The series that translates national-security and AI-policy arguments across partisan lines, because the stakes are too high for tribal shorthand.
Three Doors at Once
What Happened This Week
The White House National Policy Framework for Artificial Intelligence, issued March 20, has been on Capitol Hill for eight weeks. This week it stopped being a document and started being a strategy. On Tuesday, May 12, the Senate Commerce Committee held its third closed-door briefing on the framework's recommendations to Congress. On Wednesday, May 13, the House Armed Services Committee circulated a discussion draft of the FY 2027 National Defense Authorization Act that included, in Title XVI, a sixteen-page section on "Federal Standards for Artificial Intelligence Procurement and Use." That language, read carefully, contains a preemption clause. And on Thursday, May 14, the Department of Justice AI Litigation Task Force filed an amended complaint in xAI v. Weiser, the Colorado matter the DOJ joined on April 24, arguing that the December 2025 executive order is itself a sufficient federal interest to ground a preemption claim.
Three doors. Same hallway. Same policy outcome on the other side. The Senate ask is the legislative front: a request that Congress pass a freestanding bill incorporating the framework's preemption recommendations. The NDAA route is the must-pass front: attaching the same preemption language to defense reauthorization, which Congress has not failed to pass in sixty-three consecutive years. The executive order is the litigation front: the December 2025 EO already exists, and the DOJ is now arguing in court that the EO alone is enough to preempt state law. The strategy is not a sequence. It is parallel.
The timing this week was not a coincidence. Connecticut's Senate Bill 5, the comprehensive frontier-AI framework that passed the state Senate on April 21 by a 32 to 4 vote, is on Governor Ned Lamont's desk, with his office confirming intent to sign in the days ahead. It will become the AI Responsibility and Transparency Act the moment he signs, which is expected within days. Connecticut is already functioning as a forcing function in Washington, and the signature will lock that in. The White House moved on all three doors in the run-up to the signing so that whichever door opens first, the federal preemption posture is already in motion by the time the first state provisions actually take effect this fall.
Here's What You Need to Know in 30 Seconds
The White House is not waiting to see which approach to federal AI preemption works. It is running three simultaneously. The legislative ask is the cleanest path, the NDAA ride-along is the most likely path, the executive-order litigation is the fastest path. The two preemption attempts that already failed (a reconciliation rider in summer 2025 and a similar provision in the FY 2026 NDAA) failed because they ran on one track at a time. The framework's strategy is to make refusing all three more politically expensive than accepting one. Industry lobbyists are filing comments, drafting model language, and testifying at all three venues with the same talking points. Connecticut's bill heading to the governor's desk makes door three (the EO route) operationally urgent, because the DOJ needs a live conflict to litigate the moment the law takes effect, and the AI Responsibility and Transparency Act is exactly that. Eighteen other states are watching to see whether any of the three doors closes before they introduce parallel bills.
The Hawk Case: Three Doors Means One Will Land
The hawk position on the framework's multi-track strategy is, on the merits, defensible. The hawk's argument is that federal preemption of state AI regulation is the right policy outcome, and that the previous two attempts (the summer 2025 reconciliation rider, the FY 2026 NDAA language) failed not because preemption is wrong on the merits but because the procedural posture in each case was vulnerable. The reconciliation rider failed under the Byrd Rule because the provision was deemed not primarily fiscal. The NDAA FY 2026 language failed because it was stripped in conference after the Senate Armed Services chair concluded it would attract a filibuster threat to the underlying bill. The hawk reads both failures as procedural rather than substantive: the votes for preemption exist; the vehicles did not.
