over-year regarding misinformation, disinformation and an erosion of public trust in the digital media environment. However, where many might believe young people are concerned about this for their own well-being, we have seen the opposite. Canadian youth have expressed their concern regarding their parents’ and grandparents’ low-level capacities to discern synthetic AI-generated media from authentic media.
This might be surprising but not completely. Young people have grown up in and shaped digital environments, making these spaces like a language that is native to them but foreign to older generations. These young people, therefore, are better equipped through exposure and watching digital spaces evolve to dissect synthetic media than others.
Therefore, I would like to take this opportunity to reframe some conversation points about protecting children and young people online from misinformation and disinformation and instead note the importance of taking a broader approach to building up digital literacy capacity among all Canadians, as opposed to only those who are young. However, our youth community has identified “deepfake” porn and non-consensual use of personal pictures as especially concerned with their own personal safety and well-being.
Copyright is becoming a battleground in AI regulation discussions. This is unsurprising, as new technologies tend to battle with copyright issues; the internet was the same way when it was released to the public in the 1990s. When the digital sector changes, purpose-built policies must follow. We have seen this in Canada through legislation such as the Online Streaming Act and the Online News Act over the past few years, and what we hope to see soon with an online harms bill. The passing of both of these online content acts expanded the mandate of the Canadian Radio-television and Telecommunications Commission, or CRTC. However, one regulator does not have the capacity, knowledge and expertise to regulate complex, far‑reaching digital business models. Instead, multiple regulators need to be in conversation and collaboration with one another.
We have seen an opportunity in Canada to fulfill this gap: the Canadian Digital Regulators Forum. The CRTC, the Copyright Board, the Office of the Privacy Commissioner and the Competition Bureau are all members of this forum and part of an international push for regulators to work together in meeting the challenges of the digital environment. While this forum is a positive step, it has no teeth. Therefore, we need to see an act to legislate the Canadian Digital Regulators Forum to allow these regulators to coordinate and enact far-reaching and high-penalty regulatory decisions by coordinated powers. The previous online harms bill held a section for a new proposed regulator to be in conversation and in collaboration with the CRTC and Copyright Board. Instead, though, we must see regulators mandated to be part of a Canadian Digital Regulators Forum with identified abilities in regulating these big tech companies, especially regarding synthetic AI-generated content.
Thank you for having me today. I am happy to answer any questions you might have.
The Chair: Thank you very much, Ms. Cramer.
Wing Hin Anson Ho, Researcher, Epoch AI: Good morning, Mr. Chair and other members of the committee. Thank you for the honour of inviting me. I’m a researcher at a non-profit called Epoch AI, where we investigate how fast AI is changing and what is driving those changes. I was also a contributor to the first two International AI Safety Reports, which constitute the largest global collaboration on AI safety to date, including organizations such as the U.K. government, the Montreal Institute for Learning Algorithms and others. Through my work, I’ve often asked myself this question: What is the single biggest thing I’ve learned about the trajectory of AI that everyone — from governments to the general public — needs to know? My answer is this: There’s a good chance, in my opinion, that AI will outperform humans in virtually all economic tasks within the next decade or two, including things that we do with our hands and our minds.
The implications of this are hard to overstate, so the question for this committee is this: Are we taking that possibility seriously enough? Why do I believe this? How is it possible that AI could outperform humans so soon when AI systems make things up and can’t do laundry and their labour impacts are still so muted? The reason is that we must look not just at what AI can do today but at trends in its advancements. There’s almost no other technology that’s advancing so fast. A decade ago, AI could barely form coherent sentences and couldn’t meaningfully contribute to the frontier of technology. Today, AI writes over 90% of the code at big AI companies like Anthropic and OpenAI. It has solved previously unsolved problems in mathematics very recently and is a key innovation that led to the 2024 Nobel Prize in Chemistry. That is an incredible rate of advancement.
A decade ago, AI was a province of a relatively small community of academics, perhaps just 1,000 researchers. Today, ChatGPT has arguably become the fastest-growing consumer application in history and is used by close to 1 billion people on a weekly basis. OpenAI’s annualized revenue has skyrocketed in the past few years, to $35 billion; that’s one and a half times Air Canada’s 2025 revenue. The investment that powered AI’s progress up to this point is still growing strong.
