§ 1Introduction

The War Powers Resolution hasn't worked. It came into being in the unusual circumstance of the end of the Vietnam war, with a reviled President and a strong Congress. That political power dynamic quickly flipped, and it has stayed that way.

From President Ford forward, presidents of both parties have treated the WPR as advisory rather than binding: provisions in the U.S. Code that may shape the text of a letter to Congress but don't dictate where U.S. military forces go or when they come back. The text speaks in terms of obligation: that the President shall confer with Congress before deploying our military might; that he shall transmit a report soon after the deployment begins; that he shall withdraw our military after two months absent specific authorization to keep them in the field.

In fifty-two years, none of those three commands has had an effect in the field.

What did the Congress that passed the WPR miss? That a future Congress (as it happened, all Congresses) would be unwilling to act because it lacked the political power to do so.

When the WPR came into being, Congress had, as measured by public support, more power than the President. Probably more important, the public strongly disapproved of Nixon. It's probably not too much to say the public loathed him. It's thus understandable that the Congress that overrode Nixon's veto of the WPR would think that future Presidents would hew to its terms when it said that action was mandatory and that, if he didn't, Congress would bring him to heel. The drafters drafted in their moment, and they were right about it.

They were wrong about every moment after. The asymmetry has run the other way for every military engagement since the WPR's passage: presidents lead Congress on approval by margins that range from a handful of points to fifty. The WPR gives Congress the whip hand; the polling determines whether Congress can exercise it.

This Article narrates sixteen post-Vietnam deployments tested against the WPR's three key obligations. I assess those requirements through the lens of relative political power before the President had decided to commit our military to the field. The answers powerfully associate with a variable found nowhere in the WPR: the gap, in the President's favor, between the President's approval and Congress's at the moment of decision. Where the gap runs wide and no other variable intervenes, the statute sleeps; where it narrows or a second variable supplies what the gap withholds, the statute stirs but rarely bites.

The WPR was drafted on the premise that its textual commands would discipline the politics. They don't; they haven't; and they never have.

What makes it urgent.

That dismal record would warrant attention on its own. What makes the case urgent is the technology now arriving. Autonomous AI, once a literary conceit, the stuff of science fiction, has become a procurement reality. Ukrainian Saker Scout drones operate in modes their manufacturers describe openly, identifying and prosecuting targets without further human direction. AI-assisted target generation in Gaza has produced kill lists at a velocity that turns human review into ceremonial assent.

Closer to home, Anthropic's litigation against the Department of War over permissible uses of its frontier models is the first visible iteration of what will be a recurring conflict between commercial AI developers and the federal government. AI's arrival has materially and irrevocably changed the WPR universe.

So § 5 builds a hypothetical, lightly extrapolated from current facts on the ground, in which an integrated autonomous strike system runs an extended campaign in Yemen against Iranian proxies engaged through Houthi targets, on the legal authority of two AUMFs not much younger than some of the analysts writing OLC memos about them. The hypothetical lacks, insofar as the public record discloses, only two ingredients: the integrated platform itself and the executive decision to use it. Both sit well inside the planning horizon of every senior U.S. Central Command officer.

When the hypothetical materializes, which it inevitably will, the WPR's existing failure modes won't merely recur. Autonomous AI will compound them.

What this Article proposes.

When should Congress amend the WPR? Now. To help it, § 6 proposes five sets of repairs, woven together rather than offered as a menu.

First. Objective triggers replace presidential self-triggering with operational facts the executive can't redefine: a comprehensive definition of hostilities and a sixty-day clock that runs from those facts, not the President's decision to file a report about them. A self-executing consequence rides with the trigger: new § 1544(e) lapses appropriations supporting the operation by operation of law when the sixty-day window closes without authorization, with the Anti-Deficiency Act's penalties enforcing the lapse.

Second. Specific prior-consultation obligations replace the current statute's in every possible instance gesture with a standing committee, a forty-eight-hour pre-introduction briefing requirement, an autonomous-systems variant for ROE briefings, and a Chadha-compliant joint resolution to replace the concurrent-resolution mechanism the Court invalidated in 1983.

Third. AUMF hygiene: sunsets, geographic limits, force ceilings, enumerated targets, and capability limits, paired with a sunset of the inherited 2001 and 2002 AUMFs.

