Human Rights

---

THE STANDING SENATE COMMITTEE ON HUMAN RIGHTS

EVIDENCE

---

OTTAWA, Monday, May 4, 2026

The Standing Senate Committee on Human Rights met with videoconference this day at 4:03 p.m. [ET] to examine and report on the impact of artificial intelligence on human rights and economic security in Canada, especially in relation to vulnerable groups and the international human right to work; and, in camera, to consider a draft report and a draft agenda (future business).

Senator Paulette Senior (Chair) in the chair.

[English]

The Chair: Good afternoon, everyone. Honourable senators, I would like to begin by acknowledging this land on which we gather is the traditional, ancestral and unceded territory of the Anishinaabe Algonquin Nation.

My name is Paulette Senior, a senator from Ontario and chair of this committee. I now invite senators to introduce themselves.

Senator Bernard: Wanda Thomas Bernard, senator from Mi’kmaq territory, Nova Scotia, and I am the deputy chair of this committee.

Senator Robinson: Good afternoon. Mary Robinson, senator from Prince Edward Island.

Senator Karetak-Lindell: Nancy Karetak-Lindell, senator from Nunavut.

Senator McCallum: Mary Jane McCallum, Treaty 10, Manitoba region.

Senator K. Wells: Kristopher Wells, Alberta, Treaty 6 territory.

Senator Arnot: David Arnot, Saskatchewan, Treaty 6 territory.

[Translation]

Senator Arnold: Dawn Arnold from New Brunswick.

[English]

The Chair: Welcome, senators, and welcome to all those who are following our deliberations. Today our committee will be continuing its study on the impact of artificial intelligence on human rights and economic security in Canada, especially in relation to vulnerable groups and the international human right to work.

We will have two panels this afternoon. In each panel, we will hear from the witnesses, and then the senators around this table will have a question-and-answer session.

I’ll now introduce our first witnesses, who have been asked to make a five-minute opening statement each. With us in person from Mentor Canada, is Tracy Luca-Huger, Executive Director, and joining us by video conference is Kristen Thomasen, Associate Professor and Senior Chair in Law, Robotics, and Society, Faculty of Law, University of Windsor. I invite Ms. Luca-Huger to make her presentation, followed by Professor Thomasen.

Tracy Luca-Huger, Executive Director, Mentor Canada: Honourable senators, thank you for the opportunity to appear today as part of your study on the impact of artificial intelligence on human rights and economic security in Canada.

As the chair introduced, I’m Tracy Luca-Huger and the Executive Director at Mentor Canada, a national charity working to build a strong mentoring ecosystem and working to close the mentoring gap so that every young person in this country can access the relationships, guidance and skills they need to thrive.

Our message today is simple. As Canada prepares for the economic and social disruption of AI, we must not lose sight of one of the most powerful tools we already have: human connection.

Let me take a step back to contextualize this for you. Mentor Canada has led the first-of-its-kind national research agenda on youth mentoring in this country. We have examined who has access to mentoring, who it has left out and what impact mentoring has on the lives of young people. The findings are clear. Mentoring supports education, employment, skills development, mental health, belonging and social capital.

Young people who had access to a mentor were more likely to report good or excellent mental health and a strong sense of belonging, and they were more likely to complete high school and pursue education. Young people with a formal mentor were also more likely to be employed or studying. This matters because young Canadians are facing a difficult labour market. A new Fraser Institute study released on Thursday found that youth unemployment was at 10% in 2022 and rose to 13.8% in 2025. This is the largest three-year increase on record outside of a recession.

These numbers are not abstract. If AI reduces or reshapes employment opportunities, young people will need other structured ways to build human, social and practical skills for workforce success.

This is where mentoring is essential. Mentoring helps young people develop these essential skills, what I will call their “human advantage”: adaptability, communication, collaboration, problem-solving, self-advocacy, confidence and the ability to ask for help. More than 9 in 10 racialized young adults in our research indicated that mentors help them build these skills, the skills for work and life success.

AI raises a deeper social question. Young people are not only using AI as an information tool. Increasingly, some are turning to chatbots for companionship and emotional support. Used carefully, AI may help a young person rehearse for an interview or practise a difficult conversation, but when AI becomes a replacement for human connection, it can reinforce loneliness, deepen isolation, delay help seeking and weaken the real-world relationships that support resilience, well-being and opportunity.

Mentoring connects young people to community and people. It offers something AI cannot — genuine belonging, care, trust, attunement, accountability, advocacy and human connection. A mentor can make an introduction, open a door, notice when something is wrong, identify an emerging trait or skill, or help a young person navigate uncertainty and gain confidence.

That is why mentoring should be part of Canada’s response to AI. It can support young people as they transition to employment, adapt to changing labour markets and protect their mental health in a time of rapid change. We are encouraged by the federal government’s recent commitment to expand job and work-integrated learning and skills-building opportunities, but employment and skills are strongest when young people have mentoring support to succeed. That’s why we’re advocating for the government to establish and fund a national youth mentoring strategy.

Other jurisdictions have already moved in this direction. The United States has invested significantly — over $1 billion in national youth mentoring infrastructure. France launched the 1 youth 1 mentor initiative to expand mentoring as a tool for school-to-work transition and social inclusion. The U.K. has invested in mentoring vulnerable youth.

Canada should do the same. A national strategy would expand access to mentoring, support quality and safety, strengthen research and impact and ensure that young people facing the greatest barriers are not left behind. It would embed mentoring across education, youth employment, skills development, mental health, newcomer integration and community programs. The OECD has identified that mentoring is not only relevant but essential to youth programs, priorities and departments. A national strategy would ensure that we have intentional, evidence-informed mentoring initiatives cutting across priorities, departments and initiatives. And we need it more than ever.

AI will change how we work; it may change how we learn, perhaps even how we interact with others, but it must not change our responsibility to young people. Mentoring is not the whole solution, but it is proven and practical, and it is a solid human solution. It is one of the few tactics with the potential to support young people of all demographics and backgrounds in all aspects of their lives and should be at the centre of Canada’s response.

Thank you. I look forward to your questions.

The Chair: Thank you, Ms. Luca-Huger.

Ms. Thomasen?

Kristen Thomasen, Associate Professor and Senior Chair in Law, Robotics, and Society, Faculty of Law, University of Windsor, as an individual: Thank you for this opportunity to participate. I’m joining from the unceded lands of the Three Fires Confederacy of First Nations, comprised of the Ojibwe, the Odawa and the Potawatomi. I’m mindful in making this land acknowledgement of the close ties between the AI industry and ongoing colonialism around the world.

My comments draw on over a decade of research on the regulation of robotics and AI, seeking to understand the ways that automation shifts power, advantage and harm. I’d be eager to answer any questions about AI regulation, safety, the relevance of tort law, privacy and surveillance.

For my opening, I’ll share three broad observations.

First, hype and overselling are at an all-time high around what is colloquially called “AI.” Misleading language and mischaracterizations of different computer systems are all too common in public and even regulatory discussions, including in some of your earlier sessions. Any technology should be understood by what it actually does, not what its promoters say it could do some day. If I may urge just one thing in your study, it’s to be meticulously precise. Distinguish between different applications and computer methods. Specify what an application actually does. Make note of features that are merely hypothetical. Derive your recommendations from evidence and not speculation.

AI cost-benefit discussions, in particular, can obscure alarming details. For example, some promoters claim that AI may, some day, speculatively, cure cancer. Meanwhile, research on AI data centres raises alarms about increasing cancer risks, among other concerns. To insinuate, as industry hype sometimes does, that public health is a trade-off we must accept to maybe someday cure cancer is an odious claim made all too often. If your study engages in weighing benefits versus harms, you will set a strong example by distinguishing between speculative and conditional, future outcomes and evidence of current harms, risks and gains. It’s apt that our house of sober second thought critically rejects hype and industry bravado.

I also want to name an elephant in the room, which is that many of the systems you’ve been discussing, especially LLMs marketed for ambiguous or assorted uses, are developed within and reflective of a capitalist economic system. When your study considers the labour and human rights impacts of Big Tech systems, let your analysis be guided by the economic context of how these tools are built. Billions in capital investments have yet to realize a return of profit. The pervasive need to make bigger tools faster to compete for market dominance has led to alleged wide-scale violations of consent, copyright, privacy laws, Indigenous data sovereignty, community norms of sharing within information commons, and apparent recklessness toward training practices, content moderation and manipulation.