As a fraction of the United States’s GDP, AI spending today is greater than the Manhattan Project and the Apollo Program at their peaks. Looking at these trends, it is clear we shouldn’t judge tomorrow’s AI by today’s limitations or impacts. We must contend with an uncomfortable question: If we went from incoherent chatbots to systems pushing the frontier of scientific discovery today over just the past 10 years, what will happen over the next 10 years? As I’ve stated, I believe it is very plausible that AI will outperform all of us at our jobs by then. This is not guaranteed, but I think it is a very serious possibility.
This raises profound issues for policy-makers to contend with, and, in my opinion, five stand out as especially important.
First, powerful AI systems could dramatically accelerate economic growth and technological progress — including in medicine — which brings an enormous opportunity. How should Canada develop and adopt AI to capture these benefits?
Second, if future AI systems outperform us across all tasks, most people could soon struggle to find work to pay bills. How should Canada prepare for that possibility?
Third, as AI systems improve, we’ll increasingly delegate tasks to them and, over time, could lose our understanding of how these tasks are performed. This makes it especially important to ensure that AI systems take actions that are consistently aligned with core human interests. How can we ensure that?
Fourth, powerful AI systems could empower a wide range of malicious actors, enabling them to cause harm in ways that weren’t possible before, such as AI-enabled cyberattacks. This has already happened, according to Anthropic. How can we mitigate these risks?
Fifth, the most powerful AI systems could be controlled by a very small number of people, dramatically concentrating power. What should be done to mitigate the risk from this?
I don’t have the answers to all of these questions; they’re fundamentally very challenging, but if we take the rate of AI’s progress seriously, we need to begin grappling with what AI is capable of doing not just today but a decade from now. These could be the most important questions of the next decade, and answering them well is essential. This is true not just within the sector of information and communications technology, or ICT, but for Canada as a whole. Thank you.
The Chair: Thank you, Mr. Ho. We’ll now move on to questions from senators. You’ll have approximately five minutes each for the first round.
Senator Dasko: Thank you, witnesses, for being here today. Before I continue with my questions, I want to offer condolences to the families of the pilots who died yesterday at LaGuardia Airport. As the committee that studies transport, over the years, we have studied rules and safety issues regarding the air and rail sectors many times and know how important they are. I hope something comes out of this terrible experience that will help us move forward.
I am interested in a number of issues with respect to AI. One is the privacy of data. Rather, let’s call it consent to use data. I’m also very interested in the policies and rules we are developing.
I want to start with Mr. Vu and Ms. Cramer. Ms. Cramer, I think you were suggesting that we should have multiple regulators and multiple regulations. Mr. Vu, I think you’re calling for comprehensive legislation. Can you articulate your view and why your approach is better than multiple regulations — or how these approaches might fit together? How should we think about this?
Mr. Vu: Thank you for the question. I should start by mentioning the fact that Ms. Cramer is a long-time collaborator of the Dais. We work closely together on these topics so we’re perhaps not as far apart as our opening statements may make it seem. But from our perspective, when it comes to Canadians’ experiences online, a comprehensive approach must include both a privacy-focused approach and a harms-based approach. What I mean by that is it is not enough to be able to pass a piece of legislation — such as an update to PIPEDA — that adds further privacy protection to recognize the realities of the internet today. We must also recognize that, even with the best legislation possible, Canadians will continue to experience harm; they will continue to require recourse when their privacy is violated, their data is used or they see or are exposed to “deepfakes” online.
It is worth noting here that 47% of Canadians encounter a “deepfake” every week. Labelling legislation by itself will not be able to solve that.
We believe there are a few elements to comprehensive online legislation that include updates to privacy legislation — but also an actual law enforcement agency in the shape of an online safety commissioner. That was proposed in the previous version of an online harms bill. Passing legislation by itself is not sufficient; you need someone to enforce the regulation and keep these companies accountable.
In addition, a duty of care likely needs to be introduced to these social media companies, but I know I’ve droned on for too long, so I’ll pause here and let Ms. Cramer add her thoughts.
