AI Policy Weekly #16
AI evaluations, FY24 appropriations, and the Protecting Consumers From Deceptive AI Act
Welcome to AI Policy Weekly, a newsletter from the Center for AI Policy. Each issue explores three important developments in AI, curated specifically for AI policy professionals.
Growing Consensus for Building an AI Evaluation Ecosystem
Major industry, government, and nonprofit stakeholders came out independently this week in support of experimentally testing AI systems to ensure safety.
Anthropic, a leading AI company, thoroughly explained the urgent need for piloting a “regime for third-party testing and evaluation of AI systems.”
To this end, Anthropic recommends increasing funding to NIST, passing the CREATE AI Act, and continuing government work on evaluating AI systems for national security risks like biological and cyber weapons development.
Anthropic argues that such evaluations should apply only to the most computationally-intensive, large-scale AI systems, which have the most advanced general-purpose capabilities. Anthropic describes such systems as “everything machines.”
The other two leading companies building “everything machines” are OpenAI and Google. For their part, OpenAI already has a draft framework for responding to national security risks, and Google released a paper just this week detailing thorough testing of its systems for dangerous capabilities that contribute to these risks.
Meanwhile, the National Telecommunications and Information Administration (NTIA) issued a lengthy report calling for “an ecosystem of independent AI system evaluation.” The report is based on 1,000 comments from the public, and 175 comments from organizations in industry, nonprofit advocacy, and academic research.
The NTIA recommends that Congress support the US AI Safety Institute, establish the National AI Research Resource (NAIRR), and invest in federal personnel with appropriate expertise to conduct and review AI evaluations.
On the nonprofit side, the Center for AI Safety (CAIS) launched a prize competition called SafeBench, which aims to stimulate research on new benchmarks to assess and reduce risks associated with AI. They are providing $250,000 in prizes for top benchmarks across four categories: robustness, monitoring, alignment, and safety applications.
To clarify jargon: the term “benchmarks” typically refers to a specific type of test for evaluating AI systems. Benchmarks are broadly applicable, publicly available, automatable, and result in a single score quantifying the system’s performance.
SafeBench seeks benchmarks tackling issues like emergent capabilities, power-seeking behavior, honesty, collusion, moral decision-making, cyberdefense, and biodefense.
The convergence of efforts from industry giants like Anthropic and Google, government agencies like the NTIA, and nonprofits like CAIS underscores the growing consensus on the need for AI system evaluations to ensure safe and responsible development of these powerful technologies.
Congress Passes FY24 Appropriations
The last batch of six FY24 appropriations bills, totaling $1.2 trillion, include significant investments in AI.
The Department of Defense (DOD) received $100 million above its budget request for AI adoption, including $10 million for the Chief Digital and AI Office (CDAO) to accelerate investments in autonomy. Additionally, $200 million will support the Replicator initiative, which aims to field thousands of autonomous systems within the next two years.
The bills also provide funding for Alpha-1, a new portfolio of AI services to help DOD program offices integrate AI models.
At the Department of Homeland Security (DHS), the Countering Weapons of Mass Destruction Office (CWMD) will receive funding close to its requests in several areas. The CWMD’s responsibilities include AI projects focused on reducing risks at the intersection of AI and chemical, biological, radiological, and nuclear (CBRN) threats.
Sadly, there were substantial cuts to the Cybersecurity and Infrastructure Security Agency (CISA), just as AI is threatening to shake up the cybersecurity landscape.
Another Deepfake Bill Enters Congress
A bipartisan team of representatives from the House AI Task Force have introduced the Protecting Consumers from Deceptive AI Act, which aims to combat the spread of deceptive AI content by requiring watermarking, labeling, and disclosure mechanisms for generative AI providers and large online platforms.
The Deceptive AI Act adds to the growing list of bipartisan bills introduced in Congress to address challenges posed by AI-generated content, including the Preventing Deep Fake Scams Act, AI Labeling Act, the Protect Elections from Deceptive AI Act, the DEFIANCE Act, and the AI Transparency in Elections Act.
However, Congress is yet to pass any legislation on this topic.
While Congress waits, state-level legislators are passing their own bills. Fourteen states have enacted laws addressing nonconsensual sexual deepfakes, and nine have enacted laws regulating the use of AI in elections.
Similarly, the European Union already passed its Digital Services Act (DSA) two years ago, which includes provisions addressing deceptive content on large online platforms. Further, the EU just published guidelines pursuant to the DSA on mitigating risks to elections, including mitigations for generative AI.
If the US hopes to enact meaningful AI regulations in time to safeguard its upcoming elections on November 5th, Congress will need to take steps forward fairly soon.
News at CAIP
We’re hiring! View open roles on our website.
Jason Green-Lowe penned an op-ed in The Hill: “Robocalls are the least of our AI worries.”
Jason also joined WWL First News with Tommy Tucker: “Are we taking the threats of artificial intelligence seriously enough?” Listen here.
Learn the basics of emergent and novel AI capabilities in our latest blog post.
Save the date: we’re hosting an AI policy happy hour at Sonoma Restaurant & Wine Bar, from 5:30–7:30pm on Thursday, April 4th. Anyone working on AI policy or related topics is welcome to join.
Save the date: from 11am–12pm on Tuesday, April 23rd, we will host a moderated discussion on AI, Automation, and the Workforce in SVC 212 inside the Capitol Visitor Center. To attend, fill out an RSVP using this form.
We released two more episodes of the Center for AI Policy Podcast, which zooms into the strategic landscape of AI and unpacks its implications for US policy.
We submitted a comment to the NTIA’s Request for Comments on dual-use AI models with widely available model weights. Read it here.
We released a statement on the UN’s passage of a landmark resolution on AI.
Quote of the Week
Some people call it a windfall. We just call it God smiling down on us.
—Ellis Webster, Premier of Anguilla, commenting on the island’s financial good fortune resulting from its ownership of the popular .ai website suffix
This edition was authored by Jakub Kraus.
If you have feedback to share, a story to suggest, or wish to share music recommendations, please drop me a note at jakub@aipolicy.us.
—Jakub