The hawk further argues that the three-door strategy is the appropriate response to that history. If reconciliation will not work, try ordinary legislation. If a standalone bill cannot reach sixty votes, try riding on the must-pass NDAA. If the NDAA process is hostile, fall back to the executive order and let the courts decide whether the EO alone can ground a preemption claim. None of the three doors is a clean policy victory. All three together is, structurally, a higher probability than any one alone. The hawk reads this as competent policy execution by an administration that has correctly diagnosed the problem of state-level fragmentation and is using every legitimate procedural mechanism to address it. The framework, on this reading, is not forum-shopping. It is the executive branch declining to accept the proposition that policy failure on AI governance is acceptable.
The hawk also reads the timing on Connecticut as a feature. If the state law is going to take effect, the federal posture needs to be ready to engage it before the compliance window opens. The Connecticut AI Responsibility and Transparency Act becomes operative in January 2027. The framework's three-door strategy means that the federal preemption posture will be in litigation, in conference negotiation, and on the floor in some form by the time the law actually constrains anyone. The hawk sees this as the appropriate operational tempo. The reformer's complaint that the strategy is "too aggressive" is, in the hawk's view, an objection to the executive branch operating at the pace the problem requires.
The hawk's bottom line: If federal preemption of state AI regulation is the correct policy outcome (and the hawk believes it is), then the executive branch is doing precisely what it should do: using the legislative, NDAA, and litigation tracks in parallel to maximize the probability that one of them lands before the fragmentation problem becomes unrecoverable. The previous failures were instructive. This is the lesson the administration drew. Connecticut just made the urgency real.
The Reformer Case: Forum-Shopping Is Still Forum-Shopping
The reformer position concedes the hawk's diagnosis and rejects the prescription. The reformer agrees that federal AI legislation is overdue and that state-level fragmentation imposes real costs. What the reformer does not concede is that the executive branch is entitled to pursue the same policy outcome through three parallel mechanisms when the political branches have not produced consensus on any one of them. The objection is structural, not substantive. Three doors at once is, in plain terms, regulatory forum-shopping. It is the executive branch asking Congress, the conference committee, and the federal courts to ratify the same policy outcome on the theory that one of them will say yes for reasons the others would have said no.
The reformer further argues that the constitutional grounding on each of the three doors is weakest where the political grounding is strongest, and vice versa. Door one (a freestanding bill) is constitutionally clean but politically dead. Door two (the NDAA rider) is politically plausible but constitutionally problematic because attaching preemption language to defense reauthorization stretches the single-subject principle beyond its usual application, even by federal standards. Door three (the executive order as preemption authority) is fast but constitutionally thin: preemption doctrine has historically required either an act of Congress, a treaty, or a federal regulation issued under valid statutory authority. The EO is none of those. The DOJ's argument that the EO alone is sufficient is, in the reformer's reading, an argument the executive has not seriously made in any prior preemption case.
The reformer also points to the lobbying pattern. The same four firms (Davis-Wright, Wiley Rein, Covington, and Akin Gump) that filed comments on the framework's draft text in February are now testifying before Senate Commerce, briefing House Armed Services staff on the NDAA Title XVI language, and serving as outside counsel to the industry coalition supporting the DOJ's position in xAI v. Weiser. The reformer reads this as evidence that the multi-track strategy is being coordinated by the regulated industry rather than by the executive branch in service of any public interest. The lobbyists are not choosing between doors; they are walking through all three with the same brief. That, the reformer argues, is what forum-shopping looks like when the same client is the customer at every venue.
The reformer's bottom line: Preemption is a federal policy question that should be answered by federal legislation, not by an executive branch unwilling to accept congressional inaction as a constraint. The three-door strategy is a workaround for a deliberative process that the executive does not control and does not like. The right response to congressional gridlock on AI is to keep working on the legislative votes, not to outsource the question to the conference committee and the federal courts. The reformer is not defending Connecticut. The reformer is defending the proposition that the executive branch does not get to choose which constitutional venue produces the answer it wants.