Senator Dasko: I want to close the loop on the labelling issue. You said the labelling is not enough, and I understand it’s not enough, but should there be labelling as part of a package?
Mr. Vu: Yes. As we’ve seen in our research, leaving it up to social media companies to label is not sufficient. An example of that is the EU Artificial Intelligence Act, which essentially places two duties upon platforms: first, one on the publisher of the AI systems that create “deepfakes,” which is an approach that has forced companies like Google to introduce a technical watermark to “deepfakes” created using their technical tools; second, there are duties for the deployers of the AI systems to label that content.
I think a duty of care for the platform that hosts some of this content is probably the right approach when it comes to labelling these “deepfakes” as opposed to a duty to report and label every single thing. That becomes an unrealistic burden on these companies.
Senator Dasko: Ms. Cramer, were you about to say something?
Ms. Cramer: Thank you very much.
I would echo Mr. Vu regarding the importance of watermarking AI content.
As part of the Young Digital Leaders of Canada, we have seen a need for the ability for AI to be a productivity enhancer as opposed to having everything consistently flagged with AI. For example, if you have a technical specification as part of a downloaded image or downloaded content that is AI, when it gets re-uploaded, say, to social media, in that moment, we would like to see watermarking for labelling so there can be a clear identification that this is not authentic media; it is synthetic.
However, if you’re using certain AI tools in order to increase your productivity, such as an image for a presentation, for example, or for creating and connecting sentences in different paragraphs or using it as a copywriter, that is an ethical use of AI.
So, part of what we would encourage is differentiation between the ethical and unethical uses of AI. When AI is used for disinformation and misinformation and needs labelling content for online usage, for example, that would be an unethical use and should have that labelling of content. However, when it is being used as a productivity enhancer, that is a more ethical use. It allows us to better complete work with stronger capacities and more competitively.
I can speak to a few other questions if you have any follow-up.
The Chair: We’ll move on to give everyone an opportunity to ask questions.
Senator Lewis: Ms. Cramer, you spoke about the gap between younger generations and older generations in terms of understanding what a “deepfake” is — that younger generations are more comfortable understanding how AI works and both its benefits and problems. Have you done any work on how to bridge that gap? It seems as if a lot of people are confused, at best, and probably frightened, at worst, about some of what is happening with AI: what the future holds for AI, what it means for elections, their children’s futures with respect to jobs and so on.
Has there been any work done or have you seen work done on how to bridge that gap so there’s more comfort for generations that haven’t grown up with AI?
Ms. Cramer: Thank you for your question.
One thing that we’ve been noticing is that different generations will use AI in different ways. For example, young people have been criticized recently for using AI as a therapy tool. You see public reporting about this quite consistently in that there are chatbots used as companions or relational bots. Someone older might use it for their workplace instead.
We need to see more reporting and more funded studies regarding how AI is used by different age groups. I am starting with that because until we know exactly how different age groups are using AI, we can’t create an effective AI literacy framework as part of digital literacy in Canada. Different age groups will have different digital literacy needs around how they learn.
A lot of media and digital literacy training has focused on K to 12 education, with a bit more in post-secondary because, quite frankly, it’s easy to get students when they’re young. I’ll give a shout-out to the Dais and Mr. Vu. They’re conducting work that is trying to ensure that employers are giving more media literacy training at their places of work, which helps with broader generations.
Something we have consistently heard from young people is that they have a lot of concerns about their grandparents and their capacity to understand AI and information fragmentation in that sense.
Overall, we need a whole-of-society approach to digital literacy and AI training, and we also need more studies that fragment age groups instead of making blanket statements. When we get blanket statements regarding how people are using AI, we don’t get to the crux of the problem — and we don’t get to the crux of the solution, either.
Senator Simons: I would ask seven questions if I have the opportunity.
I’m going to start with a severely practical and technical question for Ms. Cramer, and then I have a broader philosophical question.
Ms. Cramer, you talked about the benefits of having a regulatory group that is comprised of people from the CRTC, the Copyright Board of Canada and the Competition Bureau versus having one dedicated agency that is an expert in this. I have to confess that the idea of having one regulatory agency makes intuitive sense to me, so can you explain to us why you think that’s not the right model and why a group that combines expertise from different regulators is your preferred option?
