Recognizing the potential benefits and risks of advanced AI, the White House Office of Science and Technology Policy (OSTP) has sought public input on the Biden administration’s national AI strategy. The Federation of American Scientists (FAS) is happy to recommend that federal agencies take concrete action to protect the rights and safety of Americans. As US companies create powerful frontier AI models, the federal government must direct the growth of this technology towards public good and risk mitigation.
Recommendation 1: OSTPs should work with appropriate agencies to develop and implement pre-deployment risk assessment protocols that apply to all frontier AI models.
Before launching a frontier AI system, developers should ensure its safety, reliability, and reliability through a pre-deployment risk assessment. This protocol aims to thoroughly analyze potential risks and vulnerabilities of AI models prior to deployment.
We advocate increasing funding to the National Institute of Standards and Technology (NIST) to strengthen its risk measurement capabilities and develop robust benchmarks for risk assessment of AI models. Based on NIST’s AI Risk Management Framework (RMF), evaluation metrics that incorporate a variety of cases, including open source models, academic research, and fine-tuning models that differ from large-scale laboratories such as OpenAI’s GPT-4 is standardized.
We propose that the Federal Trade Commission (FTC) implement and enforce this pre-implementation risk assessment strategy under Section 5 of the FTC Act. The FTC’s role in preventing unfair or deceptive conduct in commerce is consistent with mitigating potential risks posed by AI systems.
Recommendation 2: Federally funded AI-related projects should be required to adhere to an appropriate risk management framework.
As a significant funder of AI through contracts and grants, the US government has both a responsibility and an opportunity. Responsibility: Ensure AI applications meet high standards of risk management. Opportunity: Enhancing a culture of safety in AI development more broadly. Adherence to a risk management framework must be a prerequisite for AI projects seeking federal funding.
Currently, there are voluntary guidelines such as NIST’s AI RMF, but I propose to make them mandatory. Government agencies should require contractors to document and verify the risk management practices implemented under the contract. For institutions that do not have their own guidelines, the NIST AI RMF should be used. And NSF should require that the grantee documents compliance with her NIST AI RMF in an AI project grant application. This approach ensures that all federally funded AI initiatives maintain high standards of risk management.
Recommendation 3: The NSF should increase research and development funding for “trustworthy AI.”
“Trusted AI” refers to AI systems that are trustworthy, secure, transparent, privacy-enhanced, and fair. While the NSF is the leading non-military funder of AI R&D in the United States, our rough estimate is that investment in areas that promote trust has remained relatively flat, accounting for 10% of all AI grants. ~15% only. Given the annual AI-related budget of $800 million, we encourage the NSF to direct more of its grants toward credible AI research.
To enable this change, NSF can facilitate trusted AI research through specific solicitations. Launch targeted programs in this area. We will also include a “Trustworthy AI” section in funding applications, prompting researchers to outline the credibility of their projects. This will help assess the impact of AI projects and drive high-potential proposals in trustworthy AI. Finally, the researcher may be required or mandated to apply her NIST AI RMF during the study.
Recommendation 4: FedRAMP should be expanded to cover federally contracted AI applications.
The Federal Risk and Authorization Management Program (FedRAMP) is a government-wide effort to standardize security protocols for cloud services. Given the increasing use of AI services in federal operations, these services have similar security standards as they are responsible for managing sensitive data related to national security and personal privacy. system should be applied.
Expanding FedRAMP’s mandate to include AI services is a logical next step in securely integrating advanced technology into federal operations. Applying frameworks like FedRAMP to AI services requires establishing robust AI-specific security standards, such as secure data handling, model transparency, and robustness against adversarial attacks. Expanding the FedRAMP program will streamline the integration of AI into federal operations and avoid repeated security assessments.
Recommendation 5: The Department of Homeland Security should establish an AI incident database.
The Department of Homeland Security (DHS) should create a centralized AI incident database detailing AI-related breaches, failures, and exploits across industries. Existing authorization under the Homeland Security Act of 2002 allows DHS to assume this role. This database will deepen our understanding of the safety and security of AI systems, reduce risk, and build trust.
Voluntary reporting from AI stakeholders should be encouraged while maintaining data confidentiality. For greater effectiveness, anonymized or aggregated data should be shared with AI developers, researchers, and policy makers to better understand AI risks. DHS can use existing databases, such as those maintained by the Partnership on AI and Center for Security and Emerging Technologies, as well as contributions from global initiatives such as the Center for Financial Services Information Sharing and Analysis. You can also adapt the reporting method.
Recommendation 6: OSTP should work with agencies to streamline the process of granting agency waivers of interest to AI researchers with J-1 visas.
As the global race in AI continues, attracting and retaining a diverse and highly skilled workforce is essential. The J-1 exchange visitor program in the United States, which is popular with visiting researchers, requires some participants to return to their home country for two years before they can apply for permanent residency.
Federal agencies may waive this requirement for certain individuals through an “agency of interest” (IGA) request. Government agencies should establish a transparent and predictable process for AI researchers to apply for such exemptions. OSTP should work with government agencies to streamline this process. Taking cues from the Department of Defense’s structured application process, including dedicated web pages, application checklists, and sample sponsor letters, can be invaluable in improving the transition of AI talent to U.S. permanent residency. may prove to be
Learn more about these suggestions in public comments.
Post-corona visa interview exemption
Even after the COVID-19 emergency is over, exemptions from grooming should be a continuing policy when the risk is low.
05.11.23
|
Read in 10 minutes
How does OpenAI’s efforts to make GPT-4 “more secure” compare to the NIST AI Risk Management Framework?
Before releasing a powerful new system like GPT-4 to millions of users, I asked myself, “How do I know this system is safe, reliable, and trustworthy enough to release?” Can we do it?”
05.11.23
|
9 minute read
