OpenAI and Anthropic want AI rules. Connecticut is already writing them

Open AI and Anthropic have both called for federal regulation of artificial intelligence.

Some of the companies building the most powerful artificial intelligence systems recently reached out to global leaders at the United Nations calling for government regulation amid concerns about how fast the systems are developing and potential societal risks.

The call comes as the Trump administration has said it won't regulate AI out of concerns for what it could mean for how competitive the U.S. remains with China on the technology.

Article continues below this ad

That leaves Connecticut and other states struggling to regulate a technology that crosses every border as the companies and federal government argue this creates a conflicting patchwork of state laws.

“The industry should see the handwriting on the wall that there is going to be both state and federal engagement here,” U.S. Sen. Richard Blumenthal, D-Conn., told CT Insider. “The states are forcing the federal government to take some action here. They are providing an impetus.”

OpenAI has called for mandatory, capability-based national rules for advanced AI. Anthropic has also called for government action, proposing standards for the most powerful models based partly on the computing power used to train them. Both companies have argued that federal rules would give businesses a consistent framework and help the United States compete with China.

The Trump administration has repeatedly warned that excessive regulation could slow American innovation. The administration has also sought to limit the power of states to regulate AI independently.

State Sen. James Maroney, D-Milford, who has led a lot of regulation efforts on AI in Connecticut, said lawmakers from different states are coordinating to avoid making compliance impossible for companies. He said a national network includes hundreds of legislators sharing ideas and working toward common approaches.

“There are things that should be done federally, and then there are things that the states we will look to do together,” Maroney said.

The most serious warnings concern systems that could operate independently, evade human control or improve themselves faster than people can respond.

Sean O’Brien, founder of the Privacy Lab at Yale Law School, is less concerned about an imminent AI apocalypse than about familiar cybersecurity failures being amplified by automation.

“What permissions does the (large language model) or agent have?” O’Brien asked. “Has this been tested in a safe environment, such as a network disconnected from the internet? What access controls are in place? Who is monitoring and who is accountable?”

Connecticut’s approach has focused on transparency, consumer protection and disclosure.

Beginning Oct. 1, businesses that use AI to interact with consumers will generally have to disclose that fact. Employers using automated decision-making tools in hiring and other employment decisions will face notice requirements. Businesses that use personal data to increase online prices will have to tell consumers when an algorithm raised the price.

The state’s broader AI law requires covered generative AI providers to make synthetic content detectable, including by embedding information that can help identify whether an image, video or other digital material was created or materially altered by AI. The goal is to give consumers, platforms and investigators a way to distinguish synthetic media from authentic content, particularly as realistic deepfakes become easier to create and distribute.

It also places new requirements on AI companion services, which are designed to sustain an ongoing, humanlike relationship with users. Providers must disclose that the system is not human and establish protocols for responding when users express suicidal thoughts or other signs of self-harm. The law includes additional safeguards for minors, including measures intended to prevent companion chatbots from encouraging dangerous or self-destructive behavior.

For the companies developing the most advanced AI models, the law creates protections for employees who report potential catastrophic risks. Developers of advanced, or frontier, systems may not retaliate against workers who disclose safety concerns, and the largest companies must maintain an internal process for anonymous reporting. Those provisions are aimed at giving employees a way to raise alarms about dangerous systems before problems become public incidents.

Beyond the state’s AI law is a larger package of measures taking effect this year that also address facial recognition, data brokers, genetic privacy and surveillance pricing.

The facial-recognition provisions require businesses to disclose when the technology is being used and place limits on how related personal data can be handled. The data-privacy measures also restrict the sale of precise geolocation and genetic information and create new obligations for companies that collect and sell personal data.

The surveillance-pricing law targets online prices set using information about a shopper. In cases where a business raises a price using personal data, it must disclose that fact to the consumer.

Wayne Unger, an associate professor of law at Quinnipiac University School of Law, said state requirements need to be backed by enforcement.

“For these laws to have any effect, however, they need a strong enforcement mechanism and strict penalties to incentivize AI platforms to actually follow the laws and regulations,” Unger said.

What Connecticut could consider next

Disclosure is not the same, though, as an independent system for evaluating whether the most powerful AI models are safe before they are deployed.

