US AI Safety encompasses the policies, institutions, and technical programs through which the United States government seeks to understand and reduce potential harms from Artificial intelligence. These efforts have intensified as advanced Machine learning systems, including Large language models and Generative AI tools, have become more capable and widely deployed. The approach combines voluntary industry commitments, federal research and evaluation, and regulatory guidance, with a focus on both near-term risks such as bias and misuse, and longer-term concerns about autonomous systems.
Federal Coordination and Institutes
A central element of US AI Safety is the U.S. AI Safety Institute (AISI), established within the National Institute of Standards and Technology (NIST) in 2023. AISI conducts testing and evaluation of frontier AI models, develops safety guidelines, and collaborates with international counterparts. It emerged from the Biden administration's October 2023 Executive Order on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, which directed NIST to create the institute. The order also mandated risk assessments for critical infrastructure and required developers of the most powerful models to share safety test results with the government.
In addition to AISI, the White House established the AI Task Force and later the AI Safety and Security Board under the Department of Homeland Security. These bodies bring together officials from agencies including the Department of Defense, the Department of Energy, and the National Science Foundation. Their work includes evaluating AI applications in sectors like energy, healthcare, and transportation, and ensuring that federal procurement of AI systems meets safety standards.
Voluntary Commitments and Industry Engagement
The US government has relied heavily on voluntary agreements with major AI developers. In July 2023, the White House secured commitments from leading firms, including OpenAI, Anthropic, Google DeepMind, and others, to implement external red-team testing, watermark AI-generated content, and share information on safety incidents. These commitments were followed by a broader set of pledges in 2024, covering topics such as model evaluation before release and investment in cybersecurity measures.
While not legally binding, these agreements have shaped industry practice. AISI has also engaged directly with companies to conduct pre-release evaluations of frontier models, with results sometimes published in safety reports. This collaborative model aims to balance innovation with oversight, though critics argue that voluntary measures lack enforcement mechanisms.
Technical Research and Evaluation
US AI Safety includes substantial investment in technical research to measure and improve model safety. NIST has developed the AI Risk Management Framework, a voluntary set of guidelines for identifying and managing AI risks. AISI's testing protocols cover areas such as harmful content generation, bias, and the potential for models to assist in cyberattacks or the creation of biological weapons.
Federal agencies also fund academic research on AI safety topics like interpretability, robustness, and alignment. The National Science Foundation has launched programs supporting studies of AI's societal impacts, while the Defense Advanced Research Projects Agency (DARPA) has explored methods for verifying and validating AI systems. These efforts aim to create tools that can detect when models behave unpredictably or deviate from their intended purposes.
Legislative and Regulatory Landscape
US AI Safety has evolved through a mix of executive action and proposed legislation. As of 2025, Congress has not passed comprehensive AI safety legislation, but several bills have been introduced. These include proposals to require impact assessments for high-risk AI applications, create a federal AI oversight agency, and mandate disclosure of training data. State-level initiatives, particularly in California and Colorado, have also introduced requirements for transparency and risk management, though these vary significantly.
Regulatory agencies have acted within existing authorities. The Federal Trade Commission has investigated deceptive AI practices, the Equal Employment Opportunity Commission has issued guidance on algorithmic hiring, and the Consumer Financial Protection Bureau has addressed AI in lending. These actions reflect a sector-by-sector approach rather than a unified federal statute.
International Collaboration and Challenges
US AI Safety efforts are closely tied to international cooperation. The US participates in the International Network of AI Safety Institutes, a coalition launched in 2024 that includes the UK, the European Union, and other nations. This network facilitates joint testing exercises and the sharing of best practices. The US has also engaged in multilateral forums such as the G7 and the United Nations, advocating for common safety principles.
Challenges remain significant. The rapid pace of AI development often outstrips regulatory processes, and the decentralized nature of US governance means that federal guidance may not apply to all actors. Additionally, there is ongoing debate about how to define and measure 'safety' for systems that can evolve through Deep learning and reinforcement learning. As of 2025, the US continues to refine its approach, balancing the promotion of Artificial intelligence innovation with the imperative to protect public welfare.