Who Limits AI in America

The question in the title has an answer, and it is not a single institution.

Congress declined to take the job. The executive branch is trying to stop the states from taking it. The states took it anyway. And the people currently deciding, case by case, whether a limit holds while the case runs are federal district judges hearing three lawsuits that one company brought against state attorneys general, in one of which the United States is not an observer but a party on the company's side of the caption.

That is the whole argument, and every part of it comes from a document we opened and read in the original: the executive order's own text in the Federal Register, the Federal Trade Commission's proposed policy statement, the Congressional Record's own account of the Senate vote, a district judge's memorandum opinion, the Justice Department's filing in that case, the docket text of the two other cases, the rule and the statute one of them turns on, and the Council of Economic Advisers' report that supplies the growth number everyone quotes. We do not test products and we ran no lab. This is a reading of public documents, with the dates we read them attached.

One thing to keep in view throughout. Nothing below says the case for a single national rule is stupid. It is the strongest argument in the file, it is set out as its holders would set it out in the counter-case panel that follows this piece, and the part of it that is right is not the part being litigated.

The Senate struck the AI provision from the reconciliation bill by 99 to 1, and no statute has replaced that vote

The Congressional Record's own Daily Digest, in the Senate Chamber Action section for the legislative day of Monday 30 June 2025, records it under the heading "Adopted:" as follows: "By 99 yeas to 1 nay (Vote No. 363), Blackburn Amendment No. 2814 (to Amendment No. 2360), to strike the section relating to support for artificial intelligence." The measure was the One Big Beautiful Bill Act, H.R. 1, and the page is D678.

Three notes on that citation. The Digest files the vote under the legislative day of 30 June; coverage generally dates it 1 July, the calendar day, and we cite the legislative day because that is what the official record prints. The vote number and tally come from the Congressional Record. And the provision that amendment struck is the one usually described as a ten-year moratorium on state enforcement of AI laws, tied to broadband funding; we did not read the struck text itself, so we describe it no further than the Digest does.

Since that vote, no federal statute has preempted state AI law. We are not asserting that from an absence we assembled ourselves. We are asserting it because two federal instruments written by lawyers whose entire purpose was to find such a statute could not cite one, and both had to reach elsewhere instead. That is the subject of a later section.

The subtitle above says Congress said no twice. The second no is the one we can describe but not document: in early December 2025 the preemption language that the administration wanted attached to the annual defence authorisation was abandoned, and the bill became Public Law 119-60 on 18 December 2025 without it. We verified the public law number, the originating bill, S. 1071, and the approval date at the Government Publishing Office. We could not open a primary source for who abandoned the language or when, so we name no member and quote no one, and nothing in the argument below rests on the second refusal.

Executive Order 14365 does not regulate AI, it regulates the states that do

Read the order's operative directives in order and the pattern is unmistakable: seven directives across sections 3 through 8, and not one of them places an obligation on an AI developer. Every one places an obligation on a federal official to act against the states.

Executive Order 14365, "Ensuring a National Policy Framework for Artificial Intelligence", was signed 11 December 2025 and published in the Federal Register on 16 December 2025 at 90 FR 58499. It has nine sections; the first two state purpose and policy and the last states general provisions. Its section 3 directs the Attorney General to establish, within 30 days, an AI Litigation Task Force "whose sole responsibility shall be to challenge State AI laws inconsistent with the policy set forth in section 2 of this order, including on grounds that such laws unconstitutionally regulate interstate commerce, are preempted by existing Federal regulations, or are otherwise unlawful in the Attorney General's judgment."

Here is what each directive required and what we could establish about it, read on 12 September 2026.

The column on the right is the finding. Of those seven directives, the one that has demonstrably produced binding action in the world is the first, the one that files lawsuits. The funding conditions, the naming of onerous laws, the telecoms standard and the trade-practice statement are, on the public record we could read, announced rather than operative.

