Simon Fraser University

A concrete, shareable and adoptable, plan for governing AI as a tool (the AI we have now) and avoiding or offsetting the usual harms (misinformation, job loss, etc) and helping structure a system to stop the creation of AI as an actor (e.g., superintelligence with connections) that risks loss of control.

• Safe, secure and trustworthy AI
• AI capacity-building
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

AI is a powerful technology and merits robust and serious controls, akin to nuclear energy, chemical/pharmaceutical/aviation regulation and lethal autonomous weapon systems. We must operate with risk-based filtering and thresholds that are a light touch for those things that are not dangerous and rigorous when the AI exceeds certain levels.

Existential risk from loss of control must be addressed and cannot be skirted around or avoided.

Canada is a "middle power" and as such often has to rely on the work of others in domains like this. We have strong research capabilities and a democratic legislative process that could serve as a model and leadership for others, however. We could also host meetings and provide examples of working regulatory and legislative systems. We are - at present - sadly late to the game.

Share best practices, legitimize scholarly and practitioner warnings about the dangers, enliven and support concerned communities.

The UK has a very robust AI Safety infrastructure and Dr Bengio leads that, so you have the best connection you could ask for right there. Also, seek to involve Chinese legislators since they are the only other "frontier" lab location outside of the - likely ineffective - US.

Get industry people to agree to some hard lines/boundaries. Support and document "clean" data sets for training an AI (legal issues, fairness issues) and encourage the use of such tools and datasets.

Realistically, the only people who are effectively represented are the "tech bro" CEOs in Silicon Valley. Everyone else is lost in their wake.

Somehow convey a sense of urgency. This is not something that will wait for years like WSIS.

We suggest using CASX as guidelines for AI governance: CASX identifies four observable, independently measurable factors that together determine whether a system poses loss-of-control risk. None of them is "intelligence." **Capability** - What can the system demonstrably do in practice? Not what architecture it uses, not what training method produced it, but what behaviours and competencies does it exhibit? A system that can solve novel scientific problems, generate persuasive content, or devise multi-step strategies has crossed a different capability threshold than one that autocompletes email. **Autonomy** - To what degree does the system act without human confirmation? A system that waits for a prompt and returns a response is a tool. A system that initiates actions, pursues objectives across time, and resists interruption is operating autonomously. This is the factor most directly tied to loss of control. **Scale** - How widely and rapidly is the system deployed? A single instance in a research sandbox is a different governance problem than a million instances integrated into commercial infrastructure. Scale is the multiplier that turns an individual actor into a systemic risk. **Access** - What external systems can the system reach and affect? A model that generates text in a chat window has limited access. A model connected to financial systems, laboratory equipment, communications infrastructure, or robotic actuators has access that converts its capabilities into real-world consequences.