The **Trump administration**'s approach to AI safety, particularly concerning generative AI models, has been described by some as haphazard and, in the case of **Anthropic**, potentially unconstitutional. As explained in an amicus brief by **EFF** and allies, the Pentagon's actions against **Anthropic** are alleged to violate the First Amendment. This stems from a perceived desire to punish an uncooperative company rather than legitimate national security concerns. ### Inconsistent AI Strategy The **Trump administration**'s broader AI strategy has generally minimized regulation, prioritizing the global 'race' to develop frontier models. This has included paring back regulations intended to address serious AI threats, such as AI-enabled cyberattacks, in the name of fostering innovation. However, **Anthropic** has been repeatedly singled out for arbitrary and heavy-handed rules and sanctions. For years, the federal government relied on **Anthropic**'s models for classified systems. But after **Anthropic** reportedly resisted government demands to use its models for autonomous weapons or surveillance, the company was designated a 'supply chain risk,' effectively banning government agencies and contractors from doing business with it. ### Judicial Intervention and Retaliation Claims A court issued a preliminary injunction preventing these sanctions from taking effect, a move urged by **EFF** and other civil liberties organizations. Without this judicial action, the sanctions could have cost **Anthropic** hundreds of millions of dollars, sending a clear message that companies must comply with government wishes or face consequences. **EFF**'s brief argues these sanctions were clear retaliation for **Anthropic**'s public refusal to allow the Pentagon to use its models for fully autonomous weapons and surveillance, deeming such retaliation unconstitutional. ### Escalation: Export Controls More recently, an executive order from the **Trump administration** escalated its actions against **Anthropic** by imposing 'export controls.' These controls ban foreign nationals from using **Anthropic**'s new **Mythos** and **Fable** models, leading **Anthropic** to shut down the models entirely to comply. These extreme measures were purportedly justified by security concerns, with the administration fearing that **Anthropic**'s **Mythos**-class models could be used to find and exploit software vulnerabilities. **Anthropic** itself has contributed to anxieties, initially claiming **Mythos** was too dangerous for public release and restricting access, with its CEO calling for a pause on AI development. ### The Double Standard Critics argue that regulators should cut through the hype, not feed it. Even if **Mythos** offered a modest improvement over existing technology, other LLMs with similar offensive cybersecurity capabilities are not subject to the same export controls. Instead, the government has largely embraced a voluntary system where companies submit models for cybersecurity testing 30 days before public release. AI policy, it is argued, should be reasonably responsive to real-world risk, grounded in technological realities, and no more burdensome than necessary. The decision to impose export controls on **Mythos**-class models while other AI models face a voluntary, light-touch framework meets none of these criteria. As leading cybersecurity experts and executives explained in an open letter, these sanctions prevent developers and security teams from using the best models to find and fix vulnerabilities before adversaries, armed with nearly as capable AI, can exploit them. ### Code is Still Speech Export controls on software tools like LLMs can undermine the free flow of digital communications and technologies crucial for activists, innovators, and ordinary users. Freedom of expression, **EFF** argues, requires access to these tools, and depriving the public of the best AI threatens rights without enhancing safety. **EFF** has historically challenged government efforts to restrict the publication of non-classified software. In the 1990s, **EFF** successfully challenged export controls on encryption software, establishing the principle that 'code is speech' and is protected by the First Amendment. Courts recognized software as both a functional tool and a means of conveying ideas, knowledge, and technical know-how. While AI models introduce new questions, efforts to restrict access to them implicate similar constitutional and speech concerns as older efforts to restrict encryption. Export controls are uniquely susceptible to abuse, especially when unilaterally imposed without clear and fair standards. Whether these export controls are another attempt to punish **Anthropic** or a misguided security measure, the public ultimately loses. While legitimate cybersecurity risks from advanced AI may warrant limited regulations, any government rules must be rational and evenhanded, whether they heavily regulate the technology or promote innovation.