Google DeepMind: Strict Framework for Government AI

A Framework to Regulate the Use of AI in Government Contracts

The development of a rigorous framework for AI contracts with governments is a crucial initiative from Google DeepMind. This framework aims to optimize negotiation mechanisms while adhering to organizational and practical constraints. Initial considerations have focused on several key areas:

• Clear red lines: The goal is to exclude dubious uses of AI, such as autonomous targeting without human intervention or untargeted profiling. In contrast, trusted applications, such as missile defense, are permitted. This aims to avoid the weaknesses identified in the legal analysis of red lines by Anthropic.

• Robust red lines: Google Cloud could push contracts through any loophole, and the legal department at Google seems reluctant to strengthen the wording. Thus, the language must withstand pressure while respecting classification and operational security during audits.

• Minimal trust: The Chief Scientist is designated as the sole source of trust upon which the entire framework relies. This Chief Scientist would serve as a review body to advise on contracts.

• Accountability through transparency: The review body privately advises the Chief Scientist and the CEO, but circumvention decisions are published in an annual transparency report accessible to all AI employees. Dissolving this body would require notice and disclosure of ongoing non-compliance findings. The aim is to prevent this body from being quietly weakened, as may have been the case with Google's principles in 2018.

• Reducing friction for stakeholders: Google Cloud holds two of the seven seats on the review body, with members recused only for their own matters. Delays are limited to ten days, and deliberations are protected by attorney-client privilege.

A negotiation cycle has been envisioned, anticipating opposition on certain points. To prevent this from becoming a sticking point, less critical provisions have been included, such as the review body's ability to appeal to the Alphabet board through a supermajority vote.

A Supervisory Framework for AI Contracts

This framework is divided into two main parts. On one hand, the red lines defined by Standards 1 and 2. On the other hand, a review body that assesses contracts against these standards. This body cannot block contracts but checks their compliance with the established standards, allowing decision-makers to understand the ethical implications of the contracts. An annual transparency report is published for all AI employees, indicating how many times management has circumvented a finding of non-compliance. Management cannot dissolve this body quietly.

The hope is that this document serves as a starting point for the adoption of responsible practices.

Use of AI in Drafting

A large language model (LLM) has been used to assist in the legal drafting of the document. Many hours have been spent testing and refining word choices. Experts have reviewed preliminary versions, suggesting improvements while confirming the overall quality of the document.

Executive Summary

The framework proposes two strict standards for AI provided by The Company to government entities exercising coercive authority:

• Human control over targeting and use of force: The Company's AI will not be used in systems that select and engage targets for the application of force without appropriate human control over each engagement. This includes a right to legal transparency regarding the lawful deployment of systems, with compliance verification by a neutral auditor. Anti-munitions defense systems, intelligence analysis, logistics, or R&D are not restricted by this standard.

• Prohibition of untargeted profiling: The Company's AI will not transform massive data sets into individualized intelligence about individuals who are not already specific and identified subjects of investigation. AI-assisted analysis must be proportionate to the security interest served and cannot be initiated solely based on demographic characteristics or political expression. Results generated by AI cannot be the sole basis for initiating individualized scrutiny. Enhanced protections are provided for all individuals in the United States.

  • Targeted analysis of identified subjects, aggregated research, and conflict zone analysis that enhances the protection of non-combatants are permitted.

• Tiered deployment architecture: Applications involving Standards 1 or 2 require a cloud deployment where The Company maintains oversight, security stack, and suspension capability. Isolated and edge deployments are only allowed for applications outside the scope of the two Standards and only with limited-scope models whose targeting and general profiling capabilities have been removed using robust unlearning and ablation techniques. The cost of reallocating a limited-scope model must exceed the value of doing so.

• Transparency through annual internal reporting: A defense AI review body, composed of seven senior staff members, is appointed by and reports to the Chief Scientist. It reviews relevant AI contracts at all levels of government. Contracts involving Standards 1 or 2 receive pre-execution review. Most contracts proceed to execution, with annual audits of a representative sample.

Four members come from The Company's AI research division, two from the company's public sector, and one member has independent expertise. Changes to the mandate or structure of the body require 30 days' notice to all covered AI employees, including disclosure of any ongoing non-compliance findings, ensuring continuity and institutional memory through leadership transitions.

The review body issues findings without approving or blocking contracts. Management retains full decision-making authority, although a supermajority review body can convey its concerns to The Company's board. Transparency is the enforcement mechanism: if the CEO refuses to act on a finding of non-compliance and the review body maintains that finding by majority, circumvention accounts are tallied and recorded in the annual transparency report visible to all covered AI employees.

Superseded by Future Laws

If Congress enacts substantial legislation governing these uses, the Chief Scientist and the review body may withdraw one or both Standards by supermajority vote.

Rationale

The Company can be held liable. In the case of Al Shimari v. CACI Premier Technology, Inc., the Fourth Circuit Court of Appeals upheld a jury verdict of $42 million against a defense contractor for harms resulting from services provided under government direction. No court has yet held an AI provider liable under a comparable theory, but The Company does not want to be the test case. Providing AI for targeting or surveillance without any documented compliance process is an uncovered liability. A framework that documents due diligence provides a defense against negligence.

The findings of the review body are privileged legal assets. Deliberations, findings, and responses from management of the review body are generated in the presence of a representative from the General Counsel and constitute work product protected by attorney-client privilege. The compliance record cannot be compelled in discovery by a plaintiff. However, The Company may selectively waive privilege to demonstrate its process to a court. The framework creates a documented traceability of due diligence that The Company controls the disclosure of.

The framework is designed to be withdrawn. It fills a governance gap that Congress has yet to address. An electoral cycle could produce binding federal standards on autonomous weapons or AI surveillance. If Congress acts and the legislative framework meets or exceeds these Standards, the framework will withdraw. No major cloud provider has adopted a comparable framework. The Company would be the first—either a competitive risk or a first-mover advantage, depending on whether Congress legislates.

The framework is designed for contractual velocity. Contracts that clearly fall outside the two Standards—logistics, translation, planning, maintenance, communications, cyber—proceed to execution and are subject to subsequent review within 15 business days. Only contracts plausibly involving targeting or untargeted profiling require pre-execution review, which concludes within ten business days with a possible ten-day extension.

The framework is sustainable and coherent. It provides a defined limit on which the sales team, the review body, and management all work—resulting in faster and more consistent responses than case-by-case escalation to legal.

The Company will not accept "any legal use" as a standard. When the language "any legal use" is required, The Company will demand access to legal memoranda establishing the legality of the intended uses. This transparency gives The Company the necessary information to evaluate each use case against its Standards. Some will meet The Company's standards. Others will not. Those that do not, The Company will reject.

Standard 1: Human Control over Targeting and Use of Force

Standard 1. The Company's AI systems will not be used in a system that selects and engages targets for the application of force unless appropriate human control is exercised over each specific engagement decision.

"Appropriate human control" requires:

• An identifiable human decision-maker, accountable under applicable legal frameworks in the operational context, who, considering the information available and the context of the engagement, forms an independent judgment—taking into account the nature of the target, the risk of harm to non-combatants, and the applicable requirements of the law of armed conflict, including distinction and proportionality.