Canada - AI Safety Institute

Overview

The Canadian Artificial Intelligence Safety Institute (CAISI) is a federal research and coordination centre launched on November 12, 2024 by Innovation, Science and Economic Development Canada. CAISI's principal goal is to advance scientific understanding of risks associated with advanced AI systems and to produce practical tools, testing approaches and guidance to support safe, responsible adoption. The Institute combines two research streams: an applied and investigator‑led program implemented by the Canadian Institute for Advanced Research (CIFAR CAISI Research Program), and government‑directed projects implemented by the National Research Council of Canada. CAISI is housed within Innovation, Science and Economic Development Canada and has an initial allocation of CAD $50 million over five years. Its work complements Canada’s broader AI strategy and feeds evidence into policy instruments such as the proposed Artificial Intelligence and Data Act and voluntary codes.

Definitions

For CAISI purposes: "AI safety" refers to research and operational practices that reduce the probability of harmful outcomes from AI systems (including but not limited to misinformation, privacy leakage, cyber‑exploitation, and hazardous autonomous behaviour). "Advanced AI systems" generally denotes models or systems with capabilities that can produce high‑impact outputs or act with a degree of autonomy; CAISI uses domain‑specific definitions for project selection and testing. "Applied research" indicates investigator‑led academic or multidisciplinary research funded via CIFAR; "government‑directed projects" are operational workstreams commissioned by ISED and implemented by NRC. CAISI’s definitions are pragmatic and research‑oriented and align with international scientific reports and the International Network of AI Safety Institutes.

Governance and Institutional Framework

CAISI is administratively housed at Innovation, Science and Economic Development Canada with a dedicated office responsible for coordinating the research agenda, partnerships and international engagement. Operational delivery is split: CIFAR administers the applied/investigator research stream under a contribution agreement and an appointed scientific committee, while the National Research Council of Canada leads government‑directed technical projects tied to explicit federal priorities such as cybersecurity and public system resilience. CAISI engages Canada’s three national AI institutes—Mila, the Vector Institute and Amii—as well as academic, civil society and private sector partners. Governance mechanisms include steering committees, a scientific advisory board, periodic public reporting and funded‑project contractual obligations (reporting, ethics, data stewardship). For program announcements and governance details see the official CAISI pages on the ISED platform and the CIFAR program statement (ISED CAISI, CIFAR CAISI Research Program).

Key Focus Areas

CAISI’s agenda balances near‑term, applied risk reduction and foundational research on emergent risks. Priorities include: (1) detection and mitigation of synthetic content and misinformation produced by generative models; (2) privacy and data‑leakage risks, including memorization and extraction attacks against models; (3) robustness and adversarial resilience, especially in safety‑critical systems; (4) alignment and value‑constrained behaviours for agentic or highly autonomous systems; (5) safe deployment practices for high‑impact application domains (healthcare, critical infrastructure, finance, transportation); (6) tooling and reproducible testbeds for red‑teaming, evaluation and external audits; and (7) cross‑cutting socio‑technical research on governance, accountability and societal impacts. CAISI funds investigator‑led projects that explore the science of failure modes and funds NRC‑led work that tests and evaluates commercially or government‑operated systems against specific threat models. The institute also supports development of common testing standards, benchmarks and reproducible protocols to allow international comparability and to accelerate adoption of best practices by industry and government.

Implementation Framework

CAISI implements work through two primary streams. The first stream (applied and investigator‑led) is administered by CIFAR under a contribution agreement (publicly announced and funded) and steered by a scientific committee that selects projects through competitive calls. The second stream comprises government‑directed projects executed by the National Research Council, coordinated with other departments as required. Projects are subject to funding agreements that include deliverables, reporting, data governance and ethics requirements. CAISI maintains formal partnerships with the national AI institutes (Mila, Amii, Vector) for subject‑matter expertise and laboratory access. The institute also coordinates with international safety institutes to conduct joint testing and comparative evaluations. Program outputs include peer‑reviewed research, open‑source tools where appropriate, non‑proprietary testbeds and publicly available guidance documents for practitioners.

Monitoring and Evaluation

Monitoring is multi‑layered: project‑level performance is assessed through contractual milestones and deliverable acceptance; scientific quality is reviewed by academic peer review and by CAISI’s scientific committee; and programmatic impact is evaluated through annual reporting and periodic external reviews. CAISI intends to publish findings, toolkits and evaluation results openly where national security, privacy, or proprietary constraints do not forbid disclosure. The institute also plans to track indicators such as number of funded projects, reproducible tools produced, uptake of guidance by public sector bodies, and contribution to international safety norms. Results inform policy development, including inputs to the proposed Artificial Intelligence and Data Act and other governance instruments.

Penalties, Liability, and Appeals

CAISI is a research institute, not a statutory regulator: it does not itself create criminal or administrative penalties. Compliance obligations for recipients stem from contribution and funding agreements (for example with CIFAR or the NRC). Standard contractual remedies for non‑performance or misrepresentation include repayment of funds, withholding of future payments, contract termination and reputational consequences. Where CAISI activities expose unlawful conduct, findings may be referred to relevant regulators or law enforcement (for example privacy breaches to the Office of the Privacy Commissioner). Appeal mechanisms for funding decisions and contract disputes are typical of federal contribution programs (internal reconsideration, contractual dispute resolution and, where applicable, judicial review). CAISI also coordinates with regulator partners so that technical evidence can support enforcement under applicable statutes, such as privacy or sectoral safety regimes.

Relationship to Other Instruments

CAISI complements, but does not replace, regulatory instruments. It is explicitly designed to inform and support federal policy initiatives including the proposed Artificial Intelligence and Data Act (AIDA), the Pan‑Canadian Artificial Intelligence Strategy, and voluntary measures like Canada’s Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems. CAISI’s research outputs supply empirical evidence, test methods and technical recommendations that policy makers can incorporate into regulation, standards and procurement guidance. It also provides a platform for interagency collaboration (ISED, NRC, Privacy Commissioner, sector regulators) and for cross‑jurisdictional coordination with provincial research initiatives and international safety institutes.

International Alignment

International cooperation is core to CAISI’s mission. Canada is a founding participant in the International Network of AI Safety Institutes; CAISI is intended to coordinate joint testing, share evaluation methodologies and contribute to global safety standards. The institute references international reports and collaborates with peer institutes to harmonize approaches for model evaluation, red‑teaming, and responsible disclosure. CAISI also uses international scientific outputs to shape its research priorities and seeks to align tools and benchmarks to be compatible with work from allied jurisdictions, facilitating interoperability of test results and shared learning.

Implementation Timeline

Sources and References