That is an absence, and an absence is not an accusation. Reports get delayed, and a document we could not retrieve may exist. But the absence falls somewhere specific, and it is worth naming where: section 5(a) makes a state's eligibility for federal broadband money turn on a list that we cannot show has been published. A state cannot comply with a list it has not seen, and cannot contest its place on one either. The order also closes that door explicitly in section 9(c): it "is not intended to, and does not, create any right or benefit, substantive or procedural, enforceable at law or in equity by any party against the United States".

The child-safety carve-out everyone quotes applies to the legislative recommendation and not to the litigation

Section 8(b) of the order is narrower than its reputation, and the narrowness is textual rather than arguable. It says the "legislative recommendation called for in subsection (a) of this section shall not propose preempting otherwise lawful State AI laws relating to" child safety protections, AI compute and data centre infrastructure other than generally applicable permitting reforms, state government procurement and use of AI, and "other topics as shall be determined".

Read that against the order's own structure. The restraint attaches to what the recommendation may propose. It does not attach to section 3, which authorises the Task Force to challenge state AI laws inconsistent with the order's policy, nor to section 4, nor to the funding conditions in section 5. A reader who has only seen the carve-out summarised would reasonably conclude that child safety laws are off the table. In the text, they are off one table.

The three cases below are the test of whether that distinction is real rather than pedantic, and they resolve it. In one of them the United States filed against a state nudification statute. In another it became a plaintiff in a challenge to the Colorado law that Executive Order 14365 itself describes, in its first section, as "a new Colorado law banning ``algorithmic discrimination''".

Neither the Commission nor the Department can cite a statute that preempts state AI law, so one cites 1938 and the other cites an executive order

This is the single most load-bearing thing we found by reading the two federal instruments rather than summaries of them, and it takes three paragraphs to show.

The Federal Trade Commission's proposed policy statement, "Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems", File No. P264200, published at 91 FR 41638 on 7 July 2026, has to build its preemption theory out of the Wheeler-Lea Act. It opens its authority section with "Nearly ninety years ago, Congress gave the Commission the authority to protect consumers from 'unfair or deceptive acts or practices in or affecting commerce.'" Then it states the theory outright: "Although the FTC Act does not expressly preempt State law, State law is impliedly preempted to the extent it conflicts with a Federal regulatory scheme." The example it names is "Colorado's recently revised Artificial Intelligence Act".

The authority it gives for that proposition is not a statute either. It is a case, cited in the notice's own footnote to the sentence: Schneidewind v. ANR Pipeline Co., 485 U.S. 293, 300 (1988), quoted there for the rule that "[S]tate law is pre-empted when it actually conflicts with federal law. Such conflict will be found when it is impossible to comply with both state and federal law, or where the state law stands as an obstacle of the full purposes and objectives of Congress." So the chain the Commission is standing on runs from a 1938 Act to a 1988 case to a 2026 proposal, and there is no AI statute anywhere in it. We are not saying the theory is wrong. We are saying what it is made of, which the Commission itself sets out plainly and almost nobody quotes.

The Justice Department does the same thing from the other direction, and the order tells it to. Executive Order 14365's section 3 instructs the Task Force to argue, among other grounds, that state AI laws "are preempted by existing Federal regulations". Regulations, not statutes; that is the order's own wording. In the Minnesota court, the Department's Statement of Interest, filed 18 August 2026 under 28 U.S.C. 517, never uses the word preempt at all about H.F. 1606. It grounds the federal interest in policy: "it 'is the policy of the United States to sustain and enhance the United States' global AI dominance through a minimally burdensome national policy framework for AI'—one 'national standard,' rather than '50 discordant state ones.'" The citations for that proposition are Executive Order 14365 and Executive Order 14409. The federal statutes it does cite, the PROTECT Act and the TAKE IT DOWN Act, are cited for the opposite purpose: to show that federal law is narrower than Minnesota's, not that it displaces it.

So the status of every instrument in this story, stated plainly, because overstating one of them would discredit the rest:

One further thing sits oddly with the preemption campaign, and it comes from the administration's other AI order. Executive Order 14409's section 3(c) reads: "Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models." Its section 4 puts the federal limit at the far end instead: "The Attorney General shall prioritize the enforcement of 18 U.S.C. 1028, 18 U.S.C. 1030, 18 U.S.C. 1343, and all other applicable Federal criminal laws against anyone who utilizes AI to illegally access or damage a computer without authorization", including "employing AI agents to unlawfully access data or information that is subsequently used for a criminal or unlawful purpose."