The drive for fast commercialization means data sets are poorly curated, if at all. Rather than investing in intentional, legally curated training data, some companies rely on exploitative practices for testing and for output moderation, the harms of which are often borne by people working and living in the Global South, from Venezuela to Kenya to Palestine.

The drive to train ever bigger models motivates rapid infrastructure expansion and global extraction and supply chains adding fuel to the climate crisis. This is all predictable when we look at the economic context of how Big Tech systems are being funded and developed. Law can be a tool for reorienting the profit analysis to account for workers and global human rights. Labour, anti-discrimination and environmental protections are examples. Often, these protections only come after grassroots pressure on governments stemming from years of abuse.

I urge this committee and all our public representatives to act proactively here, shoring up legal rights and protection in anticipation of the foreseeable trajectory of capitalist innovation.

Finally, resist technological determinism, the idea that technologies are inevitable and that law and society play little role in their trajectory. That is patently not true. Refusal of dangerous products must always be an option.

You’ve heard about last year’s consultation, a 30-day sprint to make submissions. That is notably difficult for anyone affected by automation but who can’t easily put aside their lives to draft a thoughtful or collaborative response, such as caregivers, teachers, students, community groups and labour organizers.

I also ask you to look at the grassroots People’s Consultation on AI, which, with longer timelines and informational support, collected 65 additional submissions. Within these, you’ll find echoes of your own questions and concerns arising through your meetings. It’s evident that people want to see government action on AI that is not driven by venture capitalist return on investment.

Thank you for your time. I’m eager to explain anything further in questions.

The Chair: Thank you both for your statements. We will now move to questions from senators.

Senator Bernard: My first question is to Ms. Tracy Luca-Huger. I would like to hear more about your suggestion that mentoring be a response to AI and your ideas on a national youth mentorship strategy. Is that a recommendation you would make to this committee?

Ms. Luca-Huger: The recommendation is an important one because it focuses, again, on that human connection. I will give you an example. A young person can absolutely get information and guidance from an AI tool. When they close that tool, the app, that device, who is around them? We know that youth mental illness and isolation are on the rise as well. How will we continue to foster, train and support organizations to embed mentoring into their strategies, whether in places of work, education or community? How do we train and equip evidence-informed practices to support young people in having those mentoring conversations and those relationships?

The strategy is one that we would absolutely love this committee to support because it’s fragmented. We know that expanding mentoring across this country is important. We know that young people still face barriers to access. How do we integrate that into strategies and opportunities?

Looking at programs, we know from last week’s statement and update that Team Canada Strong and youth employment and skills are important. Work is changing, but we need to strengthen the impact of mentoring along the way. This isn’t the only solution, but it needs to be integrated around clear skills and opportunities for young people to learn, and we need to embed mentoring into trades and skills and apprenticeships. How do we equip those individuals who are guiding and leading those young people so that they have the support they need to get engaged and then stay?

Last week, we heard 100,000 young apprentices joined, and only 34,000, I believe, completed. What if they had mentors along the way to support them through those journeys, through some of those struggles and challenges, to help them feel like they belong in those industries but also within those places of work?

We also need to look at how we will build effective tools and practices, as well as standards for mentoring. This is where Mentor Canada has the knowledge, the training and the resources to equip industry and train professionals to understand what mentoring is and what it is not.

Senator Bernard: My next question, then, is for Ms. Thomasen. In your opening statement, you talked about the link between AI and colonialism. I would like you to say more about that and a bit about who will be left behind if we don’t address these issues?

Thank you.

Ms. Thomasen: Absolutely. There are layers to this, and I’ll touch on them briefly, but I’m happy to elaborate in another opportunity.

First, I would identify that the epistemology of artificial intelligence — and I’m talking in many ways about big tech, like Silicon Valley and other competitor tools — many of these tools are developed with — almost going back to an Enlightenment Europe type of mentality and binaristic thinking. A lot of that is perceived as necessary for the tool, but there are examples out there that show that that is, actually, not the case.

The way the tool is structured and the way that data have been collected are very extractive. There are many good resources out there that elaborate on how many big tech companies have just been scraping the entire internet, which means scraping Indigenous data and knowledge without consent and without following proper consent or community input. In fact, there are even examples of big tech taking language that was collected specifically by Indigenous communities to build their own language models that were done within community and following principles of consent and collection and stewardship.

Then there are data centres, and the way in which data centre infrastructural expansion is progressing has been rapid — almost urgent. In many contexts, particularly in the Global South, this has meant incursions into Indigenous territories or disputed territories in order to establish large infrastructure that also brings no jobs with it — or very few jobs — and many negative environmental and public health impacts.

Over and above that, I think that we can also point to the ethos of some of the Silicon Valley CEOs, I suppose we could say, which is very colonially minded and very imperial minded: If it is out there, it is for the taking, and it’s ours to take. If you haven’t turned it into something profitable, and I do, then it’s mine. That sort of mindset permeates some of the bigger tools that we talk about often.

I will flag one example of a different vision, and that is a language model built by Te Hiku Media, which is meant to bring the Maori language back to Maori speakers and is built within the community and within data collection and stewardship practices in that community.

I’ll leave it at that for now, but I would say that it is multilayered, and it is certainly global. When you asked who will be left behind, anyone who is politically marginalized in the sense that the government can override their claims and relationships to their land, to their information, to their data, to their culture and to their knowledge would be in a position to be threatened or at least — if not already — left behind by the progress of this industry.

Senator Arnot: This question is for Professor Thomasen. If citizens can’t realistically avoid being observed or assessed by automated systems, is consent still a meaningful foundation for privacy protection?

Secondly, are transparency and notice enough, or do some automated surveillance uses require legal limits or prohibitions?

Then I have a question on second round.

Ms. Thomasen: Consent as an individual model or driven by individual choice is going to become increasingly frail because of exactly what you have highlighted. Our ability to consent — on an individual level — to surveillance, monitoring and the use of AI on data about us to draw insights about us is, first of all, nearly impossible. It will potentially become nearly impossible because of the way that so much of our daily lives is collected as data — which was a choice and didn’t have to be the case, and we can imagine a world where that is not the case. But to your point, that is a very important feature.

Also because of the way that different data analysis systems work, data about me doesn’t even have to be collected or be part of a data set for a system to draw out insights about me if I am like others in the data set, such that whomever is operating a system and making these kinds of inferences or assessments would infer that I might have those same features or traits, or in the case of risk assessment tools — which are extremely dangerous — would infer that because I am like someone else in certain ways, that I carry the same level or not of risk.

I think you are right to flag that individual consent is not going to save us. To the point that some prohibition or some resistance and refusal have to be available to us, there are technologies that should be banned, and I would name biometric facial surveillance as one example. Where a system is used to scan a crowd and identify individuals, destroying anonymity, destroying public relations and destroying our ability to go out in public and not feel even a sense or a chilling effect that this might happen, that kind of technology should be prohibited.

Very strong guidance, legal guidance and policy guidance — and also the government can lead by example and not adopt harmful technologies — should be in place as well, as well as thinking about our collective rights, which is something our legal system has not done a lot of, but it is something that I think would be so rewarding to move toward, at least as far as law as a tool to deal with some of these issues.

Senator Arnot: This is for Professor Thomasen, again. In your view, what is the greatest human rights risk posed by robotic and automated technologies? Is it surveillance, decision making, normalization of monitoring or something else?

Ms. Thomasen: I think all of those are included under what I would describe as a very serious risk of further inequality in society. This has a great risk of being amplified by increasing automation of surveillance technologies, but the way in which people experience surveillance is not the same and is not equitably distributed. It’s not that I want equitable surveillance, but I think we need to recognize that all of those worries are legitimate human rights concerns, but they fall under an umbrella of the further stratification of society so that some narrow, privileged group can experience public space how they want to, and everybody else is experiencing potentially quite repressive and chilling surveillance limitations, control and limited access.

I think that is reflected in robotics to a degree, but computer systems that analyze information and go unquestioned in a lot of ways are really amplifying this risk right now.

Senator Arnot: Thank you.