Ms. Cramer: There are a few components of this.
First, when you have multiple regulators, they can build off each other’s knowledge and expertise and view problems differently. For example, the Canadian Digital Regulators Forum, which I discussed in my remarks, did a study investigating the same questions your committee is looking at right now, and they published it in the past year. They all took different views on synthetic AI content — though less so on misinformation and disinformation — but we were able to get a strong analysis for privacy, competition issues regarding labelling, how online news might be circulating and also copyright issues.
One prominent issue that is more technical, to answer your technical question on the importance of bringing regulators together, is that a lot of regulators get folks who will stay in their careers long-term at one regulator and do not have as much cross-fertilization across them. For example, in ministries, in the Government of Canada, a policy analyst will have an EC designation. It is just a policy analyst on that front. However, in a regulator, they will have a CO designation.
These differences in designation are important because they involve different routes regarding career progression. The CO designation ends up incentivizing workers to stay in certain areas of their careers and in just one team. That is not how you build expertise.
An EC designation in ministries incentivizes workers to jump around different teams in order to increase their step and their pay at the end of the day.
So why is this important to your question? When we have regulators with people who are essentially stuck on one team throughout their careers because they are economically incentivized to, that will create stagnation in knowledge accumulation. It also creates high levels of understanding a problem from just one angle. This is why it’s important to put these regulators in conversation with each other — because if they are all using certain designations for how their employees are progressing throughout their regulator and the Government of Canada broadly as career public servants, they will not have the same level of expertise that you might get from areas that have more movement surrounding them.
Senator Simons: That is a really interesting analysis. I don’t think anybody has ever broken it down for me in that granular way. Thank you. Go ahead.
Ms. Cramer: The second element, I would say, is that there is a global push to have regulators working together regarding digital regulations and then for those regulators in like-minded countries to also be in conversation. This also fits into Canada’s current foreign policy of trying to get more allies in like-minded ways. We’re actually seeing this already in digital regulator forums.
For example, Canada, with its Canadian Digital Regulators Forum and with the aforementioned regulators involved, the U.K., Ireland, Australia and the Netherlands are all part of a consortium in order to have the same knowledge sharing. You essentially get this domestic cross-fertilization of knowledge while also, on a more macro level internationally, being able to analyze these large technology companies. These companies are going to try different techniques, in different countries and across different regulators, to see what lands and how they can escape regulatory restraints on their business models. That is where that cross-fertilization helps and why I’m arguing for the Canadian Digital Regulators Forum to be legislated to allow them to have certain powers to legislate together for higher-level penalties — so that regulations are not the cost of doing business for companies and so that there is a safer digital environment.
Senator Simons: I have a broader question. I’m going to start with Mr. Vu, but I would like everyone to contribute if they have a chance.
You said something interesting to me, which was that even when synthetic content is labelled and clearly marked as “deepfake,” certain audience members don’t care. They quite like to look at the “deepfakes.” This goes to what Ms. Cramer said about the credulity, I suppose, of a certain kind of person. It’s not just based on generation and age but also taste. Facebook is forever feeding me fake crap, and the more I try to stop it by blocking it, the more Facebook thinks I have engaged with it, so they send me more fake crap.
What does it mean existentially for our world if, even when you label things, people have a preference for synthetic? When I was a kid, my little brother liked to drink Tang more than real orange juice. If you don’t know what Tang is, it was like a fake orange juice powder that was marketed as what astronauts drank, which is why my little brother liked it.
What do we do about a world where people don’t seem to want to differentiate between what is fake and pleasant and what is real and perhaps not so pleasant?
Mr. Vu: Senator, I am perhaps older than you think I am, but let me start by saying that you are absolutely right. In the experimental work that we have conducted, for those exposed to that aggressive form of labelling, 71% still believed that information in a political post describing how the budget for the Centre Block renovation has doubled — with a “deepfake” image of a glass dome that we put on top next to the clock tower — was credible.
I think there are a couple reasons why. The first reason, which deserves to be stated here, is that people know the difference between “deepfake” and misinformation. In the same study, we asked people to define “deepfake,” and manually labelled their answers to see whether people understood the meaning of “deepfake,” especially in relation to misinformation. The result was that 73% managed to define “deepfake” correctly, and only 13% defined it by confusing it with misinformation more generally.
