Establishes protocols for identifying and mitigating severe AI risks from Critical Capability Levels (CCLs) in domains like Autonomy, Biosecurity, Cybersecurity, and Machine Learning R&D. Implements security and deployment mitigations to manage risks, evolving with insights and future research.
Analysis summaries, actor details, and coverage mappings were LLM-classified and may contain errors.
This is an internal corporate policy document establishing Google DeepMind's own protocols for managing frontier AI risks. It uses voluntary language ('we will', 'we aim to') and lacks external enforcement mechanisms, but goes beyond typical soft law by establishing specific internal commitments and procedures.
The document has good coverage of approximately 10-12 subdomains, with strong focus on malicious actors (4.1, 4.2, 4.3), AI system security (2.2), competitive dynamics (6.4), AI safety failures (7.1, 7.2, 7.3), and governance structures (6.5). Coverage is concentrated in security, misuse prevention, and AI safety domains, with minimal attention to discrimination, privacy, misinformation, or socioeconomic impacts.
This is an internal corporate policy document governing Google DeepMind's own AI development activities. As an AI research and development company, Google DeepMind operates primarily in the Information sector (AI/ML development) and Scientific Research and Development Services sector. The document does not regulate external sectors but rather governs the company's internal processes for developing frontier AI models.
The document comprehensively covers Build and Use Model, Verify and Validate, Deploy, and Operate and Monitor stages. It focuses extensively on model evaluation, testing protocols, deployment mitigations, and ongoing monitoring of frontier models. There is minimal coverage of Plan and Design, and no explicit coverage of data collection and processing stages.
The document explicitly focuses on 'frontier models' and 'foundation models' as its primary scope. It does not use the terms general purpose AI, task-specific AI, or predictive AI. It mentions compute thresholds indirectly through '6x in effective compute' evaluation intervals. The framework addresses both model weights and deployed systems, with specific attention to preventing open release of model weights.
Google DeepMind
The document is authored by Google DeepMind and describes 'our first version of a set of protocols' and 'our existing suite of AI responsibility and safety practices', clearly indicating Google DeepMind as the proposer of this framework.
Google DeepMind
Google DeepMind enforces this framework on itself through internal processes. The document describes self-enforcement mechanisms including evaluation protocols, mitigation requirements, and the authority to halt deployment or development.
Google DeepMind; external parties (unspecified); independent third parties (unspecified); relevant stakeholder bodies (unspecified)
Primary monitoring is conducted by Google DeepMind itself through periodic evaluations and red-teaming. The document also mentions future involvement of external authorities and independent third parties in monitoring and oversight, though specific entities are not named.
Google DeepMind
The framework applies to Google DeepMind's own frontier model development and deployment activities. The document consistently uses 'our models', 'our evaluators', and 'our risk assessment' indicating this is an internal policy governing the company's own AI development.
10 subdomains (7 Good, 3 Minimal)