Put the two orders side by side and the federal answer to the question in our title is consistent and complete. Nobody licenses a model before release. Prosecutors act after a crime. And the states are to be sued for doing anything in between.

One company is the plaintiff in all three cases, and the United States has taken a different posture in each

The three courtrooms in the subtitle are federal district courts, one in each of three states, and the same company is the plaintiff in all three. What differs, and what almost nobody has set side by side, is what the United States did in each.

Everything in this table comes from the case documents or the docket text, read on 12 September 2026. Where the docket does not say something, the cell says so rather than filling the gap from elsewhere.

Two disciplines about that table, and they matter more than anything in it.

The California docket does not name the statute. It records a constitutional challenge to a California state statute and nothing more specific than that, so we print no bill number for it. Coverage of that case names a particular statute. We did not read that name on the docket and we are not going to supply it.

The California case is the cleanest illustration of the claim in this piece and the one nobody quotes. A preliminary injunction was denied on 4 March 2026 by minutes in chambers, and by 23 March the district case was terminated and the fight had moved to an appellate court. Note what did not happen between the complaint on 29 December 2025 and the termination on 23 March 2026: no federal statute, no agency rule, no Commerce list. A district judge held a hearing on 23 February and the question moved up one level. That is what it looks like when the arbiter of national AI policy is whichever judge drew the case.

Minnesota is the only place in this story where a product actually changed, and it took eight days

Everything above is about who may set a rule. Minnesota is where a rule met a product, and the product moved first.

H.F. 1606 prohibits a person who owns or controls a service from allowing a user "to nudify an image or video" or doing it on the user's behalf, with an exemption where the service "requires the technical skill of a user" to do it, and civil penalties up to $500,000 for each violation enforceable by the attorney general. The statute defines nudify to cover an image "altered or generated to depict an intimate part not depicted in an original unaltered image or video of an identifiable individual" where the result "is so realistic that a reasonable person would believe that the intimate part belongs to the identifiable individual." All of that is quoted from the court's memorandum opinion of 4 September 2026, which we read in full.

Eight days. That is the interval in which a frontier developer built jurisdiction-specific controls into an image pipeline, and we have it because the company put it in a declaration to persuade a judge that compliance was burdensome. The same document persuaded the judge of the opposite. The court's reasoning on irreparable harm is worth reading in its own words rather than paraphrase:

Two limits on what that table can carry. The court did not decide the constitutional question and said so: "The constitutional issues raised by the parties are complex... These issues deserve, and will receive, full consideration." And the figures the opinion recites from the legislative record, including RAINN's account of 24 million users of nudification apps in one month in 2023, the Internet Watch Foundation's 6,800 reports of generative child sexual abuse material in the first half of 2024 rising to over 400,000 through June 2025, and a study finding approximately 65 percent of the photos created in the first 11 days after one social platform launched an editing tool were photorealistic sexualised images, "amounting to a total of 3 million such images at a pace of 190 images per minute" with roughly 23,000 of children, are the court's summary of testimony and submissions. We have not read those underlying sources and report them as the court's record of what the legislature heard, not as findings of our own.

The same opinion records what the company says it does about misuse, and fairness requires printing it: xAI's complaint claims that in 2026 alone it suspended over 50,000 accounts and made over 70,000 reports to the National Center for Missing and Exploited Children, "resulting in at least 244 arrests". The state's answer was not that those numbers are false but that after-the-fact enforcement is a different remedy, and in the passage finding the balance of harms tips steeply towards the state, the court records the state's conclusion "that curtailing user violations after harmful images are circulated is not enough."

The United States filed against a law the Minnesota Senate passed unanimously, and its four objections are serious

This is the hardest fact in the Minnesota file and we are not going to soften it. On 18 August 2026 the federal government filed in a challenge to a state statute that passed one chamber of a state legislature and passed the other unanimously, and its filing cited two executive orders by name while doing so.

