Research

The CAISI Research Program at CIFAR is the independent scientific engine of the Government of Canada’s AI Safety Institute (CAISI). By bridging fundamental research with practical safety solutions, we are contributing to the safe development and deployment of advanced AI systems that are grounded in Canadian values.

Project Streams

Solution Networks

Multidisciplinary teams of researchers awarded $700,000 over two years to design and implement real-world solutions for pressing AI safety challenges.

Catalyst Projects

Early-stage funding designed to support the exploration of novel AI safety research questions examining the technical and sociotechnical implications of advanced AI systems.

Alignment Projects

A global research initiative led by the UK AI Security Institute and supported by a $1 million contribution from the Government of Canada through CIFAR. This collaboration funds critical research to ensure advanced AI systems remain reliable, responsive and fundamentally aligned with human values.

Explore

AI Alignment Project

Scaling laws, data distributions, and learning dynamics: simulated high-energy physics data as a benchmark for data in the wild

Abstract background with geometric shapes

Catalyst Project

Research

Project Streams

Solution Networks

Catalyst Projects

Alignment Projects

Explore

A unified statistical framework for quantifying rare event risks for language models

Addressing AI-Safety through Indigenous Community-based Governance

Advancing AI alignment through debate and shared normative reasoning

Adversarial robustness in knowledge graphs

Adversarial robustness of large language model (LLM) safety

Democratic Alignment of LLMs Through Economic Theory: Relative Preferences and Strategic Coordination

Economic Foundations of AI Certification

Formalizing constraints for assessing and mitigating agentic risk

Game-theoretic safety guarantees for advanced AI systems

Maintaining meaningful control: Navigating agency and oversight in AI-assisted coding

Mitigating dialect bias

On the safe use of diffusion-based foundation models

Performative Empathy and Deceptive Alignment

Repetition, Resistance, and Reinforcement: Longitudinal Effects of Conversational AI on Political Attitudes

Safe autonomous chemistry labs

Safeguarding courts from synthetic AI content

Sample-efficient online fine-tuning against resistant behaviors: statistical foundations for post-training alignment

Sampling latent explanations from LLMs for safe and interpretable reasoning

Scaling laws, data distributions, and learning dynamics: simulated high-energy physics data as a benchmark for data in the wild

Socio-Technical Solutions to Improve Information Integrity and AI Literacy

Testing Red-Team Safeguards Against AI Persuasion in Democratic Governance

Towards Socially Grounded AI Safety: Integrating Causal and Institutional Reasoning in Language Models