Where They Actually Agree
The hawk and the reformer agree on the diagnosis: federal AI legislation is overdue, the December 2025 executive order is operating in legally thin territory, and the multi-track strategy reflects an executive branch that has correctly assessed the difficulty of passing a freestanding bill. They agree that the previous two preemption attempts failed for procedural reasons rather than because preemption is a politically unpopular policy. They agree that if a single straight up-or-down vote on federal frontier-AI preemption were held tomorrow, it would likely fail in the Senate by margins that would surprise both of them.
Both sides agree that the NDAA route is the most operationally important of the three. The FY 2027 NDAA will be on the Senate floor in September or October. The Title XVI preemption language is the kind of provision that can survive in conference if the chairs of both armed services committees want it to. Both chairs have signaled, in different terms, that they consider AI a national-security priority. Both have also signaled, in different terms, that they would prefer not to have preemption be the language that holds up a defense bill. The likely outcome, both the hawk and the reformer expect, is that some watered-down version of Title XVI survives, but the language survives in a form that may or may not actually preempt anything when tested in court.
Both sides also agree, with different valences, that the industry coalition is the actual driver of this week's pattern. The hawk reads the coalition as a legitimate interest expressing a legitimate policy preference through legitimate channels. The reformer reads the coalition as the regulated entity successfully capturing all three preemption vehicles simultaneously. They are describing the same coalition and the same channels. The dispute is whether industry coordination at this scale is normal lobbying or something larger.
Where They Don't (And Shouldn't Pretend To)
On whether the EO alone can ground a preemption claim. The hawk treats the executive order as a strong statement of federal policy that courts should defer to under standard preemption doctrine. The reformer treats it as an executive policy statement with no statutory grounding and therefore no preemption authority. Both are reading the same EO. The hawk argues that the EO operationalizes existing statutory authority under the Defense Production Act and the Commerce Department's general authority over interstate commerce. The reformer argues that neither statutory hook actually authorizes preemption of state AI legislation as such. Both readings are defensible. They will be tested in xAI v. Weiser within ninety days.
On whether NDAA preemption language is constitutionally legitimate. The hawk treats Title XVI as ordinary defense legislation: AI is a national-security issue, the Defense Department is the largest single AI customer in the federal government, and Title XVI's procurement-floor language is properly within the armed services committees' jurisdiction. The reformer treats Title XVI as a vehicle for preempting state law that has nothing to do with defense procurement, attached to a defense bill because the defense bill is the one that passes. Both readings are honest. The disagreement is over how strictly the relationship between a provision and its host bill needs to be enforced. Federal courts have generally not enforced single-subject limits on federal legislation. The constitutional question is whether they should start.
On the lobbying pattern. The hawk reads the same four law firms working all three doors as ordinary industry coordination on a single policy issue. The reformer reads it as evidence that the strategy is being designed by the regulated rather than by the regulator. Both are describing the same factual pattern. The dispute is whether the pattern is normal or whether it represents a different scale of coordination than what AI policy has previously seen. There is no neutral way to score that question. It is a judgment call about what counts as ordinary lobbying.
On Connecticut as catalyst. The hawk reads the AI Responsibility and Transparency Act as exactly the kind of state-level overreach that justifies the three-door strategy. The reformer reads it as exactly the kind of state-level governance that the federalist system was designed to produce when the federal government cannot legislate. Both readings are consistent with the bill text. The disagreement is normative: should states be allowed to fill federal vacuums on AI, or should the federal government close the vacuum by any procedural means available? Neither side is going to convince the other on that question.
Here's My Two Cents
Three doors at once is precisely the kind of policy mechanism that does not appear in any AI bill and determines what AI policy looks like for the next decade. None of this strategy is in the framework's text. None of it is in Title XVI. None of it is in the DOJ's amended complaint. The strategy lives in the relationship between the three documents, and in the choice to file all three in the same week. That choice is the policy. The substantive text is downstream.