“It’s not enough, but it is important,” Blumenthal said.

For Connecticut, the next step may be expanding the state’s law from notice requirements to stronger oversight of the systems themselves. That could include third-party audits, additional reporting to the attorney general, protections for employees who raise safety concerns and restrictions on high-risk uses, Maroney said.

Maroney did not describe a specific package of bills for the next legislative session. But he said lawmakers are discussing how to respond to increasingly capable AI agents, cybersecurity threats and the possibility that new dangers will emerge faster than lawmakers can anticipate them.

Independent audits to make sure companies are following safety protocols are one of the clearest ways states can move beyond disclosure, similar to what Illinois and other states have, he said.

Illinois and California have enacted laws directed at large developers of advanced AI models. The laws require companies to create safety frameworks, report certain catastrophic risks and protect employees who disclose safety concerns. Illinois also requires independent audits.

New York has also proposed requiring safety and security protocols, testing records, public disclosures and independent audits.

Maroney also said states should be willing to prohibit particularly harmful uses of AI rather than relying exclusively on disclosure. He pointed to nonconsensual, AI-generated sexual images as one example.

The state could also look at safeguards for AI systems connected to critical infrastructure, including hospitals, water systems and utilities, such as Montana’s requirement for a kill switch in technology used in these areas.

“What if autonomous bots decide that they are going to infiltrate critical infrastructure?” Maroney said.

Blumenthal’s federal proposal

Blumenthal believes Connecticut has a role to play, but he also sees the state as part of a larger national debate.

He has proposed an Artificial Intelligence Risk Evaluation Act that would create an oversight system modeled in part on the Food and Drug Administration. The proposal would require advanced AI models and products to be reviewed by scientists and experts before deployment.

“We evaluate those risks and we determine whether to move forward and what the safeguards can be to minimize the risk,” Blumenthal said. “I think that’s the approach we need here.”

Blumenthal said a federal law should establish a strong national baseline without preventing states from adopting stronger protections.

“We cannot trust big tech to, in effect, regulate itself,” Blumenthal said. “They have a profit motive that sometimes, in effect, trumps everything else.”

Blumenthal said he began working on AI regulation three years ago, including at a 2023 Senate hearing where OpenAI CEO Sam Altman and other industry leaders testified.

At that hearing, Blumenthal said, witnesses offered different estimates for when AI systems might become more intelligent than their creators. Some said three years. Some said five. Some said 10.

“But there was no question that this danger existed then,” Blumenthal said.

The Trump administration has taken the opposite approach to state regulation.

In December, President Donald Trump signed an executive order directing the attorney general to establish an AI Litigation Task Force to challenge state laws considered excessive, unconstitutional or inconsistent with national policy.

The order called state-by-state regulation a costly patchwork and directed federal agencies to consider whether funding could be conditioned on states not enforcing laws that conflict with the administration’s approach.

In March, the White House released a proposed national framework that again called for preempting state laws that impose undue burdens on AI companies. The administration said states should retain authority over child safety, consumer protection, fraud, state procurement and other traditional areas of law, something Anthropic has echoed.

But it also said states should not regulate AI development itself, which the administration described as an interstate activity with implications for national security and foreign policy.

That has put Connecticut’s emerging regulatory system in a precarious position. The state is writing rules while the federal government is considering whether to limit the authority that makes those rules possible. The industry, however, has added another complication.

OpenAI has said federal guidelines would promote innovation and create a level playing field. The company has also described state action as a possible form of “reverse federalism,” in which states establish a baseline that Congress later adopts.

Anthropic has called for government regulation on the safety of the most powerful models while arguing that federal preemption should be “surgical.”

In a recent video, U.S. Sen. Chris Murphy, D-Conn., said the companies’ calls for regulation should not be accepted without scrutiny.

“The most alarming near-term danger is what researchers call recursive self-improvement,” Murphy said. He described systems that could build increasingly capable versions of themselves “until the technology basically just outruns our ability, humans’ ability to control it.”

“The problem is that most of the AI leaders think that they should get to decide the solution, not you,” Murphy said. “AI companies have no incentive to protect us. They answer to investors, not to the public.”