It also made four specific legal objections, and they are not decoration. Each is a real drafting difference between H.F. 1606 and federal law, quoted here from the filing we read.

The filing also observes, correctly, that Minnesota's own separate deepfake statute uses a narrower definition of intimate parts than H.F. 1606 does. An objection that a legislature used two different definitions in two of its own statutes is a good objection.

What the filing does not contain is a federal statute that displaces the Minnesota one. Its own framing is that federal law is "carefully calibrated" and the state law "sweeps more broadly than federal law", which is an argument that Minnesota went further than Congress, not that Congress forbade going further. Those are different claims, and only one of them is preemption.

In Colorado the United States is not a friend of the court, it is a co-plaintiff

On 24 April 2026 the United States did not file a brief in the Colorado case. It moved to intervene as a plaintiff, unopposed, and the court granted the motion the same day.

The motion carried two exhibits, per the docket text: a Certificate of the Acting Attorney General, and a Complaint in Intervention. That Complaint in Intervention was docketed the same day, as entry 17. The order granting intervention rests on Federal Rule of Civil Procedure 24(a)(1), and it is worth being exact about how the order handles the rule, because the difference is visible on the docket. The order quotes the rule in a flowing form of its own, rendered on the docket as "anyone to intervene whois given an unconditional right to intervene by a federal statute", with "who" and "is" collided. That is what happens when a court quotes the rule continuously and drops the subsection number. The rule as published is not continuous. It reads "(a) Intervention of Right. On timely motion, the court must permit anyone to intervene who:", then breaks to "(1) is given an unconditional right to intervene by a federal statute; or", then a second numbered subsection about an interest in the property or transaction at issue. We read that in the Government Publishing Office's edition of the Federal Rules of Civil Procedure, in the appendix to title 28, on 12 September 2026. The order relies on the first subsection.

The order then identifies the federal statute that supplies the right: the Civil Rights Act of 1964, which gives the United States an unconditional right to intervene in equal protection cases where the Attorney General certifies that the case is of general public importance.

Read the provision that description matches, because it is the part of this story that the documents make plainest and the coverage makes least. The Civil Rights Act of 1964 contains one such right, at title IX section 902, codified at 42 U.S.C. 2000h-2, which we read at the Government Publishing Office on 12 September 2026. Its heading is "Intervention by Attorney General; denial of equal protection on account of race, color, religion, sex or national origin". It applies "Whenever an action has been commenced in any court of the United States seeking relief" and specifically relief "from the denial of equal protection of the laws under the fourteenth amendment", the operative text continuing "to the Constitution on account of race, color, religion, sex or national origin". The Attorney General "may intervene in such action upon timely application" and only on a certification "that the case is of general public importance". Where he does, "the United States shall be entitled to the same relief as if it had instituted the action". The section's source credit records its origin: Public Law 88-352, title IX, section 902, 2 July 1964, and an amendment by Public Law 92-318 in 1972.

That last clause of the trigger is the one to hold onto, because it is the element that decides whether the right reaches a case at all. The statute is not a general equal-protection intervention power. It reaches the denial of equal protection on account of race, colour, religion, sex or national origin, and nothing in the docket text tells us how the certification satisfied it here. The Certificate of the Acting Attorney General is an exhibit we have not read.

So the sequence, from the docket, with nothing added.

The court's stated reasons for the standstill are on the docket and they are procedural rather than constitutional: a stay will not prejudice the plaintiff, who does not oppose the motion; it will avoid burdening the parties with the costs of litigating; it promotes judicial economy and efficiency; it does not negatively impact nonparties; and it is in the public interest.

Now set the two federal postures beside each other, and then we will stop. In Minnesota the United States filed a statement of interest under 28 U.S.C. 517, a document that binds nobody, asks for nothing and left the state to win its motion. In Colorado the United States became a party on the same side of the caption as the company, using an intervention right that Congress wrote into the Civil Rights Act of 1964 for the enforcement of equal protection, on a certification by the Acting Attorney General that the matter is of general public importance. The case was filed on 9 April. On 27 April the state agreed not to enforce its law.