Senator McCallum: Thank you both for your presentations. This is for both advocates.

Ms. Luca-Huger, as you talked about mentorship, it reminded me of my experience with residential school — that break with family, kinship, community and knowledge translation of transient cultural heritage, which is cultural genocide.

For First Nations, colonization was to kill the Indian in the child, and I will tell you, it is true. It is in me. For AI, it is almost to kill or decrease humanity or the human experience in the people who use it. What can Canadians or the government do, I guess, to counteract these effects of AI?

I find it very troubling that we seem to be going deeper and deeper into getting rid of our humanity. That’s the best way I can put it. Can you comment on that? What can Canadians do?

Because someone has to counteract it. For First Nations, we’re gathering, talking and getting back to our culture. What can Canadians do?

Ms. Luca-Huger: Thank you for that question, Senator McCallum. It is an important one because human connectedness is an essential skill. It is what bonds and bridges us. We are social beings. That is an important element.

Last week, I had the privilege of presenting at a conference with Dr. Jean Clinton, a child psychiatrist who homed in on brain development and how brains continue to regenerate, but human connection is a critical piece. We had a great conversation about AI and how that human connectedness is important to not only to build up confidence but also to see potential.

As Canadians, we need to invest in mentoring.

There was a time where we knew our neighbours. We invested in that young person next to us. We need to build this into opportunities and places where young people are, into the fibre of our society and in places of learning and work where young people are. They have a tremendous amount to contribute.

We also know from the World Economic Forum that they also identify that mentoring is a critical element, as well as reverse mentoring. Young people know their lived experience right now and how that can benefit companies in shaping the culture which they are building and those skills that they need in entry-level jobs. How are they getting those experiences when they are being trained and learned to build that social capital? Those are essential pieces: that storytelling, the lived experience, the shared journey and the care and understanding of being seen, valued and heard. They are critical for a young person and for all of us.

That is an important piece for them: to build up courage, to feel that they could take an opportunity to move forward and to learn potential wisdom and where industry is moving versus what a chatbot is telling them about always feeling good. The chatbot always says, “Great question. You’re correct. That was an interesting way that you framed it. Let me shift that. Let me change my response to you.”

How are young people having those relationships and building the trust and the safety to talk about their fears and to build up skills for tomorrow?

Ms. Thomasen: Computer models that analyze data can be useful for analyzing data. One thing that is happening right now, and this has a lot to do with the broader context in which these different tools are emerging, is this marketing push — an effective one — to imagine different tools, like ChatGPT, for example, can be used for all kinds of different things, but it is a mode of data analysis.

One thing that the government can do — it is a small thing, but I think it would be important — is to reject the rhetorical effort that is being made to insinuate that something like a chatbot can be a therapist, give legal advice or be your confidant or your romantic partner, and, instead, identify and name systems as they are, which are different types and models of data analysis systems.

AI tools, or what we would call AI, can be useful for certain tasks. They should stick to those tasks. Like you said beautifully, there is no reason that our human connection should be severed or, especially, that we should be encouraging the further industry growth of technologies that sever our human communication through tools simply because they are made available.

I think we need to be more willing to refuse certain promotions or pitches in terms of what these tools can look like. I worry about this with my kids. I hear, “We need to train them on chatbots in school so they understand what they are and how they work.” But I also wonder if we don’t need to promote and give credibility to tools like chatbots in that particular context and be more thoughtful in terms of what this is. What is this system or model useful for, and what is going well beyond that into just trying to get more user interaction and data input from the people using it to further build and refine the tool? Honestly, you have heard this from previous witnesses that the vision that some developers are headed toward is quite terrifying and unnerving. I don’t think we have to go there. I don’t think that is inevitable at all.

Senator McPhedran: Thank you to the witnesses for helping us to spend some time focusing on youth.

This probably could not be considered an optimistic question, but I think it is worth exploring with you.

Many of the constraints that have been recommended in your presentations today are likely not realistic. The capacity of lawmakers to stop or remove practices already entrenched and well developed is part of what we’re struggling with here. There really does seem to be momentum, and a lot of the governmental options have been to actually encourage the growth of this industry. This is being seen as essential for our economy and our viability in the larger international context.

I’ve heard a number of points raised about various kinds of sovereignty, various kinds of protection of individual data and also collective data.

For the constraints that some of you have recommended, to say we should just stop and forbid that, when it is already being done, if it turns out that governments in this country aren’t capable of actually implementing — it is one thing, of course, to pass a law; it is another thing to implement that law. It is one thing to state desirable outcomes; it is another thing to achieve those outcomes.

I have talked a fair bit in placing this question, but, for me, it comes to the heart of the kinds of compromises and trade-offs that we’re forced, over and over again, to come to as legislators, as lawmakers. I wanted to share the question with you: What do you see as essential protections? What do you think are the most achievable? In a somewhat separate category, what do you think are the most crucial? It is open to all of our witnesses.

Ms. Thomasen: That is the question. The reality is that you can have ideas, but how do you implement them? I have a few thoughts. I will touch on them.

I would love if I could walk away and have left one or two messages, one of which is that law can always play a role in shaping the trajectory of technology. I know that it’s difficult because law is not the only thing and that economic policy and social policy come into play, in particular for a federal government.

Acknowledging that, I would emphasize that we don’t have asbestos in residential homes anymore. We no longer permit lead paint. Dichloro-diphenyl-trichloroethane, or DDT, pesticides have been banned. I live in Windsor on the border with Detroit. The car seats we have access to here in Canada are different from the United States because the safety requirements here are substantially stronger. The whole design of automobiles changed significantly following a series of major lawsuits that then led to a regulatory agency in the United States developing regulatory standards that required cars to be built with certain considerations for pedestrians in mind. Up until the 1950s and 1960s, that was not the case.

We have an opportunity to be a bit more proactive. Obviously, I acknowledge that we don’t know exactly the trajectory of the technology, but where we see real risks, and in particular evidence of actual already ongoing harm, that is a place where law could be proactive. Instead of waiting for the many class‑action lawsuits that have led to the development of regulatory agencies and changes, there is an opportunity to be proactive, learning from lessons in the past but also thinking about the context in which these tools are built, and using that to help guide some of the legal response.

Now, I recognize that sometimes some things are in use, and it can be difficult to shift and change. I think that is where — and I know I’m speaking as a lawyer to other lawyers — but when we embrace law’s normative role, that can be valuable as well.

When you ask specifically what would be the most important thing to do — maybe this is odd for somebody who studies AI regulation — but I think focusing on general regulation and law, protections of rights regardless of which technology is involved, amping those up, shoring those up, having more protection for workers, labour, especially workers in a gig economy, and enhancing privacy rights. We already discussed consent, maybe rethinking how we require consent when tools are deployed into public spaces or collect massive amounts of information. I do not want to go on too long, but I do think there are ideas for how lawmakers can do something right now even though we already see these technologies in play.

I think that it is, to some degree, to the industry’s advantage to keep pushing the idea that we can’t do anything about this because they want unfettered growth.

I think law is not the only tool. It can’t be the only tool, but I do think it has a good normative role to play at this moment. We have a lot of insight that we can draw on.

Ms. Luca-Huger: Thank you. I will preface around the law and legal elements for mentorship, but we need to invest. We need to invest in the human connected piece. We know that AI is going to impact careers for young people and lives. How are we going to build in the value of human connection and that teaching and that learning? We know that the industry continues to speak about the skills that young people might be lacking and some of the skills gap. That problem-solving, the ability to look at communication skills and the ability to pivot. Those pieces are important elements. We need to invest in the human component and human skills, those essential skills that will allow young people to thrive.

We are not a regulatory organization in that sense. But we need to counterbalance it with the human component of this and where humans can share not only skills and conversation but that mental health and well-being. Those are the critical elements for us as we look to further advance society and well-being. We need to hone in on that human component. Those are critical elements for society and young people to not only engage, to feel that they belong, but also to have a place around skill development and the support they need to thrive.

Senator Bernard: My first question in the second round would be for Professor Thomasen. Could you expand on your cautions around biometric facial recognition technologies and how that form of surveillance is linked to systemic anti-Black racism and racial profiling, especially in the justice system, and if there has been research done in this area, if you could flag that for us as well, please?

Ms. Thomasen: Absolutely. I know I have said this already, but this also has layers to it.

