The institute?s operational model is unique in that it bifurcates its activities into two distinct streams: government-directed research (led by the National Research Council of Canada) and investigator-led research (administered by CIFAR).

Core Mandate and Mission

The primary mission of CAISI is to "advance the science of AI safety." It acts as a bridge between technical AI research and public safety policy. Its mandate focuses on:

• Risk Identification: Analyzing risks posed by advanced AI, specifically focusing on synthetic content (deepfakes, fraud) and "frontier" systems that may act unpredictably or evade human oversight.
• Trust and Adoption: Fostering public trust in AI technologies to encourage responsible adoption by Canadian industries.
• International Coordination: Serving as Canada's representative in the International Network of AI Safety Institutes, ensuring global interoperability of safety standards.

Operations and Research Streams

CAISI does not sell traditional software products. Instead, its "output" consists of research, safety evaluations, funding programs, and strategic guidance. Its work is divided into two primary operational pillars:

1. The Applied and Investigator-Led Research Stream

Managed by the Canadian Institute for Advanced Research (CIFAR), this stream focuses on tapping into Canada's academic ecosystem to solve forward-looking safety challenges. * Activity: It mobilizes experts from Canada?s three national AI institutes?Mila (Montr?al), Vector Institute (Toronto), and Amii (Edmonton)?to conduct fundamental research. * Focus Areas: * Developing technical methods to detect AI-generated content. * Creating frameworks for evaluating the robustness of Large Language Models (LLMs). * Researching "alignment" (ensuring AI systems' goals match human intent). * Grant Funding: CAISI, through CIFAR, issues Catalyst Grants to fund time-sensitive research projects that address immediate safety gaps in the AI landscape.

2. The Government-Directed Projects Stream

Managed by the National Research Council of Canada (NRC), this stream focuses on immediate, practical priorities for the federal government. * Activity: Conducting internal projects that may involve sensitive or classified information relevant to national security. * Focus Areas: * Cybersecurity: Assessing how AI can be used to attack critical infrastructure or how AI systems themselves can be hacked. * National Security: Evaluating risks related to the malicious use of AI by foreign adversaries or non-state actors. * Government Deployment: Providing technical assurance for AI tools used within federal departments.

Services and Deliverables

While CAISI acts as a regulatory and research body, it offers specific services to the government, the public, and the international community.

Safety Testing and Evaluation

CAISI conducts technical assessments of "frontier" AI models. * Pre-Deployment Testing: Collaborating with international partners to test advanced models before they are widely released to the public. * Vulnerability Assessment: Identifying "jailbreaks" (methods to bypass safety filters) and other failures in commercial AI systems.

Policy and Developer Guidance

The institute translates technical findings into actionable advice. * Developer Guidelines: issuing best practices for Canadian companies building AI, ensuring they align with the Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems. * Risk Frameworks: creating standardized metrics that developers can use to self-assess the safety of their models.

International Standardization

CAISI provides a service to the global community by harmonizing definitions and benchmarks. * Metric Unification: Working with the US and UK institutes to ensure that a "safe" rating in one country translates to another, reducing regulatory fragmentation for businesses.

Organizational Structure and Partners

CAISI is not a standalone crown corporation but a specialized initiative housed within a government department, leveraging external partnerships for execution.

• Lead Agency: Innovation, Science and Economic Development Canada (ISED).
• Primary Research Partner: CIFAR (Canadian Institute for Advanced Research).
• Technical Implementation Partner: National Research Council of Canada (NRC).
• Ecosystem Partners:
  • Mila: Quebec Artificial Intelligence Institute.
  • Vector Institute: Specializing in deep learning and machine learning.
  • Amii: Alberta Machine Intelligence Institute.

Location and Contact

As a federal initiative housed within ISED, CAISI operates out of Ottawa, with distributed research hubs across Canada's major AI centers (Montr?al, Toronto, Edmonton).