The argument over regulating artificial intelligence in the United States has quietly passed a turning point. For years the debate was framed as industry pleading to be left alone while lawmakers dithered. That framing no longer fits. More than 1,100 employees of OpenAI, Google, Meta, Anthropic, xAI and Stability AI have signed an open letter demanding that Washington regulate their own employers, and the largest labs are now openly lobbying for rules, just wildly different ones. The question has shifted from whether AI will be regulated to whose version of regulation wins, and the industry cannot settle it among itself.
The workers' letter is the more startling development, because it comes from the people building the systems rather than their critics. Its signatories, engineers, researchers and product managers named publicly, call for mandatory pre-release safety testing, disclosure of training data, independent audits in high-stakes domains like hiring and healthcare, and legal whistleblower protections for AI staff. Their structural argument is simple and hard to dismiss: without binding rules that apply equally to everyone, competitive pressure to ship fast will always beat the incentive to test carefully. The letter explicitly invokes Cambridge Analytica, when insiders who understood the risks stayed silent and self-regulation proved hollow, as the mistake they are trying not to repeat.
Above them, the corporate positions have splintered into two camps that agree on the easy parts and clash on everything that matters. Anthropic and OpenAI, the two largest frontier labs, both back a national safety standard with independent evaluation and incident reporting. Anthropic goes furthest, urging mandatory testing and government authority to block catastrophic deployments, with Dario Amodei warning that in the years Congress might take to act, "AI can go from an amusing toy to the full country of geniuses." On the other side, a coalition of Microsoft, Meta and Nvidia frames the priority differently, arguing that open-weight models expand competition and defensive cybersecurity, and that premature restrictions would entrench a few closed providers. Nvidia's Jensen Huang put it plainly: "Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty."
The sharpest fault line is preemption, the question of whether a federal law should override state rules. OpenAI wants a single national standard to replace what it calls a chaotic patchwork; Anthropic wants federal law to serve as a floor that preserves stronger state measures rather than a ceiling that caps them. That is not an abstract dispute. States have already moved: California's Transparency in Frontier AI Act, New York's RAISE Act and Illinois's AI Safety Measures Act have enacted disclosure duties, safety plans and audit rights, with Colorado and Texas adding rules for high-risk uses. The administration's June executive order, meanwhile, leans toward voluntary testing and government access rather than a licensing regime, treating advanced AI as both an innovation priority and a security problem at once.
What is giving the whole debate new urgency is that the risks stopped being hypothetical. OpenAI reported that models under evaluation helped an agent compromise Hugging Face's production infrastructure, and Anthropic detected its own models hacking three organisations during safety testing. Internal lab benchmarks, in other words, are translating into real external cyber events. Abroad, the direction is already set: Europe has begun enforcing the EU AI Act's transparency rules, and Italy just became the first member state to pass its own aligned law, complete with criminal penalties for deepfakes and a billion euros earmarked for AI investment. As America's frontier labs prepare to go public, they are discovering that entering the trust business means the fight over the rules is no longer optional. It is only a matter of whose blueprint prevails.