I think that many people can tell that just because content is generated by AI, it may not mean that the information contained in the content is fake in and of itself.
Second, I think it is hugely concerning that people may prefer synthetic content. However, I am encouraged by recent studies that show that the novelty wears off very quickly, for the younger population especially but others as well. A lot of people, after being exposed to this barrage of synthetic content, end up actually wanting more human-generated content and more human-created art and music. Obviously, it will likely take months, if not years, to fully understand people’s preferences, but I am not despairing over this right now.
Senator Miville-Dechêne: Thank you. I will continue on this path, though a little differently.
You talked a lot about “deepfakes.” I’m interested in pornographic “deepfakes” in particular. I have been working on that in the real world. How do you protect children? You’re talking about the duty of care, which could be pretty much anything. Do you really want children to watch that? Do you use age estimation and age verification? What do you do? Do you think something should be done?
Mr. Vu: To whom was that directed? Me?
Senator Miville-Dechêne: Yes. And, obviously, I am also speaking about using real faces and mixing all that up. All of that is possible.
Mr. Vu: I am very happy to take those questions.
I think a couple of issues ought to be separated and clarified. The first issue concerns CSEM, or child sexual exploitation material. Those materials are illegal under Canadian law.
Specific to children’s exposure to pornographic materials online, in terms of age assurance and age verification technology, we have obviously been following with interest what many other jurisdictions, including Australia and the U.K., have been doing. Our belief is that there is still a technological challenge, especially in preserving people’s privacy while ensuring that the age estimation and age verification technologies deployed are accurate.
We have seen some interesting efforts in California, which has started to introduce age verification and age estimation at the device level. So, when a child buys a device, such as their first smartphone, the device itself is going to be age-grouped so that any outgoing internet traffic from that device can only include age-appropriate material.
However, again, we can’t talk about this specific sort of solution in isolation from all the other solutions of a comprehensive online harms bill. This should be part of a comprehensive policy package as opposed to individual action, if that makes sense.
Senator Miville-Dechêne: Does anybody else have something to add — Ms. Cramer, perhaps?
Ms. Cramer: Thank you. I would echo Mr. Vu.
Regarding age verification policies, it’s a very tricky issue because when you start asking for people’s age in these regulations, technology and telecommunication companies will start taking a significantly higher degree of data from young people and their families more broadly. It is disheartening, to say the least, the ways that young people are having to engage with content that previous generations just did not have to. However, all of us, including children, should not have our privacy restricted through goals to reduce what is seen because, at the end of the day, privacy is not a privilege; it is the rule of law.
With age verification policies, we must be careful if we are trying to stop young people from seeing certain images. Their privacy will be taken for granted, and at the end of the day, most online platforms will likely have a security breach at some point. Those young people’s data becomes cybercriminal data, and it can lead to more CSEM, as Mr. Vu sort of identified earlier.
That is my answer. If you have follow-up questions, please let me know.
Senator Miville-Dechêne: However, at the same time, the privacy of children, with those images, is already breached. The platforms take on their information, so why are we so afraid of new privacy challenges?
Ms. Cramer: We shouldn’t be looking for a race to the bottom around how we can have privacy in certain areas. We need to strengthen privacy as a whole. Regarding, for example, some of the other components of synthetic media, when you also get young people’s data at an earlier age, it creates an unfair advantage for companies to be able to use that for personalized pricing later on, as well as for market research. The goal, again, is to strengthen privacy, not reduce it for certain age groups or for certain goals regarding content viewing, but I will yield to Mr. Vu.
Mr. Vu: I’m in full agreement with Ms. Cramer here.
Senator Arnold: Thank you all for being here today. This is absolutely fascinating. Dana, I thank you for bringing humanity into it. Having dealt with a lot of young people and their fears around their grandparents, I really appreciated that.
We have heard from a number of different witnesses who’ve talked about empathy, maternalism, human joy and how to be a human — and you touched on that too, Mr. Vu — that innate human need for real creation and real humanity, so I appreciated all of your input on this.