We are not going to characterise that, because the documents do not need help. We will note two things that a reader should hold onto, both of which cut against reading it too hard.

The first is that nobody struck anything down. The standstill came from a joint motion, and the court's reasons are about cost and judicial economy. A state that agrees to stop enforcing can also resume. The accurate sentence is not that a judge stopped Colorado's law; it is that Colorado stopped enforcing it after the federal government joined the case against it.

The second is what that means anyway. For a developer or a deployer subject to that statute today, an enforcement standstill of unstated duration and a repeal are the same document. The legislature that passed it is not the body that ended it, and the body that ended it did not rule on whether it was lawful.

Is the patchwork argument right?

Partly, and the part that is right is not the part being litigated. The definitional mismatch is real: the same generated image can be lawful under federal law and a $500,000 violation in Minnesota, and no amount of geofencing fixes a definition.

But look at what the record actually establishes about cost. The one company litigating this told a federal judge that compliance was burdensome and produced, as evidence, a declaration the judge found "vague and conclusory", describing work that took eight days and had already been completed. The order's own premise is that the patchwork "makes compliance more challenging, particularly for start-ups". We could not find a single filed figure for what it costs a start-up. The plaintiff best placed to produce one produced an eight-day engineering sprint, and the argument for the start-up is being made by the party least like one.

That asymmetry is the honest state of the question. The patchwork imposes a real cost that nobody in this file has quantified, and the entity carrying the argument is the entity least burdened by it. If someone publishes the small-developer number, we will print it, and the balance struck in this section moves.

There is one more thing the order's premise does not survive. The administration's count, quoted by the Commission in the Federal Register, is that "State legislatures have introduced over 1,000 different AI bills[.]" Introduced is not enacted, and enacted is not in force with operative obligations. The denominator that would tell a developer how many regimes actually exist is not published anywhere we could find, by any party to this argument.

The growth figure quoted beside this order was estimated six months before the order existed

The Council of Economic Advisers' January 2026 report, "Artificial Intelligence and the Great Divergence", is the administration's own economic case, and its deregulation paragraph is the source of the number that appears in almost every account of this policy. On page 25, in a section headed "Deregulation", it says: "As previously studied by CEA, deregulatory efforts by the Trump Administration aim to ameliorate these issues and deliver benefits to the AI sector along with other sectors throughout the economy. CEA estimates that these deregulatory efforts are capable of delivering meaningful productivity gains that translate into an extra 0.3 to 0.8 percentage points of GDP growth each year for two decades—a cumulative increase of approximately 6 to 17 percent by 2045."

That is the administration's claim, attributed, and we are not evaluating it. What reading the page adds is the provenance of the estimate, which is visible only in the footnotes.

So the number cited beside the state-preemption order was published roughly six months before that order was signed, was computed for the administration's deregulatory programme generally rather than for AI preemption, and sits in a report that nowhere prices the thing the order is about. None of that makes the estimate wrong. It does mean that anyone citing 6 to 17 percent as the payoff from preempting state AI law is citing something that was not an estimate of preempting state AI law.

The money is on one side of this and the polling is on the other

Neither of these decides anything, and we are putting them in one short section for that reason. They describe the environment the decisions are being made in, and we make no causal claim from either.

The political committee registered with the Federal Election Commission as LEADING THE FUTURE, committee ID C00916114, reports the following for the 2026 cycle. We read these at the Commission's own interface on 12 September 2026.

One detail matters more than the totals, because getting it wrong inflates the figure by $25m. The two firm contributions carry the memo text "SEE ATTRIBUTION BELOW", and the individual partner lines that follow them, including two at $12,500,000 in the name "ANDREESSEN, MARC L.", carry memo code X and the memo text "PARTNERSHIP ATTRIBUTION". Those are attributions of the firm's money to its partners, not additional money, and they must not be added to the total. If a larger network figure is reported elsewhere, it is someone else's figure and it counts something else. We did not examine what the $44.8m in disbursements was spent on, and we are not suggesting that any of this money produced any of the documents above.

