Google DeepMind
AI-generated text. This page was generated using artificial intelligence.
Google DeepMind is a Google AI research organization formed in April 2023 by combining DeepMind and Google Research's Brain team. Its contribution to AI policy includes published safety protocols, model-evaluation practices and engagement with governments and other research organizations. This profile focuses on those governance activities rather than its general product and scientific research.[1][2]
Organizational scope
The 2023 announcement describes Google DeepMind as a single research unit working with Google's product teams and Alphabet's resources. That relationship matters when interpreting policy activity: a Google-wide lobbying position or litigation filing should not automatically be attributed to DeepMind alone. Conversely, a framework published by the lab may describe practices used across relevant Google models.[1]
The Frontier Safety Framework sits alongside Google's broader AI principles and responsibility practices. It is a company governance document that can inform wider debate about evaluations, risk thresholds and oversight; it is not itself a statute or a regulator's certificate of compliance.[2]
UK AI Safety Summit materials
Google DeepMind's 2023 summit materials describe approaches to evaluating risks, protecting model weights, monitoring deployed systems and identifying synthetic content. The account calls for collaboration on shared risks and discusses responsible disclosure, including work with the Frontier Model Forum. These are the company's reported practices and proposals at the time of the summit.[3]
For synthetic media, the materials consider watermarking, metadata and digital signatures, recognizing limitations and the need for policy and public-literacy measures. They also discuss research access and consultation with outside experts. References to particular products or planned releases belong to the 2023 account. They do not independently demonstrate that every announced safeguard was effective or remains unchanged.[3]
Capability thresholds and risk management
Version 3.1 of the Frontier Safety Framework, published April 17, 2026, distinguishes Critical Capability Levels from Tracked Capability Levels. The former concern capabilities associated with severe harm absent mitigation; the latter are intended to capture significant risks at lower capability thresholds. The framework covers misuse as well as machine-learning research and development and misalignment.[4]
Its misuse domains include chemical, biological, radiological and nuclear risks, cyber risks and harmful manipulation. The process links identifying capabilities with detection protocols, mitigation plans and, where appropriate, outside involvement. The document acknowledges evolving research and says some mitigations are most effective when adopted across the industry. These qualifications matter: publication of a threshold does not establish that all relevant risks can be measured reliably or that another developer uses the same threshold.[4]
Changes in the published framework
The September 2025 update added a harmful-manipulation capability level and revised the treatment of risks from models capable of accelerating AI research. It discussed scenarios in which misaligned systems could interfere with operators' ability to direct, modify or stop them. The account also extended safety-case review to certain large-scale internal deployments, recognizing that risk can arise before external release.[5]
The April 2026 update added Tracked Capability Levels and further explanation of the risk-management process. Distinguishing those two dates avoids attributing later changes to the original 2025 publication. The updates illustrate an iterative approach to private governance; they do not establish that the framework is equivalent to every jurisdiction's legal obligations.[5]
Selected contributions
- Safety and responsibility materials for the UK AI Safety Summit (2023).[3]
- Frontier Safety Framework publication series, beginning in May 2024.[2]
- Framework expansion covering manipulation and revised misalignment protocols (September 2025).[5]
- Version 3.1 and tracked capability levels (April 2026).[4]
Related articles
- AVERI — Independent auditing research and a confidential evaluation pilot with Google DeepMind.
- Policy
- United Kingdom policy on catastrophic AI risk
- Frontier Model Forum
- Partnership on AI
- OpenAI
- Anthropic
References
- ↑ 1.0 1.1 1.2 Google DeepMind, Announcing Google DeepMind, April 20, 2023. Accessed September 12, 2026.
- ↑ 2.0 2.1 2.2 2.3 Google DeepMind, Frontier Safety. Accessed September 12, 2026.
- ↑ 3.0 3.1 3.2 Google DeepMind, AI Safety Summit Policies, 2023. Accessed September 12, 2026.
- ↑ 4.0 4.1 4.2 Google DeepMind, Frontier Safety Framework, version 3.1, April 17, 2026. Accessed September 12, 2026.
- ↑ 5.0 5.1 5.2 Google DeepMind, Strengthening Our Frontier Safety Framework, September 22, 2025, updated April 17, 2026. Accessed September 12, 2026.