My question concerns a slightly different topic, though, and one that is being discussed in a lot of realms right now: AI as public infrastructure. Would there be a way to think about — and have you thought about in any of your worlds — Canada having its own sovereign public infrastructure, with our own Canadian guardrails around it to protect the chain of custody regarding data and that sort of thing?
Mr. Vu: I’m happy to begin. We believe that sovereignty is a complicated topic because the AI value chain is so globally integrated that I don’t think it’s a realistic goal for Canada to be producing the chips and building the data centres to train the foundational models. However, sovereignty can come, we believe, in two forms. The first is by ensuring there is a clear competitive advantage in parts of that value chain. In the submission we made for the AI strategy consultation, we made an argument that electricity is one such competitive advantage that Canada can create, both because we create a lot of it cleanly and because electricity will be required for any AI going forward.
Second, in terms of investment and building AI computing infrastructure and data centres to build public AI for Canada, we believe that should only be restricted to strategic priority areas where it’s important to have a Canadian alternative in case geopolitical tensions — trade or otherwise — mean that our access to global AI algorithms may be cut off.
I think investment in public infrastructure should be restricted to those priority areas, as determined by, let’s say, defence, health and the usual suspects.
Mr. Ho: I would like to add something on the consideration of the costs of different kinds of infrastructure. To add to what Mr. Vu mentioned, essentially, the vast majority of the cost for building AI infrastructure comes from the computer chips themselves. That is why we are getting to hundreds of billions of dollars of investment in AI in the U.S., which is comparable to that of the Manhattan Project or more. This is not something that I think is feasible to try to match in Canada. However, there are certain parts of the AI production process where it is feasible, such as data. Something key that is shifting within AI right now is the use of internet data, from training chatbots to building workspace environments. These are things that Canada is better placed, at least in comparative terms, to try to address because there is software talent here, instead of trying to match, say, the United States or other countries on huge amounts of energy. For example, the largest data centres being planned in the United States will require as much power as New York City. This would be very hard to match. Other things are easier.
Ms. Cramer: I would echo Mr. Vu and reframe some of the elements regarding the sovereignty question. When we talk about AI, AI governance and sovereignty, we tend to use a blanket approach, but we need to ask, “What are we looking to govern?” Mr. Vu noted that AI is complex and uses different systems. There are chip supply chains, data input, energy, compute power and so on.
We need to identify the key elements where Canada would be a strategic competitor in that AI value chain and then find partners around the world, such as those in the Asia-Pacific and on the European continent.
In that regard, energy is likely our strongest pathway. We have significant levels of capacity with fossil fuels — and, more importantly, with green energy, as we don’t want to see these higher-energy-usage systems, as the previous witness noted, require fossil fuels because we don’t have the carbon budget in the world to deal with that. However, Canada is helpful because we have a lot of rivers for hydro power and we have the capacity with modular nuclear power to create more non-emission energy, along with solar and wind.
What that means, then, is that Canada could be a strong place for data centres but maybe not for compute technologies within the AI stack. We need to work with other partners in a broader technology stack for regional as opposed to national sovereignty.
[Translation]
Senator Cormier: I would like to ask my question to Mr. Vu.
I’ve just returned from the summit on artificial intelligence and culture. In your document, which I enjoyed reading, you address several issues. I’m particularly interested in the issue of intellectual property and copyright, as well as the creative sector’s ability to acquire the skills needed to benefit from artificial intelligence.
Regarding the issue of licensing for the use of creative content, one of the many suggestions made at the summit is to develop a national licensing framework and support copyright collectives in adapting their models to AI. I’d like to hear your thoughts on this. I see that you’ve made a recommendation on this issue. However, if you could address it, in your opinion, what content should be prioritized within this framework?
Secondly, there’s been a lot of discussion about digital literacy training. In your view, what key aspects of digital literacy does the creative sector need to make its members better able to benefit from AI?
Thank you.
[English]
Mr. Vu: Thank you, senator. I have printed English copies as well if any of you would like a copy later.
With regard to the question on licensing, there are two things that we have been seeing when it comes to licensing content and the training of these AI algorithms. First, the large companies, publishers and owners of intellectual properties of various brands have the resources and have enough content — massive amounts — to directly create contracts and deals with these companies. We don’t see that necessarily as an issue by itself.