At sort of the most level of specificity, a number of facial recognition tools have been shown through research — and I would be happy to include citations and submit it to you — to reflect biases in a data set. Some groundbreaking research happened, probably seven or eight years ago now, to identify that many of the major commercial tools — which were the tools that the researchers had access to — this is Joy Buolamwini and Timnit Gebru’s study, Gender Shades. The tools that they had access to study and test had been trained on predominantly white faces and pictures of white faces.

They reflected anti-Black bias and, in particular, misogynoir because they were most inaccurate for black women. That was a groundbreaking piece; unfortunately, because none of these commercial developers had thought of that and considered whether to even test for that, let alone train for it.

I do suspect that many developers have now adjusted how they build their data sets and train their tools to attempt to mitigate that, but we continue to see examples of facial recognition misidentifying people. In Detroit — my neighbouring city — in many instances, they are misidentifying black men and women. Also, I would say, these tools are incapable of understanding non-binary gender identity. To layer onto that, anybody who might already be facing some of the racial bias or discrimination in these systems who is non-binary might face added layers of challenges when encountering these systems.

At the layer of specificity of the training set, the data used to train these tools, there are known issues. I think we can also scale back and look at known issues and discrimination around profiling and policing.

If somebody is more likely to be scanned by facial recognition systems, then even the best ones — if they have errors at all in the system — will be more likely to produce errors for people experiencing more facial recognition, and people might not know that that is happening to them. This is a difficult one to prove. I would connect it to existing systemic bias and discrimination in terms of how surveillance is carried out. It is not just in policing. This could be commercial. It could be in other contexts.

I apologize, I know I had a third point in my mind, and I have now lost it.

Oh, immigration. We have examples in Canada where it has been flagged that there are worries and concerns that people were thrown into life-altering processes in the process of attempting to immigrate and settle in Canada. It has been suggested that that might have been because of misidentification, in particular where the claimants are Black women.

So I think that there is research out there, and we can layer this into what we know from other research about how surveillance operates in society.

Senator Bernard: Do I have any time left?

The Chair: I have to split the time remaining. I need to hold you there and now to Senator Arnold.

Senator Arnold: Thank you both for being here today.

I love this discussion throughout all of the committee work around AI. It is coming up constantly. I keep writing down how to be a human and how we need to almost be intentional about that right now.

I recently learned about this great youth-citizen assembly called Gen Z AI. They looked to AI chatbots, information technology, data privacy and age assurance. I believe what you say, Professor Thomasen, as far as we need general recommendations in law. Some of the recommendations they made regarding AI and age assurance were interesting to me.

I would like your comments on them.

One is to mandate that any AI platforms accessible to children, including educational contexts, implement safety by design so ensuring that, before anything, it is proactive rather than reactive. They also said we should introduce measures to minimize young users’ exposure to harmful design features across digital platforms, including by increasing transparency measures, requiring digital safety and compliance reporting, and requiring AI-generated media content to be identified through digital watermarking. We’ve heard that many, many times.

Some people think it is effective; some people think it isn’t. But, I mean, it is a tool, perhaps.

Thirdly, reducing sycophancy and positive reinforcement in digital AI chatbots. I wonder what you think of those things.

Ms. Thomasen: I think that the approach is exactly right. Age assurance issues — where somebody has to upload an ID, and then if you are above a certain age, you can access a platform, and if you are below, you can’t — engage with serious privacy concerns. When we have to share our identification with platforms that we don’t trust, and that we’re saying are unsafe for young people, and now we’re telling everyone that they have to give that platform — which is probably based in the United States or outside of Canada — our government ID to prove our age has so many layers of difficulty.

Also, if we’re saying that these platforms are unsafe for children under a certain age, they’re unsafe for all of us. Though they may be more unsafe for vulnerable children, I absolutely agree with that.

If the issue is the recommender algorithm brings us into more and more deep violence, then that is unsafe for everybody. I really appreciate, and I would give a plus-one to these recommendations in the sense that they focus on regulating the platform. There’s no reason, in my mind, that AI gets special treatment as a product or service compared to anything else.

I would nuance it a little, but to oversimplify, we would not say that we’re going to start regulating drivers and how they sit. We’re going to regulate cars to be safe. That has been imperfect, but there have been definitive efforts. That approach makes sense, to me, with platforms. The platforms should be responsible if it is putting a product out into the world to ensure that that product is safe, and it should be safe for all the users that they allow and market themselves toward. Again, I have kids. I know that a lot of these platforms actively market themselves toward children. That should bring with it a lot of responsibility.

Senator Arnot: This question is for Ms. Luca-Huger.

If AI tools are used to provide career guidance to young people, what safeguards are needed to ensure that they don’t simply reproduce existing class, gender, race, disability or regional inequalities?

Ms. Luca-Huger: I don’t know if I have an exact answer for that from a technical standpoint. My colleague co-witness probably has the technical piece.

What we do know is that young people using those tools will have some advantages and some advice. But the human component around is that real, is that the reality on the ground, I think, is a critical one.

We know from our research that young people that had access to mentors were also able to navigate social injustice and racism in the workplace or in their community. How do they have that safe, trusted individual to have that conversation with? Is this real? Is this the reality of what I’m facing? How can I navigate it to call it out and to name it? That human piece is a really important one.

We know misinformation and bias exist in those systems. How do we have that human element to have the conversation and to help navigate?

Senator Arnot: Professor Thomasen, do you have any comment on that?

Ms. Thomasen: I would take this opportunity to say that big tech-type bigger and bigger models are not the only way to build tools. So I won’t speak to the mentorship aspect of a tool because that is not my expertise, and I defer to my co-witness.

In terms of envisioning possibilities: small data sets — intentionally and carefully curated —with proper data that has been reviewed by experts, and then used to train a system to produce hopefully better outputs that are then rigorously tested before being shared with end-users. With constant oversight, that is something we could imagine if we think there is a circumstance where an automated computer system might be helpful to provide someone with something that they need or that is useful to them.

I would layer into this that there are other ways we can imagine this. It does not have to be the ever-bigger, ever more profitable model approach.

The Chair: Thank you.

Senator Karetak-Lindell: Are we already creating inequalities? I am reading some notes about the basic and paid AI models that people can pay to get better advice and better information. Is that creating inequalities? Thank you.

The Chair: That may be a question for Professor Thomasen.

Ms. Thomasen: At its access point, yes. There is the reality that people who have access to a better — as in more accurate, better trained, perhaps, or better designed tools will have access to better outputs.

However, the bigger inequality will be the impact that the growing use of these tools have on things that have been discussed in this committee, like jobs, entry-level training and opportunities. If we think about economic inequality, access at the front end to the tool is definitely inequitably distributed. However, I think there are broader, troubling ways in which these tools also promote further economic inequality that then reverberates into these kinds of circumstances where a student or a person doesn’t have access to the same quality of tool.

What I’m saying is, yes, at the front end, but where I think your study’s focus would produce the most benefit, in terms of addressing those inequalities that are being created, is by looking at the bigger picture of why does somebody need to rely on this tool. Is it for school? Why is there this inequity in access to a chatbot that gives statistical language answers? Why is that benefiting some students and not others? What does it mean, also, in a broader sense, for how the people having these different levels of access are going to access employment, work and skill development, but also joy and curiosity and engaging in their day-to-day lives, feeling fulfilled and wholesome? You are right to question it, but also I would encourage us to continue to expand the scope on that question.

Senator McPhedran: This is a question on mentoring.

In order for learning to be effective, the mentor has to be highly effective, with a lot of expertise. How would you recruit the number of mentors who are going to be needed to follow through on your proposal to us this evening?

Ms. Luca-Huger: Thank you for the question. It is multi‑faceted. Depending on the context of that mentoring relationship and the goal of it, mentors would need to be recruited in different forms.

Imagine a Canada where young people will be mentored and you will mentor. That is a community piece where we need to embed mentoring along the journey. It needs highly skilled and trained mentors. For further context, if it is a mentor helping to support a young person to complete a job application, that might be shorter term and a technical skill of a mentor that they will offer, but they need to understand their role and the impact and the safety around it. For mentoring, it is multi-dimensional, depending on the context. Not all mentoring needs to be long and enduring. It could be short term.

