Voluntary AI Commitments Announced by Leading Companies in 2023
On July 21, 2023, the Biden-Harris Administration secured voluntary commitments from seven leading AI companies—Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI—to prioritize safety, security, and trust in AI development. These commitments marked a critical step toward responsible AI innovation, setting a baseline for future frontier-safety pledges in the industry. The commitments were announced during a White House meeting, emphasizing the shared responsibility of companies to ensure AI products are safe before public release.
The commitments were part of a broader effort by the Biden-Harris Administration to manage AI risks while fostering innovation. The administration had previously convened AI experts, published the Blueprint for an AI Bill of Rights, and signed an Executive Order to address bias and discrimination in AI systems. These actions reinforced the government’s commitment to protecting Americans from AI-related harms while encouraging industry collaboration.
The voluntary nature of these commitments underscored the administration’s approach to balancing regulation with industry self-governance. While the commitments were not legally binding, they established a framework for responsible AI development that later influenced global discussions on AI governance. The White House also consulted with international partners, including Australia, Brazil, Canada, and the UK, to align these commitments with broader global efforts to govern AI responsibly.
Core Commitments: Safety, Security, and Transparency
These commitments addressed critical risks, including biosecurity threats, cybersecurity vulnerabilities, and societal harms such as bias and deception.
- Internal and External Security Testing: Companies committed to conducting rigorous security testing of AI systems before public release. This included collaboration with independent experts to assess risks related to biosecurity, cybersecurity, and broader societal impacts.
- Information Sharing and Collaboration: The companies agreed to share best practices, insights on safeguard circumvention attempts, and technical collaboration with governments, civil society, and academia. This transparency aimed to strengthen collective efforts in managing AI risks across sectors.
- Cybersecurity and Model Weight Protection: To prevent unauthorized access to proprietary AI model weights—the foundational components of AI systems—the companies committed to investing in cybersecurity and insider threat safeguards. Model weights were to be released only under controlled conditions, with security risks carefully evaluated.
- Third-Party Vulnerability Reporting: Companies pledged to establish mechanisms for third-party discovery and reporting of vulnerabilities in their AI systems. This included creating channels for external researchers and users to report issues post-release, ensuring rapid identification and resolution of security flaws.
- AI-Generated Content Attribution: To combat fraud and deception, the companies committed to developing technical solutions—such as watermarking—to help users identify AI-generated content. This measure aimed to preserve trust in digital media while enabling creative applications of AI.
- Public Reporting of AI Capabilities and Limitations: Companies agreed to disclose the capabilities, limitations, and appropriate/inappropriate use cases of their AI systems. These reports would cover both security risks and societal impacts, including fairness, bias, and privacy concerns, providing transparency to users and regulators.
Societal and International Implications of the Commitments
Beyond technical safeguards, the commitments addressed broader societal risks posed by AI, including bias, discrimination, and privacy violations. The companies acknowledged the historical challenges of AI systems amplifying harmful biases and committed to prioritizing research in these areas. This included developing AI models that mitigate bias, protect privacy, and align with ethical standards.
The commitments were discussed with allies and partners, including members of the G-7 Hiroshima Process, the UK’s AI Safety Summit, and India’s Global Partnership on AI. These discussions aimed to create a unified global framework for AI development, ensuring that U.S. efforts complemented broader international initiatives.
Domestically, the administration continued to advance AI policy through legislative and executive actions. This included the development of an AI-focused Executive Order, bipartisan legislation to regulate AI, and investments in AI research, such as the $140 million allocation to National AI Research Institutes. These steps reinforced the U.S. position as a leader in responsible AI innovation, balancing technological advancement with public safety.
Broader Context: Administration’s AI Governance Strategy
The 2023 commitments were part of a multi-faceted strategy by the Biden-Harris Administration to address AI risks. Earlier in 2023, Vice President Harris convened leaders from consumer protection, labor, and civil rights sectors to discuss AI-related risks, reaffirming the administration’s commitment to protecting Americans from discrimination and harm. Additionally, President Biden met with AI experts in San Francisco to explore opportunities and risks in AI development.
The administration’s Blueprint for an AI Bill of Rights outlined principles for protecting Americans’ rights in the age of AI, while the Executive Order directed federal agencies to eliminate bias in AI design and use.
The Office of Management and Budget was also preparing draft policy guidance for federal agencies to ensure AI systems were developed, procured, and used with safeguards for public rights and safety. This guidance aimed to standardize AI governance across federal operations, further aligning with the voluntary commitments made by private companies.
Required actions
- Internal and external red-teaming before release; sharing information on risks; investing in cybersecurity and insider-threat safeguards for model weights - Watermarking or provenance for AI-generated audio and visual content; public reporting of capabilities and limitations; prioritising research on societal risks