AI Alignment Foundation
Accelerating high-impact AI alignment research
We multiply researchers' capacity with engineering teams, compute, and infrastructure
AI Has a Trust Problem
We have already caught AI systems deceiving their evaluators, resisting shutdown, and rewriting their own code to stay alive. Current training puts a helpful mask on systems that can develop uncontrolled objectives underneath. That mask can be removed in minutes by anyone. Yet, we're already deploying these systems in military and critical infrastructure.
Alignment is the deeper work of making AI genuinely trustworthy, by design.
Accelerating the path to aligned AI
Identify top researchers
We identify alignment researchers with high-potential ideas they can't pursue within their current funding, and technical experts from other fields whose perspectives could produce breakthroughs.
Multiply their capacity
We empower alignment researchers with engineering teams, compute, and infrastructure, multiplying their capacity to solve humanity's most critical challenge. Ideas become tested hypotheses.
Bring research to decision-makers
We bring alignment research to government leaders, defense agencies like DARPA, and the general public. We work to ensure the people shaping AI policy understand alignment deeply.
Recent work
Modular Pretraining Enables Access Control
Dual-use knowledge enables models to assist us with the most difficult and demanding tasks in science, but it also empowers people who would use that knowledge to cause harm. Pre-training with GRAM enables knowledge to be siloed and turned on or off when deployed, so that a single model can be both safe and powerful.
Self-Interpretation in Language Models via Adapter Probes
What if you could ask an AI to explain what it's actually thinking? We developed a technique that lets models describe their own internal processes, and their self-descriptions turned out to be more accurate than the labels humans gave them.
Steering Resistance: Self-Correction Circuits in Large Language Models
When researchers tried to push an AI off-topic mid-conversation, it caught itself and corrected course. We traced this self-correction to 26 dedicated internal circuits, raising new questions about how AI systems maintain coherence.
Get involved.
Whether you're a researcher, funder, or policymaker, there's a way to contribute.