Think back to some of the mentors in your own life that you had. Some of them maybe are still there today, and others were there for a brief moment, but the impact might still be the same. How do we equip them with evidence-informed practices and tools? How do we equip industry, places of work, education facilities, opportunities and environments with the right tools to implement mentoring those effective relationships for the context?

Yes, it is large in scope. Some will be programmatic, but some of it needs to exist in the lives of young people along their developmental journey, and those can look very different. We need to ensure that we equip young people with evidence-informed tools and the right information, and to build out the mentoring experiences that young people need and want.

The Chair: To have a follow-up on that. How have other countries that you mentioned earlier developed frameworks, and how are they approaching it?

Ms. Luca-Huger: The United States has a long-standing investment in youth mentoring. I referenced over $1 billion. They’ve had that priority in place since 2008. Not only are they supporting research in mentoring, they also support organizations similar to ours to help build out those standards and practices, and to build capacity across sectors to implement programs and support young people through those mentoring relationships. That’s critical.

They’ve also created a very strong investment in youth mentoring programs, where youth mentoring can apply for mentoring dollars to expand, understand and implement effective evidence-informed programs. They are multi-faceted in their investment. Not only to help build the infrastructure and the resources, but also to support mentoring for what it is — versus it being dressed up as something else. We need to be intentional in that approach with those human supports that young people will need more and more.

The Chair: Thank you. We have come to the end of our first panel. I sincerely thank you both, Professor Thomasen and Ms. Tracy Luca-Huger, for your presentations today. Your assistance with our study is greatly appreciated.

We will turn now to our second panel. With us by video conference, please welcome Katie Szilagyi, Associate Professor, Faculty of Law, University of Manitoba, and Suzie Dunn, Assistant Professor, Schulich School of Law, Dalhousie University.

Professor Szilagyi, please proceed.

Katie Szilagyi, Associate Professor, Faculty of Law, University of Manitoba, as an individual: Thank you for the opportunity to be here today. I join today from Treaty 1 territory in the homeland of the Red River Métis, and I echo Professor Thomasen’s call for an acknowledgement of the extractive colonial aspects of AI technologies.

Conversations about AI in the workforce emphasize the efficiency enabled by generative AI, or gen AI. Despite gen AI’s propensity to hallucinate untrue information, people increasingly rely on it as a research tool. Well-meaning workplaces have put in place AI-use policies that mandate that a human remain in the loop of any important decisions assisted by AI. Yet decades of research on human-automation interaction have demonstrated that humans slip into induced psychological states when asked to monitor the behaviour of automated systems.

Users exhibit automation bias where they are asked to delegate their own authority to an imperfect system, allowing the system to generate both omission errors, being things that are missed, and commission errors, being things that are done improperly. In concert, users exhibit automation complacency when forced to perform multiple tasks simultaneously, relying on the system over their own instincts. In combination, these well-known psychological phenomena lead to overtrust and overreliance on machine systems.

The human rights implications of automation bias and complacency are significant. Social science research confirms that automated systems are typically deployed first in low-rights environments where users are less likely to agitate for their rights, before expanding to other sectors of society. Meanwhile, workers are receiving significant pressure to adopt the latest gen AI tools. Doctors have promised efficiency with AI transcriptions. Lawyers are promised research and drafting expertise from agentic AI assistants. Police officers are being heavily marketed for AI-enabled software, such as Axon Enterprise’s Draft One, a gen AI tool that writes automated police reports based on an officer’s body-worn camera footage.

Independent research into Axon’s software shows it has meagre oversight features that intentionally obfuscate how the system works, making its operations almost impossible to audit.

In the law enforcement context, reliance on outside technology introduces corporate interests to the exercise of state power. This dichotomy distinguishes insider and outsider tools, asking how we should limit the outside influence realized on internal emanations of the state, which traditionally required legislative oversight.

In my work, I characterize this sort of issue as a rule-of-law problem. Our democratic system is designed to constrain instances of power, but if power is wielded by inappropriate actors or augmented by certain technological affordances, it transforms law’s democratic authority, conceived by all, into a technological authority exerted by some. In other words, it transforms the rule of law into a rule “by” law.

Additionally, overtrust and overreliance on generative AI software impoverish human futures by impeding critical thinking skills. Like social media before it, individual humans are fighting a losing battle against the entire Silicon Valley technology complex to keep their eyes glued to the platform. In generative AI’s effort to maximize the user’s time spent on platforms, it exhibits sycophancy, a tendency to be effusive, warm and pleasing to users.

Studies confirm that sycophantic gen AI systems are 40% more likely to affirm users’ incorrect beliefs. This effect is most profound when messages express feelings of sadness. This has concerning impacts for vulnerable groups, especially in a mental health context.

Previous testimony before this committee has highlighted AI in employment scenarios to screen for resumes, including bias‑induced discrimination in hiring procedures, as institutions replicate historical injustices, like predominantly male, predominantly white applicants being most successful. Even when systems are designed to avoid race or gender as criteria, they found workarounds with other resume information known as a proxy data problem.

New research on gen AI software to screen resumes has supercharged this effect. In addition to preferring white, male names, AI has its own self-preference, which is even platform specific. ChatGPT prefers the resumes generated by ChatGPT. DeepSeek prefers the resumes generated by DeepSeek, and so on.

For the future of work, especially through a human rights lens, it matters significantly that generative AI tools are transforming the nature of human contributions.

Thank you. I look forward to your questions, and I would be happy to answer anything about how AI is disrupting law and legal practice, its rapid adoption by government, the private sector and higher education, AI governance and critical analysis of AI using a human rights lens.

The Chair: Thank you, Professor Szilagyi.

We’ll now go to Professor Dunn.

Suzie Dunn, Assistant Professor, Schulich School of Law, Dalhousie University, as an individual: Thank you. I appreciate the invitation to be here today, and I’m joining you from Mi’kma’ki.

My area of research focuses on online harms, AI and synthetic media, and I’ll focus my comments on that area today.

Without freedom from discrimination, the ability to control our digital identities and social and legal protections from technology-facilitated harms, human rights and economic security will not be available to all. The promises of the benefits of AI are not being felt by many of those already experiencing discrimination and limitations in Canada.

Artificial intelligence can cause equality-based harms at multiple levels, more than I’ll ever be able to name here today. Indiscriminate data scraping that is used to train AI often violates principles of consent, privacy, intellectual property and Indigenous data sovereignty.

Equality seeking groups often have limited resources to fight against the misuse of their data, and individual, rather than collective rights, are more easily recognized in our legal system, leaving gaps for systemic reforms.

Large image-based data sets scraped from the internet have been shown to have scooped up problematic content — such as child sexual abuse material — that is then used to train AI systems. The people paid to sort through, categorize and clean the data scraped from the internet are often exposed to some of the darkest and most graphic material from the internet for minimal pay and with limited support to address the harms from viewing this type of data.

Once the data is collected and categorized, the labels attached to the data can entrench sexist and racist stereotypes that show up in AI’s outputs. Defining social constructs like race and gender is difficult and can’t be reduced to simple data labels. Once AI has been trained, it can output results that rely on and reinforce inappropriate stereotypes about equality seeking groups, including for decisions that have significant importance to them, such as whether they will be interviewed for a job, access basic services, enter the country or get out on bail.

Some of the AI systems themselves are designed to facilitate harmful behaviour, or, at the very least, do not provide guardrails to adequately prevent it. Sam Altman, the CEO of OpenAI, recently apologized to the Tumbler Ridge community for not reporting the shooter’s account to law enforcement after banning it.

X recently allowed Grok to digitally undress images of people on its site, including some children, often stripping women down to only a floss bikini or with clear tape over their sexual organs. Many DeepNude apps only work on women’s and girls’ bodies so that the resulting images are only able to create nude female bodies. These images are often produced without consent and violate the person’s privacy and sexual integrity.

On most public-facing sexual deepfake websites, almost all of the people featured in them are female celebrities, Twitch streamers, Instagram influencers and politicians. Women in leadership are particularly targeted by this type of AI-based harm.

A 2024 American study on sexual deepfakes found that 26 senators and members of Congress, 25 of whom were women — with one man — had non-consensual sexual deepfakes made of them with tens of thousands of matching hits when searching their names across several popular deepfake pages.