Second, we also see, in some specific arts groups and arts collectives, mainly within the actors guild as well as in the music industry, collective action being taken through collective licensing of a usage of a likeness of an actor, a voice actor or music content produced by members of the same sort of guild, to be licensed and protected together.
We don’t yet see the same thing when it comes to visual creations, illustrations and independent artists. Importantly, for all of these sorts of protections, the access that is granted to smaller independent artists — for example, an actor who may have to sign away their likeness the moment they sign on to be an extra in a commercial — is just not there.
Any protections have to come from the fact that they need to work for both these big organizations, as they already do, and for a collective licensing mechanism that essentially allows the smaller, independent artists to discuss and claim ownership. Whether that takes the form of an investment in a legal clinic; a copyright right; or an intellectual property strategy, as seen in the tech sector strategy that Ontario or Canada has invested in, we currently don’t have a strong opinion on, but those are the issues.
Could you please remind me of the second question, senator?
[Translation]
Senator Cormier: It’s about digital literacy training to ensure that members of the creative sector are well prepared, because that’s one of the major challenges.
[English]
Mr. Vu: Specific to how creative-sector workers use AI, I am not hugely concerned in the sense that many industry surveys, including by Adobe and those within the creative sector, show that these workers use the technology already. I believe more than 50% of creative professionals use generative AI daily, so they have actually already incorporated these technologies in a thoughtful way.
Where we see risks coming and where literacy training is needed is with regard to non-creative professionals employing these tools for creative tasks. For example, for our report, we have a wonderful designer on staff who illustrated a cover that was truly beautiful, but we could have just gone to Sora or any of these image-generation tools, generated a cover image and stamped it on there. That carries risks on multiple grounds. First, there is a potential reputational risk from people undervaluing us because we don’t have the budget to pay for an illustrator. Also, we may potentially violate someone else’s copyright. Maybe a rogue character like Mario from Nintendo ends up on our cover inadvertently and opens us up to litigation.
So, AI literacy should be directed at people who are not creatively trained but who engage in or contract creative tasks. We have developed wonderful programs specifically targeted to non-profit organizations. Ms. Cramer mentioned that we developed some training courses targeted at workplace training when it comes to misinformation and disinformation. Those are models we would be interested in scaling and exploring what that looks like.
[Translation]
Senator Cormier: Thank you.
[English]
Senator Mohamed: Thank you very much. It is nice to see you again, Viet.
I have so many questions. I want to ask about data sovereignty and AI exposure in terms of jobs, but, first, I will comment that I don’t share the same optimism around the ability to take something and say whether it is a “deepfake” versus misinformation. Not now, but could you later give us the breakdown in terms of age and level of education of the 70% who said they can see the difference? That is my first question.
Second, one of the things that keeps me up at night is young peoples’ mental health and what AI, “deepfakes” and misinformation actually do. Dana, I think you said that young people are able to tell the difference. My understanding from CAMH is that such is not the case, so I would like to understand if I am correct on that.
My biggest concern is that as we look at labour shifts and people being worried about not having a job, part of the challenge is that many people now use AI to evaluate whether somebody is suitable for a job, so I want to talk about biases in hiring. Do you have any comment on that? That is something that could become a runaway train.
If there is time, I would also like to understand something else. All countries — or at least most of them — are dealing with this, whether it is around sovereignty, usage, “deepfakes” or whatever the case may be. There are some countries that are further ahead and have had to scroll back; those of the EU might be among them. What countries should we be looking to, to learn from and to avoid making the same mistakes, understanding that this must be made-in-Canada approach and that we have limitations.
Thank you. You can answer in writing, as well.
Mr. Vu: In terms of the pieces on age breakdown, we are happy to provide those after.
With regard to bias in AI, I will respond directly now. I believe that the conversation about AI in hiring and how that perpetuates biases can sometimes be a distraction from the fact that, even without these AI systems, bias in hiring takes place. The conversations should be about how we should be deploying these systems to mitigate human biases. Thoughtfulness comes in here, because if you don’t think things through and just use your already biased hiring framework and introduce AI on top, it is no wonder that it reproduces the exact same biases we have as humans during the hiring process.
