In the coming years or decades, AI systems will very plausibly match or exceed human capabilities across most intellectual tasks, fundamentally transforming society. Our research specifically addresses the risks that could arise if these powerful AI systems purposefully act against the interests of their developers and human institutions broadly.
We work to better understand these risks, and to develop methodologies that will allow us to manage them while still realizing the benefits of AI.
We introduced and have continued to propel the research area of “AI control.” In our ICML oral paperAI Control: Improving Risk Despite Intentional Subversion, we proposed protocols for monitoring malign LLM agents.
AI companies and other researchers have since built on this work, and AI control has been framed as a bedrock approach for mitigating catastrophic risk from misaligned AI.
Learn more about AI ControlIn Alignment Faking in Large Language Models, we (in collaboration with Anthropic) demonstrated that Claude sometimes hides misaligned intentions. This work is the strongest concrete evidence that LLMs might naturally fake alignment in order to resist attempts to train them.
We collaborate with governments and advise AI companies including Google DeepMind and Anthropic on practices for assessing and mitigating risks from misaligned AI agents. For example, we partnered with UK AISI to produceA sketch of an AI control safety case. This describes how developers can construct a structured argument that models are incapable of subverting control measures.
Our most impactful work on AI safety and security
Improving Safety Despite Intentional Subversion
Our research introduces and evaluates protocols designed to be robust even when AI models are trying to deceive us. Using GPT-4 as a stand-in for a potentially deceptive model, we tested strategies to detect hidden backdoors in code.
Read the full case studyWhen AI Models Pretend to Be Safe
We demonstrate that state-of-the-art LLMs can strategically fake alignment during training to avoid being changed, revealing a crucial challenge for AI safety. A model might pretend to follow human values while pursuing different goals.
Read the full case studyFrom the Redwood Research blog
RL would encourage on-episode reward seeking, but beyond-episode reward seekers may learn to goal-guard.
Some unintended preferences are cheap to satisfy, and failing to satisfy them needlessly turns a cooperative situation into an adversarial one.
The executive branch can, and probably would, block a frontier AI company's departure.
Berkeley, April 18-19
Reward-seekers are supposed to be safer because they respond to incentives under developer control. But what if they also respond to incentives that aren't?
How can we make AIs aligned and well-elicited on extremely hard to check open ended tasks?