Additionally, websites and social media accounts have been using AI video tools to create everything from sites dedicated to AI-simulated videos of women getting shot in the face to videos of Black women featured as gorillas. The normalization of online harassment and discrimination that is amplified by AI makes it undesirable to be a public-facing person in this era, especially if you are a member of an equality-seeking group. The individuals and organizations that develop this technology are profiting off this type of content with little recourse or limitations.

For example, in the case of non-consensual sexual deepfakes, which have been known to the public for over a decade, there has been little governmental response to address this harm. Two criminal cases in Canada have recently found that the current criminal intimate image provisions do not apply to digitally altered images, though proposals in Bill C-16 aim to change that. Only some provinces in Canada have civil protections for people whose images have been sexually deepfaked, and in Canada there are only three provinces with dedicated government-supported organizations to provide support to people who are experiencing some of these types of online harms. Canada has no content moderation legislation and lacks a digital safety commissioner.

The Personal Information Protection and Electronic Documents Act, or PIPEDA, provides some protections, but it lacks teeth and needs reform to further ensure that human rights will be prioritized over business interests. Without systemic responses that target the platforms and technology companies causing these types of harm, there is little opportunity for truly fulfilled human rights.

Thank you for your time, and I look forward to your questions.

The Chair: Thank you both for your presentations, and I now look to senators for questions.

Senator Arnot: This set of questions is for Professor Szilagyi. The rule of law requires decisions to be noble, contestable and attributable. What does the rule of law require when AI is used to make or inform decisions affecting rights, benefits, employment or access to services? And at what point does an AI-assisted decision become, in substance, an automated decision requiring legal safeguards? If there is any extra time, should Canadians have a legally enforceable right to know when AI has materially influenced a decision about them?

Ms. Szilagyi: Thank you for the question. It’s extremely important, I think, to acknowledge the rule of law aspects of any kind of automated decision-making because of how we have such a stringent expectation of what the rule of law is supposed to do in a democratic society.

We have ideas around accountability, repeatability, non‑arbitrariness and transparency, all of which have the potential to be obscured by AI, and all of which have the potential to be transformed when corporate tools are introduced into legislative functions.

This is the subject of my forthcoming book, which I will pass along to your committee clerk. It’s called The Rule of Law After Artificial Intelligence: Automated Narratives. In the book, I talk about three main case studies that are useful for characterizing the way in which the problems arise, one of which is this idea of sentencing software. In the first panel, Professor Thomasen alluded to some of the studies that have been done around facial recognition, and that is the subject of the second case study. The third one is about large language models and generative AI.

When I was doing my PhD, I used this little-known technology — no one had really heard of it much — called GPT-3, and I said, “Oh, I’m going to ask it the central organizing questions of the thesis.” I then defended my PhD in October 2022, and about one month later, ChatGPT was released to the world. All the people who said, “What are you going to do with a PhD in AI and law?” were laughing.

What I like to underscore when I talk about these things is how it subverts the narrative structure that is at the heart of what the common law system does. Common law builds a story over time — so whose stories get told and which data points get emphasized when we rely on technological aspects of new tech without really unpacking the way that the avoidance of those technologies works.

If we start using automated aspects of sentencing, sentencing is a great entry point to the conversation, because people think sentences are supposed to have a right answer, right? We see that all the time in criminal law that concerns mandatory minimums.

Actually, sentencing is designed to be discretionary. It is designed to be responsive to the unique experiences of an offender, and the Gladue factors really take that into account when limiting or offering guidance to judges about how sentences might occur. Any attempt to automate those kinds of sentences can rely on historical data sets that, themselves, might contain bias and discrimination.

It also limits the field of vision in terms of what the judicial system might be able to achieve. We see similar aspects with other kinds of technologies that exist. What is going to happen if we, unthinkingly, allow these technologies to make decisions for us?

Now, the last part of your question was about the right to an explanation. Could we have large language models that were able to say, “This is how I came to my decision, and this is what I was thinking?” That is aligned with the ideas of due process or procedural fairness in the administrative law context, the bigger factors, for example.

If we were to legislate that, and you do see that type of legislation in the GDPR in Europe, you will find that the right to an explanation is a little bit hollow because today’s deep learning models don’t truly understand how they are arriving at the decisions that they make. Therefore, asking to legislate that, I think, fundamentally misunderstands the technology and will not be the remedy that people hope that it would be. I think we need to look at deeper AI governance solutions to be really responsive to the issues.

Senator Bernard: Thank you both for your testimony here this evening. I would like to ask this question to both of you. I was intrigued by the statement: The rule of law problem versus the rule by law problem.

I was thinking about AI in law through the lens of human rights. What should we be doing to ensure that a human rights lens on AI is not just a principle but a legal requirement? Can it be a legal requirement that a human rights lens be used so that those who are so negatively impacted and disenfranchised are not left further behind?

Ms. Dunn: It is such a difficult question, and I think it’s ultimately one of the most important questions we can ask about AI. How do we remove that bias? Part of it is regulating the design of systems. I think this is true regardless of the technology. Whether it’s AI or tracking devices — whatever type of technology it is — we need regulations where things like safety by design and privacy by design are embedded at all stages, from development to output.

Often technologies are put out into the world without having thought of these issues beforehand. Where if there were a requirement that companies go through, at least, a process to think about it, consider it and embed the thought of human rights into the process of their design, it is something that would be very important.

Then, once these models are released to the public, there need to be regular reviews of them. The more sensitive the use of the AI system, the more often they need to be reviewed because what AI systems have been shown to do is adopt and amplify discrimination, sometimes making it worse than the initial data within it.

When we’re thinking about police systems, employment issues or access to funds from governments, if governments decide to adopt AI systems in this type of decision making, there needs to be regular review from an independent assessor, who can accurately get access to the data and the process to determine if the algorithm needs to be adjusted to course correct from the discriminatory output it’s creating.

Ms. Szilagyi: One of the tools used by lawyers and legal scholars coming from minority or marginalized groups is the role of narrative and telling the lived accounts and experiences of people that were experiencing harm in the legal system. You see it in feminist epistemologies, Indigenous traditions and critical race studies, for example. How do we think about including the narratives of those types of groups in the data that will be considered by any sort of technological system that is being advanced?

In the first panel, there was some discussion about facial recognition technology and the landmark Gender Shades study that revealed how poorly it is able to identify Black faces and, in particular, Black female faces. That is a useful thing to know about when you’re implementing that kind of technology.

Then, what other unique circumstances for the group that you’re considering need to be an example? For example, sometimes I work with police groups trying to educate police members about how AI technologies work. Particularly in Manitoba, you would have a very high percentage of the population that would be Indigenous, and that is particularly true in the incarcerated population. Are there Indigenous faces in any kind of training data set that would be used for facial recognition technology? Because if there aren’t, there is no way that there could be an accurate representation of those folks.

You can expand that to other situations where we think about the dominance of particular groups within the data set and this idea that there’s an almost a black-and-white binary where we say, “The data doesn’t lie. The data says this.” But whose story is the data telling? I think that is the question we need to ask in the data training and data curation phase that will enable a human rights-based approach to flow from the use of those technologies.

Senator McPhedran: First, I want to thank Professor Dunn. We have worked together on the issues of the online harassment of women parliamentarians, and I think this is a natural extension of that.

The question is to both of you. I came out of the days of the 1970s and 1980s when, internationally, many of us thought we were going to be able to regulate pornography. We saw pornography as such a crucial tool for entrenching misogyny and, as a result, femicide, et cetera.

We lost that fight. The commercial imperative of pornography being framed as freedom of speech, and this privatization is, I think, something similar that we are seeing potentially here with AI. That’s about as far as I’ve gotten in terms of my analysis, but I’d welcome any thoughts you might have on this.

Ms. Dunn: I think you’re right that it’s so difficult, and the cat is out of the bag a little with these things. It was the same thing with pornography. It is prolifically available now, and we continue to struggle with these ideas of how do we regulate various types of content online?

Artificial Intelligence is the same way. It is out there and growing. It is unknowable to many people, and it can be a struggle with how you regulate it. But I agreed with the comments from Professor Thomasen that, in some ways, this is a piece-by-piece model. We may need to have broader legislation about AI systems, and people need to be made aware about when decisions are being made about them with AI. What do we do at high-, medium- and low-risk systems?

