OTTAWA — The Liberal government is seeking to add binding rules for generative AI systems to its proposed law for the technology, and bring the algorithms underpinning social platforms and search engines under its scope, Innovation Minister François-Philippe Champagne informed MPs earlier this week. Here’s what you need to know: 

The backstory: Champagne’s letter proposes amendments to Bill C-27, which he introduced in June 2022. Its Artificial Intelligence and Data Act (AIDA) sets rules for developers and users of high-impact systems. Both industry and civil society said the proposed law was too vague. The House of Commons industry committee is currently reviewing the bill; in September, Champagne promised changes in response to the feedback. 

Generating new rules: The government wants to add to its law requirements for “general-purpose systems,” multi-use instruments like ChatGPT and others built atop foundational models. Before launching such products, developers will need to assess potential risks and take steps to mitigate any harms or biases.

Pictures of Pope Francis in Balenciaga and Drake-The Weeknd singles have grabbed the attention of policymakers worried about more nefarious deepfakes in political or societal contexts. “The technical feasibility of watermarking synthetic media is currently a work in progress,” Champagne’s letter acknowledges. But companies building tools that produce text, images, video and other content that could pass for human-created will still need to try to ensure readers and viewers can figure out that the material is machine-made. Chatbots and person-seeming products will also need to let users know they’re interacting with AI.

The amendments make compulsory some of the main commitments in Ottawa’s voluntary code of conduct for generative AI, launched in September.

Define “high-impact”: The government will keep a schedule for the types of systems to which AIDA’s obligations will apply. It’s starting with an initial list of seven, including tools used in employment and health-care decisions, or by cops. The high-impact designation will also apply to AI that’s used to prioritize or moderate content on search engines and social platforms. (Ottawa’s long-promised law to tackle what it calls online harms will also affect tech giants’ algorithms for ranking and taking down posts.)

Digital rights scholars and groups have expressed concern that AIDA only focuses on AI’s negative impacts on individuals, missing how it can hurt marginalized communities or societies. While the amendments don’t redefine harm, they would allow cabinet to consider collective ones in deciding whether to add a new category of systems to its high-impact list.

If the changes are adopted, the law will differentiate more between makers and users of AI systems. AI sellers like Cohere and IBM have called for distinctions, arguing they don’t control what customers do with their models. 

Developers will need to create a “model card” with details about their machine-learning products—Ottawa will set out the specifics in future regulations. Open-source repository Hugging Face pioneered such tags, which list how the model is supposed to be used, performance indicators and the datasets on which it was trained.