I read the multi-track approach as competent execution of a policy the administration cannot persuade Congress to enact through ordinary means. That is a description, not a judgment. Every administration since the 1970s has used some version of this approach when the legislative branch declined to legislate on a priority. The Obama administration ran simultaneous tracks on student-loan forgiveness through agency rulemaking, executive action, and litigation posture. The first Trump administration ran simultaneous tracks on immigration through executive orders, agency guidance, and DOJ litigation. The Biden administration ran simultaneous tracks on antitrust through FTC rulemaking, DOJ enforcement, and proposed legislation. The pattern is bipartisan. The pattern is the response of every modern executive branch to a modern Congress that has not legislated on a priority issue in any sustained way for the better part of two decades.
The thing that makes the AI version distinctive is the industry coordination. On student loans, the regulated population was borrowers, and borrowers do not lobby. On immigration, the regulated population was immigrants, and immigrants do not lobby. On antitrust, the regulated population was a small number of platforms, and the platforms did lobby, but the platforms did not have a unified position. On AI preemption, the regulated population is a small number of frontier-model developers who do lobby, have a unified position, and are paying the same four firms to work all three doors. That is what is new. It is not the multi-track strategy. It is the alignment of the lobbying coalition with the strategy at a level of precision that I have not seen on a comparable issue.
The right reading, if you are trying to position yourself for what happens next, is to track which door is closest to landing and prepare for the consequences specific to that door. Door one (a freestanding bill) is the slowest and least likely. The Senate Commerce briefings have produced no markup. Door two (NDAA Title XVI) is the most likely and the most consequential, because if Title XVI survives conference in any form that includes the preemption language, the resulting statute will be the strongest federal preemption authority on AI that exists, and it will exist in a defense bill that nobody wanted to read carefully. Door three (the EO route via xAI v. Weiser) is the fastest and the most legally precarious. The district court in Colorado will rule in summer 2026; the Tenth Circuit will rule by mid-2027; the Supreme Court will hear it, if it takes it, in the 2027 term. The preemption doctrine that emerges from door three will be, for better or worse, the operative doctrine on AI preemption for the rest of the decade.
Connecticut is the forcing function in all three. The AI Responsibility and Transparency Act becomes operative in January 2027. Between now and then, the federal preemption posture has to land on at least one door, or the operative federal-state line on AI regulation is going to be drawn case by case in district courts that nobody currently expects to be making AI policy. The administration knows this. The industry coalition knows this. The Connecticut attorney general knows this. The thing you should watch over the next six months is which of the three doors actually opens. The thing you should not do is assume that the slowest and most visible door (the freestanding bill) is the door that matters. The door that matters is the one nobody is watching, which is currently Title XVI. By the time the NDAA conference report drops in November, the question of federal AI preemption will likely have been answered in one of two ways: either Title XVI made it in, and federal preemption is statutory and durable; or Title XVI was stripped, and federal preemption is whatever the DOJ can persuade a district judge to grant it. There is no third outcome that runs through ordinary legislation. The framework has effectively retired that path by routing past it.
Position yourself accordingly. If you work for a state government, your window to legislate on AI before federal preemption arrives is approximately seven months. If you work for an AI developer, your compliance team should be modeling both outcomes (Title XVI passes; Title XVI fails and the EO is litigated). If you work in Congress, the question worth asking the staff director of House Armed Services this summer is not whether Title XVI will survive markup. It is whether the language will survive conference in a form that actually preempts state law. The answer to that question will determine the next decade of AI governance in the United States. The answer will be given in a conference report that will not be read carefully by anyone outside the four law firms that helped write it.
That is what is happening this week. Three doors. Same hallway. The doors are being worked simultaneously because the administration has correctly concluded that working them in sequence does not produce the outcome. I do not know if this is good. I know it is what is happening.
Related Briefings
Anna R. Dudley writes on national security, intelligence policy, and the places where hawks and reformers need to find each other. Bipartisan Translation is the weekly series for the conversation that is not happening on cable news. Subscribe at annardudley.substack.com.