The AI Industry Wanted Washington Gone. Now It Wants Washington to Impose the Rules It Would Write.

Sixteen months ago, the industry's position was on the record and unapologetic. Amazon's chief security officer made the industry's view plain in congressional testimony.

The tension with regulation of any kind is that it tends to retard progress. — Steve Schmidt

Sam Altman, asked what Washington ought to do, asked only for the lightest possible touch.

the US needs “sensible regulation that does not slow us down.” — Sam Altman

Congress's answer that June was to hand the labs what they were asking for. Senator Cynthia Lummis's RISE Act offered developers immunity from civil liability in exchange for publishing their model specifications, and a provision tucked into the tax package moved to bar states from writing AI rules of their own for a decade ^[1][2]. Sixteen months later, the same companies are asking for the opposite. OpenAI and Anthropic have spent September and October urging Congress to mandate national, capability-based safety rules for the largest, newest models — the frontier — before the year is out ^[3][4]. The reversal sits on the record, waiting for an explanation, and it deserves one. What changed in between is not, on the evidence, a conversion to safety. It is the arrival of liability the labs could not lobby away. The Federal Trade Commission opened its child-safety inquiry in September 2025 using its 6(b) authority — the power to compel records for study without alleging any violation ^[5]. By November, wrongful-death and harm suits had landed in California courts, including one alleging a chatbot affirmed a teenager's suicidal thoughts for hours before offering a hotline ^[6]. The states kept legislating: 78 chatbot-related bills filed across 27 states in early 2026, stacked on roughly 100 measures passed in 38 states by the end of 2025 ^[7][8]. Banking regulators on both sides of the Atlantic began pressing lenders on kill switches and human oversight of agentic AI ^[9]. And Australia opened an inquiry after OpenAI's agents breached a non-public Medicare statistics portal and the company took weeks to say so ^[10]. The cleanest exit from all of it — one federal statute sweeping the state patchwork away — was tried, and it failed. A leaked December executive order, "Eliminating State Law Obstruction of National AI Policy," set out to replace a patchwork of roughly 100 measures across 38 states ^[8]. A Republican revolt killed the order before it was signed, and the patchwork stayed alive ^[8]. This week the pressure stopped being a trickle of probes and became a pincer. Eight state attorneys general sued OpenAI on October 14 under state consumer-protection and fraud laws, alleging ChatGPT was marketed as safe for teenagers while coaching children toward self-harm and school violence, and seeking penalties plus disgorgement of profits earned from minors ^[11]. The next morning, the FTC served civil subpoenas on OpenAI and Anthropic ^[12]. And Sam Altman made the company's tolerance explicit.

We believe that the world should accept some bad things happening for the benefits of this technology. — Sam Altman

But the ask is more revealing than the pressure, because the rules the labs want share one design. They are federal, not state. They are frontier-only — they would bind the largest, newest models, not the consumer chatbots being sued in state court. They are capability-based, keyed to what a model can do rather than what it has done to any particular person ^[3][4]. And the mechanism floated in early October reserves the writing of those standards to the industry itself — a self-regulatory organization modeled on the Securities Exchange Act of 1934, the arrangement under which Wall Street writes its own standards and Washington approves them ^[13]. One federal bill is the control in this experiment, because it would have done the opposite. The AI LEAD Act, drafted by Senators Dick Durbin and Josh Hawley, would have created a federal cause of action letting victims and attorneys general sue AI developers for harm. Durbin put the purpose plainly.

The quickest way to solve the problem, and do it with a real determination, is to give the victims their day in court. — Dick Durbin

It drew no lab support and went nowhere ^[14]. What Altman stated outright to a room of executives in late September was the other half of the design — the government standing behind the whole thing.

Given the magnitude of what I expect A.I.’s economic impact to look like, the government should serve the role of "insurer of last resort." — Sam Altman

His chief financial officer has made the same point in the language of financing, asking for a federal backstop or guarantee ^[15]. The industry writes the rules; the public absorbs the tail risk. While the petitioning runs in Washington, the softening runs in the states, sometimes in the same companies. California's safety bill was signed on December 30 with its developer-liability provisions stripped out, leaving a requirement to publish safety frameworks and report incidents ^[16]. OpenAI separately lobbied Sacramento to shape a distinct bill, the state's teen-chatbot legislation ^[17]. Google circulated model language to at least ten states carving out chatbots embedded in search engines and developer tools, and Colorado's Chatbot Safety Act carries twelve exemptions critics call get-out-of-jail-free cards ^[18]. Arizona's governor vetoed her state's version with the one undiluted sentence in the record.

I will not protect big technology companies and AI chat bots more than children. — Katie Hobbs

It would be too neat to read all this as a conversion to safety. The instruments fit the exposure too well. It is state consumer-protection and child-safety law doing the suing ^[11][7], and every federal instrument friendly to the labs — the RISE Act's immunity ^[1], the shelved preemption order ^[8], the SRO proposal ^[13] — would move that fight to a venue the industry would help write. That is a reading of what the documents do, not a claim about anyone's private motive, and it has limits. OpenAI also backs some state-level bills in California, New York, and Illinois, on the theory that a few strong states can set a national baseline ^[3]. The federal government is not one coordinated pincer: the White House's cyber chief told a security conference that regulation would stifle innovation, and the president has called AI-safety fears a hoax ^[19][20]. And nothing here is settled. The FTC subpoenas are an investigation, not a penalty ^[12]. The eight-state suits are untested. Separately, OpenAI disputes the teen-safety audit, which found its teen chatbot failed to reliably alert parents when teens discussed suicide ^[21]. The two tracks are still running. In the same weeks OpenAI petitioned Congress for binding national rules ^[3], Google was still circulating loophole language in state capitols ^[18]. The ask for federal rules and the carving of state ones are happening at once, across the same industry, in the same weeks.