Thank you for inviting me to speak with you.
AI holds extraordinary promise, but the focus of my remarks today will be on AI safety. We are at a keen inflection point where AI is now embedded in every facet of society. However, trust in AI is low, and very little is being invested by the public and private sectors into safety. This is a problem. The buzz term is “trustworthy AI”, but this term is often used strategically. It sounds like safety, but it isn't. Thereby, it becomes a way to avoid implementing meaningful safety measures. The focus is on end-of-the-world scenarios rather than on the everyday ways that AI is already causing harm.
For the purpose of today, I want to emphasize one key message: AI does not merely reproduce existing inequality; it makes it worse. This is because AI can be used to create, scale, personalize and conceal discrimination, and it can be a tool for violence. The safety culture needs to be flipped so that AI safety, at least, is designed up front into the system, based on understanding the people and social practices at play. I do not believe this paradigm shift can be achieved through corporate governance alone, and existing laws are insufficient to meet the moment.
Let me situate this by explaining the AI harm pathway.
First is inputs and data training, and I'll offer two examples. Generative engine optimization, GEO, is the practice of creating and shaping online content so that it's attractive to generative AI systems and more likely to then inform AI outputs. It may be as simple as convincing the AI to recommend your company's mascara, to reproduce false and damaging information about a political opponent, or to reinforce negative stereotypes about women, 2SLGBTQI+ individuals and other marginalized groups.
Sometimes that discrimination in the inputs and outputs becomes the basis of decision-making harm, such as relying on AI as a recruiting tool. We all know the story of Amazon's AI tool being scrapped because it trained on old résumés, which were mostly from men, and taught itself that male candidates were preferable.
A key feature of both of these scenarios is that the bias was buried into the system.
Second is system design and operation. This is particularly acute on social media. The platforms are designed to keep users engaged in using their products and services. That's how they make their money. A few things happen here that lead to amplification of harmful content and of fuel-targeted harassment and self-harm. The platforms collect enormous amounts of information about user habits that inform the inferences made about them and that inform what algorithms and advertisers push to them. People are responsive to emotional and rage-filled content, so this pushes the most extreme content to the top of our feeds. Bad actors have figured out that they can monetize hate, and it has become a lucrative industry, fuelling discrimination and gender-based violence.
Design features make the problem worse. Children are being pushed beauty filters and are being encouraged to get plastic surgery, reinforcing unrealistic social expectations. It's an unbearable burden for children, and it's one that crosses gender lines.
Third is AI misuse. Deepfake and synthetic image abuse illustrate the problem particularly clearly. It's now easy, using AI, to create nude and nearly nude images, audio or video content. Some of this is illegal under criminal and civil law, particularly if it depicts children.
Grok is illustrative of the broader risks. It's an AI tool that was used to nudify images of mostly women, but it was also used to create other degrading images, such as women covered in bruises, tied up and gagged. Grok is integrated into X, so X's algorithm fuelled distribution of these images to wide audiences. Experiences like this send a message to women that they must be silent and invisible and that if they dare to exist in public, they will be punished and humiliated. The constant threat of sexual violence and the lack of available remedies chills meaningful participation in public life.
I'll close with the question about a federal response.
As it stands, Canadian law is fragmented and disproportionately places the burden on individuals to manage these harms. Canada does not presently have a comprehensive federal AI statute. Existing laws address parts of the problem, as will the digital safety act in Bill C-34.
A federal response should focus on shifting the burden for safety from victims to the actors in the AI supply chain: AI developers, AI platforms, and organizations that integrate AI into their products and services.
I'll close with this: The federal government should not treat innovation and safety as opposing objectives.
Thank you.