Fourth. Institutional repairs that interrupt the executive's interpretive ratchet, through a publication requirement for OLC and comparable executive-branch opinions, paired with a non-citation rule for opinions kept past the publication deadline, and the creation of an Office of Congressional Legal Counsel on War Powers staffed to produce competing interpretive positions on the legislative side.

Fifth. A service-member track. New § 1544a gives a service member ordered into an unlawful deployment a private cause of action for declaratory and injunctive relief, expedited proceedings, and a defense to a court-martial proceeding for refusal where the deployment is unauthorized.

Appendix A implements all of this as a redline against the existing text, in a form Congress could enact next session.

What the proposals can't do.

Most significant, they don't manufacture political will, and no statute can. They don't close the visibility deficit autonomous AI will widen, and no statute can. And of course they don't repeal Article III.

What they do is raise the price of executive evasion by closing the legal escape routes the executive has used, for fifty years, to slip the WPR's commands without political cost. A President willing to pay the new, higher price can still pay it. The point is to make noncompliance visible, and to steepen its political cost to the point that brings the executive to heel.

§ 2The WPR in theory

The WPR's three awkwardly matched parents were a mother (the constitution) that's ambiguous about distributing war powers, a father (the executive) that, over the course of nearly two hundred years, converted that ambiguity into accreted power unilaterally to deploy American military power abroad, and a stepfather (the judiciary) that concluded it wasn't going to disturb whatever arrangement the original parents had come to.

The Constitution distributes war powers across two branches and then goes quiet about sequencing and boundaries. The historical record suggests the ambiguity was the result of a compromise, an effort at the Philadelphia Convention to preserve presidential power to repel sudden attacks while vesting the initiation of sustained conflict in Congress through the declare-war clause. The power to repel sudden attacks turned out to be elastic. Jefferson sought authorization after the fact for the Barbary actions. Polk framed his march into disputed Mexican territory as war started by Mexico. Lincoln blockaded Confederate ports before Congress convened. Korea was the leap: Truman ordered more than five and a half million Americans into a conflict that killed thirty-six thousand of them, cited a Security Council resolution, and called it a police action.

By the late 1960s, with American forces in Vietnam fighting under no declaration of war, the Court was repeatedly invited to resolve the constitutionality of the conflict and repeatedly declined, dismissing on political-question and standing grounds. The retreat wasn't uniform in doctrine, but it was uniform in result: in the dozen years between Mitchell v. United States and the war's end, no plaintiff who challenged the war's legality obtained relief that survived appellate review beyond a few weeks.

This picture was filled out in Youngstown. Justice Jackson's concurrence sorted presidential power into three categories, and he conceded what no opinion before or since has conceded so candidly: the constitutional text can't tidily resolve disputes about war powers. Between the poles of express authorization and express prohibition sits the second category, the famous zone of twilight in which Congress has largely held its peace for most of our history. There the distribution of authority is uncertain, and outcomes turn not on abstract doctrine but on, in Jackson's phrase, the imperatives of events and contemporary imponderables.

That's perhaps the most candid acknowledgment in the United States Reports that constitutional law in this domain runs out before the hard cases begin. For a legislature trying to replace twilight with text, that's a warning: the ambiguity is a feature of separated powers a statute can mitigate but not eliminate.

Those were the circumstances the WPR walked into. The conference committee produced a hybrid that satisfied neither the Senate's enumeration impulse nor the House's outer-limit instinct. Senator Eagleton, instrumental in drafting the Senate version, voted against final passage, arguing that by authorizing sixty days of presidential war-making without any congressional action, the WPR had expanded executive authority rather than constrained it. Nixon vetoed the WPR on October 24, 1973; the Senate overrode 75-18, the House 284-135, and the WPR became law on November 7.

Here's the irony. Eagleton's critique and Nixon's veto message identified the same defect from opposite ends: Nixon said the statute wrongly constrained presidential authority; Eagleton said it wrongly expanded it. In the end, both were wrong: as we shall see, in practice, the WPR has had almost no substantive effect on presidential authority.

§ 4Four findings, organized into one variable

The sixteen cases narrated in § 3 (and catalogued in detail at the cases dashboard) pattern to four findings. They organize around one variable the statute doesn't mention.

