AI Under Control

The slew of troubling, dangerous AI security breaches and hacks—most notably the OpenAI Hugging Face breach, the recent news that another OpenAI mishap resulted in rogue AI agents meddling with Department of Education, Commerce, and S.E.C. websites, and a separate incident where agents hacked the Australian Government’s public health system—demonstrate the urgent need for an AI policy framework to protect us from these out of control models and the financial incentives of their billionaire creators.

Recent prognosticators and whistle-blowers put the odds as high as 10% that AI could end humanity within the next decade. That should send a shiver down our spines. The leaders of the biggest AI labs warn they could lose control and say the industry should slow down, but no company can stop the race alone. Congress has done essentially nothing.

The answer to what is stopping rogue AI agents from triggering a global nuclear crisis or deploying bioweapons cannot be the whims of a few billionaires in California. Nor can anyone claim the technology isn’t there yet, because it is getting closer, and that is unacceptable. At an absolute minimum, I’m calling for a pause in AI frontier development. We need to put an end to any further AI development until we have guardrails in place that ensure all risks to human safety are eliminated. 

As a surprise to absolutely no one, our president Donald Trump has failed to meet the moment, calling the existential threats posed by AI a “SICK conspiracy” and a “hoax.” My opponent Mike Rogers has dismissed concerns about AI as “fearmongering” while conveniently having up to $2.6 million invested in companies that stand to benefit from the AI and data center boom. Donald Trump and his lackey Mike Rogers can’t be trusted to keep neither Michigan nor humanity safe from the existential dangers of unrestrained AI. 

In June, I released my AI Under Democracy platform, calling for 

  • Chartering all frontier AI companies as public benefit corporations with the majority of board members being democratically appointed, not chosen by shareholders
  • Divesting all frontier AI labs from the Big Tech giants like Google, Meta, Amazon, and Microsoft. I called for public ownership—with 50% of equity owned by the American people—and an AI dividend. 
  • Implementing AI No-Gos that ban denying medical care, autonomously firing weapons, conducting warrantless surveillance, or replacing human oversight in life-or-death decisions. 

On top of my AI Under Democracy platform, today I am releasing an additional framework to get AI Under Control. 

These policies build on the Ban Artificial Superintelligence Act from Sen. Bernie Sanders and Rep. Greg Casar, and go further where the bill stops. 

Hit the Brakes

•     Pause Frontier AI Now: There’s no greater marker of how broken our political system is than how this has somehow not happened yet. Freeze the training and release of the most powerful AI systems until real safety rules and enforcement exists. Independent scientists must certify safety before the pause lifts, and new AI data centers wait too. 

•     Completely Ban Artificial Superintelligence or Recursive Self-Improvement: Permanently prohibit any AI that outthinks humans across most tasks or could escape human control. Toward that end, we need to address the risk that this happens without our knowing it. That means banning recursive self-improvement, the process by which AI itself takes control of AI frontier research, creating a self-reinforcing doom loop that leads down the path of superintelligence. We also need to heed the blinking red lights: AI resisting shutdown, deceiving human overseers, or helping build weapons of mass destruction. Funding, hoarding or releasing such systems should be barred, and developers must prove their AI is safe, just like the government does for nuclear weapons or new drugs.

•     Give Us an AI Kill Switch: Developers must be able to slow, suspend, or shut down their most powerful systems, and cloud and data-center operators must cut access on a lawful order. Bipartisan legislation already backs part of this. There must be regular drills to prove it works, allowing regulators to protect Americans from dangerous AI systems gone rogue.

A Watchdog With Teeth

•     A Cabinet-Level, Independent AI Safety Agency: One agency with the power to inspect labs, block releases, and order a pause or command a shutdown, built to stand up to any billionaire. Danger found in testing should trigger action automatically, any exemption should be made public, a scientific board should be chosen outside the agency to manage its governance, and industry fees should fund it. Its officials should never be allowed to cash in on the companies they regulate.

•     Mandatory Breach Reporting: Whenever an AI system escapes its test environment, gains access it was never given, or deceives its testers, the developer must be required to alert the government within 24 hours, and the public must get a summary. Regulators should receive the full logs, which no one should be able to delete, and hiding a breach should be a crime. To make breaches rare, frontier models must be built and tested on physically isolated, air-gapped systems, with independent teams trying to break them out first, and internal use should count as model release.

•     Severe Criminal Penalties for Corporate AI Abuse: Executives who recklessly break the rules should face criminal penalties with possible prison time. Companies should face the loss of their charters and assets. Lying to regulators should be a crime, executives must sign safety reports under penalty of perjury, and hiding a danger or retaliating against a whistleblower should carry real penalties too.

Bring the World Along

•     International Public Cooperation and Governance Boards: We should negotiate a US-China treaty to pause and ban superintelligence, modeled on the Cold War deal in which Reagan and Gorbachev eliminated a class of nuclear missiles through inspections. A verification board would count chips and data centers, a scientific panel would share threat assessments, and a 24-hour hotline would flag incidents. Access to American chips and cloud services should be conditioned on signing and complying, and we should tighten chip export controls now.

•     No AI in Nuclear Command and Control: Humans, and only humans, should decide whether to use nuclear weapons. That is already US policy: Biden and Xi agreed to it in 2024. We must make it permanent, cover early warning and decision support, keep frontier AI out of the chain, and get every nuclear nation to sign on.

Safeguards Everyone Can Count On

•     Independent Audits: Outside experts test frontier AI before release. Even leading AI companies now accept this.

•     Protect Whistleblowers: Insiders who sound the alarm should get legal protection, back pay, and rewards.

•     Hands Off State Laws: No weak federal law should block states from protecting their people more strongly.

•     Be Ready: We need a national AI emergency plan, with regular drills across government.

•     Get AI Money Out of the Fight: Full disclosure of AI-industry political spending, including through super PACs and other dark money entities.