What Are the Government's Demands?
Through a landmark Executive Order and subsequent actions, the White House has mandated new standards for AI safety. A key requirement is that developers of the most powerful AI models, often called 'frontier models', must report their safety test results
to the federal government. This includes conducting extensive 'red-team' testing, where experts actively try to make the AI fail or produce harmful content to find vulnerabilities before public release. The National Institute of Standards and Technology (NIST) is tasked with creating the standards for this testing. The overall goal is to manage the significant risks posed by advanced AI, from algorithmic discrimination and privacy violations to threats against national security.
The Enforcement Conundrum
This is where the picture gets complicated. While the government can demand safety reports under the Defense Production Act, the primary bodies created to oversee this are still finding their footing. The US AI Safety Institute (AISI), housed within NIST, was established to evaluate AI models and advance the science of AI safety. However, the AISI's core function is research and voluntary testing; it does not have statutory enforcement power to compel companies to act or to punish them for non-compliance. Its role is to publish guidance and best practices, not to issue binding rules or penalties, a stark contrast to the regulatory powers of bodies like the EU AI Office.
So, Who Is the Watchdog?
The AISI's primary power is one of expertise and influence. By developing rigorous testing benchmarks and collaborating with international partners, it aims to create a global consensus on what 'safe AI' looks like. The institute works with a consortium of over 280 organizations from industry, academia, and civil society to develop these guidelines. The idea is to foster an 'ecosystem of safety' where companies are held accountable by a network of experts, and where voluntary commitments gradually transition into industry-wide standards. However, some critics worry these institutes may focus too narrowly on specific risks while neglecting others, like fairness and bias.
The Missing Piece: Penalties
For the most part, the federal framework for AI safety currently lacks explicit, universally applicable financial penalties for non-compliance. While a recent White House accord with major AI companies was described as 'morally binding', it is not legally binding and specifies no enforcement mechanism or penalties. Real teeth are emerging more at the state level. California, for example, has passed several laws with significant civil penalties for the misuse of AI, particularly concerning minors. One law includes penalties up to $15,000 per affected child for intentional violations, while another focused on frontier models allows for fines up to $1 million per violation for large companies. Other states like Colorado and New York have also implemented rules with penalties for specific applications of AI, such as in hiring and employment.
A Framework Under Construction
The current US strategy appears to be a mix of federal guidance and state-level enforcement. The federal government sets the tone and direction, promoting innovation while establishing a baseline for safety through NIST and the AISI. This approach encourages companies to adopt safer practices voluntarily, with the implicit threat that if they don't, Congress or state legislatures will step in with more rigid laws. Several federal bills have been proposed that would create stricter liability for AI companies and give the government more direct power, including a 'government-controlled kill switch' for dangerous models. This suggests that while the current system relies heavily on voluntary action, a more punitive framework could be on the horizon if self-regulation fails to protect the public.
