The age of open and unregulated frontier artificial intelligence ended over the summer.

In June, Americans (and the rest of the world) lost access to Anthropic’s latest public model, Fable, after the federal government implemented export controls under the Export Control Reform Act (ERCA). As applied, ERCA required Anthropic to obtain a license from the federal government before exporting or transferring the model to "all destinations worldwide" and "all foreign persons" including those residing in the United States.

At the time, the basis for the export control, per David Sacks, was a concern over the purported cybersecurity threat posed by Fable. While some referred to the export control as an AI "Kill Switch," it's worth clarifying that the federal government did not act directly on the model. And the control did not apply to American citizens. Instead, Anthropic shuttered the model and chose to rescind access to Americans in response to the export control. The administration and Anthropic eventually negotiated a re-release.

This series of events represented a dramatic shift in the Trump administration's approach to regulating frontier AI. The administration had previously employed a relatively laissez-faire approach, and a recent executive order, for example, stressed that participation of frontier AI labs in its regulatory efforts was to be strictly "voluntary." This made the administration's apparent shift to apparent preclearance for frontier models surprising.

Of course, whether the Trump administration will stringently impose preclearance or licensing for all frontier AI models remains unclear. The summer’s conflict notwithstanding, Anthropic might favor that shift. The company has long suggested that governments should possess top-down, pervasive control of AI. It's for this reason that a common framing of the spat between Anthropic and the Trump administration was that Anthropic merely reaped that which it sowed. That's fair. But construing this as a discrete disagreement between the president and an outspoken opponent that "has an inability to communicate effectively" with this administration is tenuous.

As discussed, the basis for the Trump administration's export control was due to purported cybersecurity risks posed by Fable. But that concern applies equally to models released by other companies, like OpenAI, xAI, and Google, as demonstrated by OpenAI’s “Hugging Face incident.” Thus, any model possessing Fable-like capabilities will likely face scrutiny from the federal executive. And in the event a lab releases a model without the government's go-ahead, they risk a similar, heavy-handed response.

Companies typically respond to incentives, especially incentives backed by government force. Perhaps that's why OpenAI cooperated with the administration as the company released its own new models over the summer. The Financial Times reported, for example, that OpenAI has worked with the Trump administration to implement the AI executive order. Others, like Google, xAI, and Microsoft have agreed to "allow the US government to evaluate artificial intelligence models before they are publicly available." In sum, each company developing frontier (or near frontier) AI in the United States has agreed to let the government vet its models prior to public deployment. Some have argued this amounts to a “de facto licensing system.”

Assuming the administration is willing to continually enforce its agreements with actions like export controls, preclearance and licensing does appear to have arrived. This suggests that deployment of frontier AI is no longer free and open--a company can't just release a powerful model to the public.

In the immediate term, the Trump administration seems committed to pre-deployment review for arguably legitimate concerns related to cybersecurity. But what about the next administration? What happens when preclearance is turned over to actors that have broader aims and believe AI can effectuate their desired political ends through distortion of the truth?

For example, it is well-documented that during the Covid pandemic, the Biden administration attempted to pressure social media companies into censoring truthful information. My own law firm, the New Civil Liberties Alliance, sued the Biden administration over this censorship effort. Moreover, governments around the world adhere to the principle that ideas and speech alone can be dangerous. So, a government suggesting that "AI safety" requires the suppression of "dangerous" ideas is not a substantial leap. My concern, to put it plainly, is that a future administration will wield an AI preclearance or licensing power to similarly manipulate or suppress the truthful information contained in AI output.

An objection to that sort of interference is that the First Amendment may preclude such efforts. I've suggested this elsewhere. After all, the Supreme Court has long held that "if there is any fixed star in our constitutional constellation...it is the principle that the government may not interfere with "an uninhibited marketplace of ideas.'" First Amendment law professor Eugene Volokh has also argued that the public has a strong First Amendment interest in receiving "meaningful communication," including the sort of communication people receive from AI. The First Amendment will therefore limit some inevitable government efforts to dictate LLM output.

But the First Amendment isn't sufficient protection. Courts are inherently ad hoc institutions, and they lack the ability to respond to all government action. More problematically, it's not even clear the public would be aware that model responses were being influenced by the government. How AI generates its responses--and the sort of guardrails in place--often operate as a black box. Most people will likely assume the model is being relatively objective and truthful, and they won't always know if the AI is being steered away from sharing certain topics or information. This means that the public might be completely unaware of future censorship, as it is with current and pervasive censorship of the internet.

At bottom, the question is no longer whether AI will be regulated. It is, and political actors will continue to regulate the technology. The relevant question is how we regulate AI. Will we merely try to constrain its potentially dangerous applications? Or will we use preclearance to engage in manipulation of the truth and social engineering? It's unclear. But as it always has, political power, and who controls its relevant levers, matters a lot for the future of AI and the unfettered, uncensored availability of information for the general public.