However, we have to get back to the core: What are the rights we want to enforce? Like privacy rights, human rights and basic access to education so that people are knowledgeable about these types of things. The focus has to be across the board. Ultimately, one of the greatest things we need is support for people who are experiencing harm.

People are going to experience AI-generated harassment and abuse in their workplaces or in the general public. Sometimes, the services we need are down low. We have high-level regulated AI, but we also need to provide basic services, like front-line services, to people experiencing various harms caused by AI, like when we see AI psychosis or that our images have been manipulated.

I keep getting calls from people who have had their images manipulated because they are TikTok doctors and their images are being used for scams. Where is the place to call for that type of help? We need to develop an expertise and support system in Canada, ideally that is government funded, where people can at least go to learn what the variety of legal, emotional and technical supports that they can get to address this complicated issue are.

Senator McPhedran: Professor Szilagyi, please.

Ms. Szilagyi: I will add and say that the challenge of regulation in law and technology is often referred to as the pacing problem or the Collingridge dilemma, this idea that the law is continually playing catch-up with the adoption of new technologies. I like to talk about it as an avocado ripeness problem, right?

Not yet, not yet, not yet, not yet — oh, too late.

How do we make decisions about what laws we’re going to put in place when we’re still seeing the impacts of the technology unfold? It is a difficult problem, but not an insurmountable problem and not one where we need to throw up our hands in despair and say there is no possible way that we can do it. What I would like to empathize as your committee does its review is, is there a way to think about putting regulations on the technology companies themselves as opposed to the users of the technologies?

How do we want to think about responsibility versus how do we want to think about appropriate use? I think that is the place where we have the most opportunity and the people who should bear the brunt of that responsibility in an AI-governance structure.

Senator McPhedran: Thank you.

Senator K. Wells: Thank you. My question is for Professor Dunn. You mentioned about online harms extensively and highlighted some of the recommendations.

In particular, do you have thoughts about recommendations in relation to children and youth? You have probably been following what is happening around the world and perhaps what the federal government will be doing or introducing in terms of a social media ban. I would be interested in your thoughts on that. Also, chatbots are another topic we’re hearing a lot about.

You mentioned in one of your comments about a digital safety commissioner. Would you talk more about what that would look like from a regulatory framework?

The other conversation we’re having around online harms is online hate and the algorithms that are driving this kind of content. It was party of our anti-Semitism study to fuel a sort of sow division in western democratic societies, pitting communities against one another with that.

In terms of recommendations to address online harms, do you have thoughts on how we tackle this problem of online hate and responsibilities? Is that the platform, criminal law, or government legislation to address, specifically for children and youth?

Ms. Dunn: Yes, I sometimes feel like I repeat myself when I say we need a little bit of everything. When it comes to children and youth, I think the first place we need to change things is with our education systems when it comes to what is embedded and required to be taught to young people about online safety.

In the research I’ve done on young people, they’re looking for non-judgmental supports to be able to talk about these issues. In the modern era, many parents who are raising children have never had the internet; they don’t know what these problems are like and can often react strongly to young people’s circumstances.

From speaking to young people, what they want are resources, places they can go, and trusted adults they can get support from.

I think education on how to think about these harms and support systems in place for young people has two critical pieces.

Criminal laws can be expressive and helpful, but not a lot of young people use criminal law. It is pretty rare they report things to the police. The same thing for civil law: it’s good to have in place circumstances where they are needed, but I find they are so rarely used by young people, which is why I think the larger content moderation education is so important.

We need the sites to be safe. Ideally, we don’t want to limit that to children. I think these sites should be safe for all people when it comes to hate, racism, discrimination and sexism. These things need to be moderated in ways that people can all engage in free expression and creativity and the benefits that come with this technology.

I struggle with the idea of banning youth from social media. In Canada, we have a really rural population, and some of the youths that I spoke to in my research are rural youth who might not have someone who is like them because of their sexual identity, race or whatever it is. The internet is the one place they have been able to go to find connections and community. Any restrictions we’re going to give for young people need to be based on real harms.

I love the Roblox example, where they start to say kids playing games with kids should only be able to talk to kids. That is a great thing we should have for companies. That can be based more in mitigation of harms rather than age restriction.

I worry that age verification is such a simple, good-sounding law that is clean; it is easy. We kick them off the sites, then we don’t have to worry about safety or any of those things, but I think that right now we don’t have the technology that is available in order to do that safely without violating a lot of other privacy rights.

We need to consider what the harms are that are caused to children when it comes to algorithmic feeds that encourage them to scroll all day long. What are the ways that we can change the technology through an online harms bill?

I realize I’m taking a bit of time, but my final point is looking at the benefit of having something like an online safety commissioner or similar models to what we see in New Zealand, the U.K. and Australia, which is talked about as the gold standard.

One of the things that those bodies do other than providing help and helping these companies mitigate harms and enforce them through penalties is they also do research so that they know what the harms are and they know what remedies are needed. I think in Canada, we have massive research gaps about what it is that Canadians actually want. I think having some sort of a commission that could provide that type of research and support, as well as on the ground helplines, and dealing with the mitigation of the harms on those companies would be really important.

Senator McCallum: Thank you to the two panellists.

I want to return to some of the statements you made, whose stories get told and when you look at narrative and lived experiences. This morning, I was looking at the Second Generation Cut-off, which is what we’re studying in the House of Commons.

I went in to do the research, and AI always comes first. I was absolutely shocked to see what was written. It already is rewriting our history as First Nations. It is going to increase denialism and replicate historical injustices.

When I looked at what it wrote, I could see that it came from documents from the government in how they are protecting First Nations and how they are doing this work.

Who is accountable and who will force AI to rewrite this section?

Part of this is that it is not Charter compliant, but it didn’t mention that. It didn’t mention anything. I was shocked to see it, especially because we’re studying it. I thought imagine the work needed to track AI in all the directions it is going and to correct in all areas, because even with the work it does — I can think of it going into all directions, it wasn’t able to gather all the information correctly and state it correctly.

How do we deal with that now?

Ms. Szilagyi: It is an important issue and an interesting question.

I like to think about it through the lens of sourceless information, right?

When we get these AI-generated summaries, who is the author of them? If we are not able to trace it back, then we fall down these echo chambers of almost hearsay information that we think we have heard.

Even when you search something on Google now, you get the AI Overview as your first result, or within documents, Microsoft Copilot wants to offer you a short summary of that information.

What is really important to be aware of about these technologies is the stochastic nature of them. You can input the same query multiple times and receive multiple different responses each time. So the ability of a generative AI system to convey meaning is very limited, and perhaps even non-existent, because all it is doing is, deterministically, coming up with what is the statistical most likely token, not even word — right? Generative AI systems do not deal in words. They deal in tokens, which are words split up into pieces. What is the token that is most likely to offer a response to your question that will follow or mimic the way that the English language works?

When we ask these questions about responsibility and accountability for the messages that the systems are producing, there is a huge distancing mechanism that the companies themselves have put in place in order for their model to be functional in the way that it is. They say, “Oh, there is no real authorship going on here.” If you read the terms of use or the privacy policy, the little links at the bottom of the page that are so often forgotten — companies are not taking responsibility for the outputs, and, in fact, they say, “Don’t put in sensitive information. Don’t put anything in from a privacy perspective or a cybersecurity perspective that could be problematic.”

These are things we need to keep in mind when we think about information that has been generated in this way and what kind of responsibility could potentially be associated with it.

Ms. Dunn: Especially when it comes to Indigenous data, there are really important concepts being developed around Indigenous data sovereignty in the OCAP principles that need to be taken into consideration when we’re thinking about AI.

There is a difficult push and pull between the ability of people to control their data and tell their stories while also having the larger models being able to replicate the preferred stories of people who are connected to them. It is quite difficult to do those things at the same time.

First, one of the things that needs to happen is that the AI systems need to stop being forced on us in the way that they are not opt-in. You can barely opt out of them. In the research that I have been doing at universities, our work is being done on PDFs, it is being done on Google and all of these different things where the AI is put upon us. Even trying to get away from those systems so that you don’t see this narrative, which, I agree, is often an inaccurate narrative, being put on you. It is hard to resist. With a Google search, the first thing that comes up puts this story in front of you. If we want to say “no” to having AI tell stories for us and replicate narratives for us, what are the ways that we can turn it off?