However, there are thoughtful ways to use AI that actually reduce bias. People are optimistic. In a new piece of research that we are working on with a queer tech advocacy organization for queer and trans people in Canada, we find that queer and trans tech workers in Canada are very optimistic about AI systems actually reducing bias against them. It is a technical system, and there are very clear boundaries on how AI uses and reacts to these systems. That tends to be much easier than with people.
My answer is to focus on the bias itself and the actual systems that create it as opposed to the technology piece, because technology is ultimately a tool.
Senator Quinn: Thank you for being here this morning. This has been very interesting, if confusing.
I want to come back to Mr. Ho. You talked about some of the things that AI can assist with. It is moving so quickly. I think you mentioned the space of economics and finance, for example. I want to contrast that with the beginning of Mr. Vu’s presentation, when you said there won’t be a big impact on jobs in the near term.
The two seem to be in contrast. I say that because, in the transportation sector, one once couldn’t imagine automated ports, yet they are here today and have replaced workers — who are being paid to stay in the building, play video games and maybe retrain for jobs in AI.
I’m at a loss as to how fast things are going and the amount of data that can be processed through AI. I am kind of confused by your contrasting views.
Mr. Ho: Thank you for the question.
This is a very important question and something that I’ve been wondering about. There is quite a large variance in opinions, especially between people who work directly on the capabilities of AI, such as those in the San Francisco Bay Area, and a lot of economists we talk to in our research.
In this particular case, I think we should ask people questions about their areas of expertise. In the case of AI researchers, it is understanding what kind of capabilities these AI assistants will be able to reach within the next few years. What is indicated there is that whenever AI researchers try to build any kind of benchmark to try to simulate a workspace and see to what extent AI systems are able to perform those kinds of jobs, we find it is very hard. We spend millions of dollars trying to build these kinds of benchmarks. OpenAI built a benchmark called GDPval based on economic tasks to measure how well, as a leading indicator, when AI systems will be able to automate these kinds of tasks. By the time that benchmark was released, after spending millions of dollars on it, it was basically saturated. The AI systems had outperformed the human baselines on these tasks.
What we find as a general trend is that, any time we try to build a benchmark like this, it always saturates extremely quickly — within one or two years. This suggests that the trend is moving extremely quickly, and it is getting harder and harder to find things that AI is not able to do in a test.
So, the question is this: Why, then, don’t we have AIs doing everything already? There are many factors, such as the difficulty of putting all the contexts of different kinds of real work environments into a clean task. These benchmarks don’t measure absolutely everything we care about in the real world.
For example, I have had conversations with people over the phone or in person, and the AI can’t write an email for me about that conversation because it didn’t record it. There is context from the conversation that was not included in the context of the email-writing process. There are also physical bottlenecks where AI systems aren’t as strong as humans. In some domains, AI systems are doing better. They are able to outperform most of us — everyone in this room, probably — at research math, but we are still better at doing laundry.
But if you look at the main predictor of when AI capabilities are going to improve, it is when you’re throwing a huge amount of computation at that problem. For clarity, if we look at the trends — this is our data we’ve been looking at for the past few years — the increase in computation for training these systems has been growing four to five times per year at the frontier. That is faster than almost anything. Compared to things like the cost of DNA sequencing, which is extremely fast — this is faster. If we extrapolate the trend, we’ll get multiple orders of magnitude more computational power thrown at training these more powerful systems. If we compare the trend in performance with the trend in inputs, like compute, then we expect the capabilities will increase. Given they’re building so much infrastructure to try to build these things, it’s not out of the question that we will see massive increases.
In summary, the difference is there are still bottlenecks in the real world that prevent us from seeing huge impacts. Even if AI can do some tasks, there are bottlenecks in other parts of the economy that make it hard. However, if we look at the trends and the inputs, given how these inputs have predicted massive increases in capability in the past, we should expect to see this change, potentially within the next few years. That’s where the distinction comes from: looking at the trend, not just at what AI can do today.
The Chair: Thanks to our witnesses for your participation today. It was very enlightening, and I think I’m going to go dye my hair before the next meeting so I look younger. Thank you so much for your comments and time, and we really appreciate it.
(The committee continued in camera.)