AI safety research
for real-world
deployment contexts.
We study how advanced AI systems, including future AGI, behave when assumptions of abundant compute, stable governance, and expert oversight no longer hold. This work reduces existential risk globally.
Core research priorities.
Our research focuses on safety challenges that emerge in real-world deployment contexts, particularly those that global labs often overlook.
AI Safety Under Resource Constraints
How do AI safety guarantees change when deployed in environments with unreliable infrastructure, limited compute, and weak institutional oversight? We study robustness, alignment, and interpretability under scarcity.
Key Problems
- •Alignment under partial observability and limited feedback
- •Safety guarantees without reliable human intervention
- •Interpretability methods usable by non-experts
- •Robustness to infrastructure failures and adversarial conditions
Impact
Reveals failure modes that affect AI deployment globally, not just in Africa.
Multilingual AI Alignment
Most AI systems are trained primarily on English. How do alignment failures manifest in multilingual contexts? We study value alignment across Zambian languages (Bemba, Nyanja, Tonga, Lozi, etc.).
Key Problems
- •Value learning from non-English cultural contexts
- •Detecting misalignment in low-resource languages
- •Cross-cultural value alignment challenges
- •Bias amplification in multilingual systems
Impact
Ensures AI systems are aligned with diverse human values, not just Western norms.
Frontier Model Evaluation & Red Teaming
We evaluate cutting-edge AI systems for dangerous capabilities, misalignment, and potential misuse. Our team has red teamed GPT-4o, Claude models, and other frontier systems.
Key Problems
- •Identifying emergent dangerous capabilities
- •Testing for deceptive alignment
- •Evaluating robustness to adversarial attacks
- •Assessing misuse potential in African contexts
Impact
Directly improves safety of deployed AI systems used by billions of people.
AI Governance for Weak Institutions
How do you govern advanced AI when regulatory capacity is limited? We develop practical governance frameworks that work in resource-constrained environments.
Key Problems
- •Governance mechanisms that don't require strong institutions
- •International coordination for AI safety
- •Compute governance and model access controls
- •Liability frameworks for AI harms
Impact
Creates governance models applicable to most of the world, not just wealthy nations.
AGI Preparedness in Africa
What happens when AGI is deployed in contexts with limited institutional capacity? We study AGI as a state capacity substitute and the unique risks this poses.
Key Problems
- •AGI deployment risks in fragile institutions
- •Power asymmetries from unequal AGI access
- •Economic disruption and adaptation strategies
- •Ensuring African participation in AGI governance
Impact
Ensures AGI development considers global deployment contexts, reducing catastrophic risks.
Questions we're working to answer.
These are fundamental questions in AI safety that remain unsolved. We welcome collaboration from researchers worldwide.
Technical Alignment
- How do we align AI systems with values that vary across cultures?
- Can we develop interpretability methods that work in low-resource settings?
- How do we ensure robustness when we can't assume reliable infrastructure?
- What safety guarantees can we provide without constant human oversight?
Deployment Safety
- How do we detect misalignment in deployed systems with limited monitoring?
- What are the early warning signs of catastrophic AI failures?
- How do we build safety infrastructure that scales to AGI?
- Can we create fail-safes that work even when institutions are weak?
Governance & Policy
- How do we govern AI when regulatory capacity is limited?
- What international coordination mechanisms can reduce existential risk?
- How do we ensure equitable participation in AI governance?
- What liability frameworks prevent race-to-the-bottom dynamics?
How our research reduces existential risk.
We don't just publish papers. We create tools, benchmarks, and frameworks that frontier labs and governments can adopt to make AI systems safer.
Research
Conduct technical research on AI safety under real-world constraints. Publish findings in top venues.
Tools & Benchmarks
Convert research into evaluation tools, safety benchmarks, and practical frameworks.
Adoption
Work with frontier labs, governments, and organizations to adopt safety practices.
Global Impact
Reduce existential risk by making AI systems safer in the contexts where they're most likely to fail.
Collaborate with us.
We welcome collaboration from researchers, institutions, and organizations working on AI safety. If our research agenda aligns with your interests, let's work together.