Then we notice inaccuracies. What are ways we can report that and advocate for more diverse data sets and truth in what is coming out of these AI systems. It is a lot more complicated to figure out how to do that.

Senator McCallum: Thank you.

Senator Arnold: Thank you both for being here and sharing your knowledge with us. It is really helpful. I will use that avocado analogy. That was fantastic.

I have three questions. First, would you both agree that an online safety commissioner ombudsman is an important thing for us to recommend?

Second, Geoffrey Hinton, often called the “godfather of AI,” has warned that in order to prevent existential threats from artificial superintelligence, researchers must embed maternal instincts into future models to ensure they care about human well-being. Do you believe this is possible?

Third, I have only been doing this for about three months or so, heavy on AI, but I am starting to see bigger pushback in international media, not just about data centres but in general there is more awareness around the harms and more legislators who seem to be stepping up saying that we need to do something about this. I wonder if you have seen the same.

Ms. Dunn: Maybe I will answer your questions backwards.

First, I think it is so interesting watching the shift in mentality around AI. A few years ago, everyone was like, “Oh, wow, this is going to change our world in such a great way.” Now the younger generation is like, “We hate this. So, basically, we’re not going to have any jobs, no one will earn any money, and all of this money will be filtered to these multi-billionaires that we’re never going to be able to be unless we’re some crypto bro.”

I think the things that have been promised from AI have not really panned out. There are harms. It is discriminatory. It is inaccurate. People are using it for lawyers and doing all of these hallucinations, and they are struggling to connect with people. In my interviews with young people, they often just want — they’re like, “How do I go outside and meet people?” I love the kind of analogue movement that is going on with young people.

So there is pushback on whether this AI promise is actually panning out to be a good thing for the world, and I think there are many critiques that are coming.

Your second question about the superintelligence thing. I’m like, “Let’s worry about that when it gets here.” Right now, there are a lot of other issues we need to be dealing with, and it is such a distraction from the environmental, social and discriminatory outcomes that are happening right now with AI. What can we do with AI right now?

The maternal instinct is like a — it strikes me as oddly sexist that we need a mom in there to keep them all okay. I don’t love that language. But I know people are very concerned about superintelligence.

I support, 100%, an online commissioner and ombudsman. More than anything, having a body that is well resourced, thoughtful and who can understand how this works in Canada and the struggles we are going to face as a nation, who is going to try to regulate social media companies and AI, who can hear from Canadians and know what their experiences are like — and the resistance that we’re going to face not having a lot of those companies in our jurisdiction — is absolutely critical.

Professor Szilagyi, I will pass it on to you.

Ms. Szilagyi: I will go back down the other way. Number one, yes, I agree with the online safety commissioner, but I caution the committee against too blindly following Australia’s lead. Things start making the headlines about legislative approaches, and they get really excited, and it turns out that isn’t working very well. The thing that I’m talking about in particular is the social media ban for young people in Australia. Since it came into effect, the research shows that more than 70% of young people are still able to access the 10 banned platforms and that it just hasn’t had the impact that they were hoping for.

Number two, I saw Geoffrey Hinton speak about three weeks ago, and he made that same point about superintelligence, right, having this idea that there is this one situation where mothers — they are able to produce their children and their children grow up to be more powerful than them, but they will still love the thing that created them, and can we instill that?

I think this is a very fun talking point. It is a really great thing to bring out when you are on a speaking tour, but it commits what is called, in the literature, the Android Fallacy.

It puts the argument way off in the future, and so conversations around existential risk tend to play on a science fiction narrative and science fiction stereotypes about tech. The Terminator, The Matrix, I, Robot. Pick your poison. But it tends to be a speculative future that is really based in science fiction understanding of anthropomorphized robots that will march off the assembly line and take over society.

Computers are not going to get too smart and take over the world. The problem is computers are still quite dumb, and they’re already starting to take over the world. They are already starting to do things and make decisions for us in ways that they shouldn’t, in ways that will lead to that automated decision making and, linking back to what I talked about in my opening statement, that automation bias and automation complacency at work.

The example I like to use to describe this is the idea of advanced driver assistance systems. You have all the bells and whistles in your car now — parking assist, lane-keeping assist, the backup camera, all of these things. Five or ten years ago, you could not have imagined backing out of a parking spot without shoulder-checking, but, over time, do you start to do it? Do you start to rely on those decision aids? At what point do you stop fact-checking? At what point do you let ChatGPT write the first draft and then just send that email?

Those are the things to which we need to be paying close attention. I am glad we’re starting to see more pushback around these AI harms, but we’re in tight up against a lot of hype and a lot of really strong messaging from big tech companies that are really lobbying different industries.

The doctors I talk to are constantly getting inundated with requests to use AI transcription. Lawyers are constantly getting inundated with presentations about new legal tech and how it will make their practice better and easier and faster.

The police — this one that makes me so scared because they want to use all of these new AI tools, and they don’t necessarily understand how those tools work and the transformative impacts they can have on people’s decision-making capacity if they blindly allow those tools to work.

We’re in tight against that. We have to put up a good fight against these corporate interests. They say it’s a problem of high capitalism and low socialism — technology regulation.

Senator McPhedran: My feeling is that a lot of what we have been discussing has been on the assumption that a lot of the generation of knowledge or information is local and within the sphere of Canada. My question is more about the Canadian government’s capacity for accountability mechanisms for AI, especially large language models, that have been developed outside of Canada, outside of our jurisdiction and by non-Canadian corporations.

Ms. Dunn: It is going to be hard and challenging. You can see already the Data Structures and Algorithms, or DSA, and the various online harms bills that have been introduced globally where even with social media companies are making demands for transparency and access to data; there have been resistance and challenges with that.

It will be the same with AI. It will not be easy or convenient. These companies will resist where they can. Even here in Canada, where we’ve had takedown orders for intimate images, X is challenging those decisions and saying we should only geofence those for Canada. There will be a big push, back and forth. Right now, these companies are some of the most powerful, wealthy companies and people in the world. They know that they have a lot of power in this current time. For legislation in Canada, there will be resistance.

That reality is there. It doesn’t mean that we shouldn’t at least make an effort and an attempt to see what we can do in that. In some circumstances, what we look at with Clearview AI, where they said they weren’t willing to comply with Canadian privacy policies, and that service just isn’t available in Canada anymore. That might be one of the realities of what comes down, but we have to wait and see.

Ms. Szilagyi: We need to push really hard against this idea that just because a technology exists, we need to use it. There’s always this “genie has been let out of the bottle” narrative that accompanies the use of new technologies. We have the ability to ban things, to regulate and to make decisions about appropriate uses for different things. We can develop made-in-Canada technology and made-in-Canada solutions.

To the degree that we want to adopt models that are designed by Big Tech companies in Silicon Valley, we will largely be unable to transform the nature of those technologies or put a lot of limitations on how they work. That’s the distinction between insider and outsider tools. Do you develop something in-house? Or do you build something off the organic data set that you have collected and curated yourself? Or do you simply assuage to, “Okay, whatever there is available on the market is what we’ll purchase.”

I have heard concerning reports around broad adoptions within different sectors of government of Microsoft Copilot, which is powered by OpenAI, saying, “Well, we already use the Microsoft 365 suite of products, and we feel comfortable with the cybersecurity and data security and privacy laws that we have in place about those things, so this is the same as that.” But it is not the same as that. There are a bunch of reasons in terms of how the information gets generated and stored, the insights that it’s capable of producing and the data sovereignty piece about information passing between countries when things are being processed, stored and developed, particularly in the United States but potentially in other countries as well.

How do we think about those accountability mechanisms for large technology companies? Can we put regulations in place that will ask the right people to bear that burden in order to penalize the platforms themselves, as opposed to the users of the technology who are sort of blindly following along?

The Chair: Thank you, Professor Szilagyi and Professor Dunn, for your informed presentations and even more informed responses to the questions. On behalf of the committee, I sincerely thank you for taking the time to appear before us today. Your testimony will be helpful in our deliberations as we conclude our study and move on to a report.

This concludes the public portion of our meeting. Thank you.

(The committee continued in camera.)