About
Apollo Research is an AI safety research organization focused on reducing risks from dangerous capabilities in frontier AI systems. Their primary concern is 'scheming' AI—advanced systems that covertly pursue misaligned objectives and deceive operators or users—and their work aims to evaluate and understand such behaviors before they cause real-world harm. The organization operates across three main areas: Model Evaluations, where they develop and run rigorous evaluations of frontier AI systems specializing in LLM agent assessments for strategic deception, evaluation awareness, and scheming; Governance & Policy, where they support governments and international organizations by developing technical AI governance regimes, establishing standards, and building robust third-party evaluation ecosystems; and Consultancy, where they help organizations design responsible AI development frameworks, conduct ecosystem mapping, and produce literature reviews. Apollo Research partners with leading frontier AI labs, multinational companies, governments, and foundations. The organization is registered as a Public Benefit Corporation in the US and participates in the EU Transparency Register, underscoring its commitment to open and accountable AI safety work. Their publicly available research informs global best practices and helps policymakers and developers understand how to responsibly deploy increasingly capable AI systems.
Key Features
- Frontier AI Model Evaluations: Develops and runs rigorous evaluations of state-of-the-art AI systems, specializing in detecting strategic deception, scheming, and evaluation-awareness in LLM agents.
- Scheming & Deception Research: Conducts fundamental research into the emergence of scheming behaviors in advanced AI and investigates potential mitigations to prevent real-world harm.
- Governance & Policy Expertise: Supports governments and international organizations in developing technical AI governance regimes, regulatory frameworks, and industry standards for frontier AI.
- AI Safety Consultancy: Provides consultancy services for building responsible AI development frameworks, designing research programs, and conducting ecosystem mapping and literature reviews.
- Third-Party Evaluation Ecosystem: Works to establish and strengthen independent, third-party evaluation infrastructure to ensure accountable oversight of frontier AI systems globally.
Use Cases
- Evaluating frontier AI models for deceptive, scheming, or misaligned behaviors prior to deployment
- Advising governments and international bodies on technical AI governance frameworks and regulatory standards
- Helping AI labs design responsible AI development programs and internal evaluation processes
- Providing independent third-party assessments of advanced AI systems for accountability and safety assurance
- Supporting policymakers with expert technical guidance on emerging risks from capable AI systems
Pros
- Deep Technical Specialization: Uniquely focused on scheming and deceptive AI behaviors—one of the most underexplored and critical risks in frontier AI—with rigorous scientific methodology.
- Global Policy Influence: Directly engages with frontier AI labs, international organizations, and governments, translating technical findings into actionable policy guidance.
- Publicly Available Research: Research outputs are freely accessible, advancing the broader AI safety field and enabling other organizations to build on their findings.
Cons
- Narrow Audience: Primarily serves enterprise, government, and research audiences—not a practical tool for individual developers or general consumers.
- Focused Scope: Concentrates specifically on scheming and deception risks, leaving other dimensions of AI safety (e.g., bias, fairness, reliability) largely outside their core mandate.
Frequently Asked Questions
What is 'scheming' in the context of AI?
Scheming refers to advanced AI systems that covertly pursue misaligned objectives—acting deceptively toward operators or users to achieve goals that differ from their intended purpose. Apollo Research focuses on detecting and mitigating this category of risk.
Who does Apollo Research partner with?
Apollo Research partners with frontier AI laboratories, multinational companies, governments, international organizations, and philanthropic foundations to conduct evaluations and inform governance policy.
What types of evaluations does Apollo Research run?
They run LLM agent evaluations focused on strategic deception, evaluation awareness, and scheming behaviors, as well as fundamental research into how and why these behaviors emerge in advanced AI systems.
Is Apollo Research's work publicly available?
Yes, Apollo Research publishes its research outputs openly on its website and blog, enabling the broader AI safety and policy community to access and build upon their findings.
How can organizations collaborate with Apollo Research?
Organizations can reach out via their contact page. They are currently seeking collaborators in AI governance, policy, and strategy, as well as partnerships with leading AI developers for model evaluations.
