By Sam Schechner, Meghan Bobrowsky and Amrith Ramkumar
SAN FRANCISCO — Like much of the tech world, startup founder Quinn Slack was riled up.
It was just hours after the Trump administration on Tuesday said it was lifting its weekslong export ban on Anthropic's powerful artificial-intelligence model dubbed Fable. But the reversal had done little to ease Slack's fears about government control over AI.
He told a crowd of roughly 50 tech workers at a rally Tuesday evening called "Freedom of Intelligence" that the saga confirmed his worries: The government could, and would, gatekeep access to AI models. "That is a dangerous precedent to have set," said Slack, chief executive of AI-agent startup Amp.
Yes, the Fable ban may be over. But America's debate over the degree to which the federal government should control access to cutting-edge AI tools is just heating up.
There is growing awareness of just how powerful new AI tools are, but little agreement over how they should be controlled.
Spooked by the potential for new models from Anthropic and OpenAI to help bad actors find unknown software vulnerabilities to launch cyberattacks, the Trump administration recently created a new de facto approval process. It drew lightning bolts from across the tech-policy spectrum for flipping its approach to AI oversight, moving from an earlier hands-off approach.
"U.S. labs are getting the message that they should make sure that their models are never very good at cyber evaluations, lest they land in endless model purgatory," Alex Stamos, chief product officer of AI security firm Corridor, wrote on X after the Fable ban was lifted.
The effect of the administration's actions "was to put the entire industry on a shaky foundation," said Will Rinehart, a senior fellow at the right-leaning American Enterprise Institute.
"We have entered a world where a frontier model is considered 'secure' when the government says so," he said.
At the Freedom of Intelligence rally, where one woman was wearing a "Free Fable" T-shirt, some attendees said they think there could be room for some regulation on how AI is used. But one argued it was too early to do much regulation since people still don't understand the models' full capabilities.
Opponents of AI regulation have blamed Anthropic for what they describe as fearmongering, particularly around its Mythos models, arguing that it contributed to the Fable crackdown and new oversight. Anthropic has long advocated for AI safety measures.
Proponents of AI safety rules are also unhappy. Some say the administration's previous laissez-faire approach left the government flat-footed and ill-equipped to respond when powerful AI models posing foreseeable risks started to emerge. They also argue that the administration's current review process is opaque and unpredictable.
In order to lift the mid-June ban, the U.S. pushed Anthropic to broaden the company's guardrails limiting the model's ability to respond to harmful requests such as writing malware. Anthropic conceded but said that would lead to blocking more benign requests than before.
In one sign companies want more of a say in evaluations, Anthropic said it is working with other tech giants including Amazon, Microsoft and Google to develop a framework for gauging the severity of so-called jailbreaks that bypass safety measures. The Trump administration is working to implement an executive order focused on some of the same subjects.
Tensions have been running high since a spat earlier this year between the Pentagon and Anthropic over the use of AI in the military showed how thorny the decisions can be. That fight spawned multiple ongoing lawsuits. The Defense Department has said two-thirds of the operations that were using Claude are using other AI models and that it wants to phase out Anthropic's technology over time. Anthropic has said it believes it is important to work with the U.S. government.
Last week, before the White House lifted its export ban on Fable, OpenAI said it was limiting access to its newest AI models, dubbed GPT-5.6, helping spark a new round of hand-wringing. The company previously had discussions with the Trump administration about the models' potential risks.
Administration officials and advisers have said they are balancing innovation and security and trying to address safety risks in the most minimally invasive way possible.
American AI boosters argue that the bans and limitations would hurt the U.S. in its efforts to stay ahead of China in a geopolitical AI race as Chinese models improve.
"We are the most pro-innovation, anti-red-tape administration that has ever been in existence when it comes to technology," Sriram Krishnan, a former White House AI adviser who recently left the role after working on some of the Anthropic issues, said on CNBC on Wednesday. "At the same time, we need to make sure that our critical systems are protected and safe."
Some executives say the recent back and forth reflects a bumpy transition period while the executive order President Trump signed recently gets implemented. The order gave cybersecurity officials 30 days to accelerate efforts using AI to patch security vulnerabilities and share information about security threats.
When OpenAI announced the restricted release of GPT-5.6, it said that it didn't think the current system should become a long-term approach. The company said Wednesday that it appreciates the administration's work on "a durable framework for future frontier model releases."
Write to Sam Schechner at Sam.Schechner@wsj.com, Meghan Bobrowsky at meghan.bobrowsky@wsj.com and Amrith Ramkumar at amrith.ramkumar@wsj.com