On the other side, the Annenberg Public Policy Center's Institutions of Democracy survey, fielded 17 February to 20 March 2026 with a nationally representative sample of 1,330 US citizens aged 18 and over by SSRS for APPC, published 13 May 2026 and read by us on 12 September 2026:

The release does not publish subgroup sample sizes, so no precision should be read into the party rows beyond their order. What survives that caveat is the direction: in this one survey, a majority of every named group said too little, not too much. An executive order whose premise is that state regulation is excessive is operating against a public that, in the only fielded measurement we verified, says the opposite.

What a vendor could adopt and a regulator could point at

Each row below is specific enough to be implemented by one party without anyone else's permission. Rows 1 and 5 are the two we would take first, because they are the two that would let anyone outside this argument check it.

So who limits AI in America

Today, and on the documents rather than on anyone's intentions: a state legislature writes the limit, a company decides in eight days whether to comply with it, a federal district judge decides whether the limit holds while the case runs, and the United States shows up on the company's side, in one state as a commentator and in another as a co-plaintiff. Congress is not in that sentence, and neither is any federal instrument that is final.

That is not the same as saying nobody limits AI. Minnesota does, right now, for anyone serving Minnesota. It is saying that the American answer to who decides has been arrived at by default, through a route nobody would have designed: a recorded 99 to 1 vote that settled nothing, an executive order whose only visibly operative directive is the one that files lawsuits, a Commission reaching for implied preemption because express preemption does not exist, a Department whose task force is instructed to argue that state laws are preempted by existing federal regulations rather than by any statute, a civil-rights intervention right from 1964 used to make the federal government a plaintiff against a state AI statute, and three courtrooms in three states where the arbiter of national AI policy is whichever judge drew the case.

The strongest objection to all of that is the patchwork, and it is set out in the counter-case panel that follows this piece rather than answered here in a sentence. We think it is serious, we think the definitional mismatch between Minnesota and the TAKE IT DOWN Act proves it, and we think the fix for it is the one instrument in this entire file that nobody has produced: a statute. Every other route on offer asks a court, an agency or a grant officer to do the work of a legislature, and the one time this question went to a legislature, the answer was 99 to 1.

What we read, and what we could not open

Everything below was read on 12 September 2026 unless stated otherwise. We do not test products, we run no lab, and no part of this piece rests on our own research. It is a reading of documents their publishers put in public.

What we could not open, and where a correction would most likely come from.

The Commerce evaluation of state AI laws required by section 4 and due 11 March 2026. We looked in the Federal Register's full-text search, including an exact-phrase search for "State AI laws" which returns only the order and the FTC notice; in the FCC's own 2026 Federal Register documents; and in general web search. We found no published evaluation and no notice of one. We are not saying it does not exist. We are saying we could not find it, and that the funding condition in section 5(a) depends on it.

The pleadings and orders in the California and Colorado cases. We read the docket text for both and no filing from either. That means we have not read the United States' Complaint in Intervention, the Certificate of the Acting Attorney General, the order granting intervention beyond the passage the docket carries, the joint motion behind the enforcement standstill, or the California minutes denying the injunction. Every statement about those two cases in this piece is a statement about what the docket records.

The statute challenged in California. The docket does not name it. Coverage does; we did not read that name on the docket, so no bill number or section appears here for that case.

The appellate proceedings. We did not open the Ninth Circuit docket in the California appeal and give no number, schedule or ruling for it. An appeal from the Minnesota order of 4 September is also reported; we did not open an appellate docket for it either, and our line that the Minnesota law is not enjoined rests on the district court's own order denying the injunction.

The Attorney General's memorandum of 9 January 2026 establishing the Task Force. It exists and it is dated, because the United States cited it in a filing we read. We quote nothing from it.

The White House "National Policy Framework for Artificial Intelligence" of 20 March 2026, cited twice in the Commission's notice. We describe none of its contents.

The second congressional refusal. We could not open a primary source for the December 2025 decision to leave preemption out of the defence bill, so we name no member, quote no one, and rest no part of the argument on it.

If any of those contradicts a line above, we would like to be shown it, and the line will change.