§ 5The next WPR frontier: autonomous AI

The sixteen cases exposed how the WPR could be bent, from its very beginning. A new frontier will go further and simply break the WPR: the use of autonomous artificial intelligence to select and engage targets in foreign combat.

The omission isn't the drafters' fault. The technology didn't exist; the category didn't exist; the vocabulary to describe the category didn't exist. Autonomous AI was the stuff of science fiction, the realm of Isaac Asimov or Robert Heinlein. A statute written when the state of the art in battlefield computing was the HP-35 calculator can't be expected to have contemplated an armed drone that picks its own targets.

The issue isn't distant. It's here now, on three fronts that don't depend on each other. Ukraine is the world's first sustained laboratory for attritable autonomous strike. Israel's operations in Gaza supply the first sustained public record of large-scale AI-assisted targeting at the population level. The United States is the third front, and the most legally consequential for purposes of this Article: Anthropic's dispute with the federal government, OpenAI's immediate substitution, and the Pentagon's standing position that AI systems should be available for any lawful purpose with the boundary of lawful set by the executive's own legal interpretations. Three fronts, three pressures, one diagnosis. A 1973 statute drafted for human pilots in human cockpits doesn't reach this operational environment.

A hypothetical from the near future.

It's January 2027. Houthi forces in Yemen have escalated attacks on Red Sea shipping for the third consecutive year, and Iranian materiel support has produced a new generation of anti-ship ballistic missiles. The administration, citing the 2001 AUMF and Article II, deploys to U.S. Central Command an integrated targeting platform I'll call Project Sentinel: a constellation of long-loiter ISR drones, attritable strike drones, and a centralized machine-learning targeting suite running on a classified DoD network. Sentinel ingests SIGINT, IMINT, OSINT, and MASINT in real time, classifies threat signatures against a continuously trained model, and, once a threshold confidence score is met, prosecutes targets within a defined Yemeni geography.

The rules of engagement authorize Sentinel to engage missile-launch preparation against U.S. or allied vessels, command-and-control nodes within a defined Yemeni geography, and, upon attack on U.S. forces, escalate to a tier authorizing GBU-57 Massive Ordnance Penetrators against a pre-identified bunker network. The escalation tier is automatic. No human approves the second wave. The Pentagon's General Counsel has concluded, in a classified opinion, that the standing ROE satisfy the appropriate levels of human judgment over the use of force standard in DoD Directive 3000.09. So has the Office of Legal Counsel.

Run the hypothetical through the existing WPR.

Six provisions, six failures. Plus a seventh the sixteen cases couldn't have anticipated: the system did the killing.

Not just Soleimani, with a drone.

One might reasonably ask: isn't this Soleimani all over again? The January 2020 strike put no U.S. personnel in theater, killed its target through a machine, and triggered Iranian retaliation that produced American casualties at al-Asad. The hypothetical rearranges the same elements.

It isn't, along the main axis the WPR cares about: the decision locus. Soleimani involved a human targeting decision at every step. The President ordered the strike; a targeting cell confirmed the target; an operator released the weapon. The MQ-9 is a remotely piloted aircraft, not an autonomous system. Every link in the chain had an identifiable human making an identifiable decision, which is why the Soleimani legal justification reads as a conventional command-authority analysis.

Sentinel inverts that. The President authorized deployment months earlier; the Pentagon approved the ROE in the summer; the algorithm selected the specific target in October on its own confidence threshold. No human officer decided to strike this launch site at this moment. The decision the WPR's triggers presuppose, the human act of introducing forces, the human judgment about circumstances necessitating the introduction, isn't present at the moment of engagement.

A skeptic will press that the decision is still the President's, just made earlier. That's right, and it's the problem. Every operative WPR trigger assumes the act of introducing forces is roughly contemporaneous with the act of using them; when the two acts separate by months, and the use-of-force decisions are made by a system rather than a human, the statute's temporal structure breaks in ways Soleimani didn't test.

A more sophisticated skeptic suggests a press-to-confirm step. Before each engagement, an officer reviews and presses a button. The decision locus is human; the attribution problem dissolves; the legal architecture is restored.

Not quite. The human-confirmation step is meaningful only if the human can meaningfully decline. When the officer reviews a target the algorithm has selected, at the velocity the algorithm operates, against a confidence score the officer can't independently verify, on a target list the officer didn't compile, the confirmation function approaches what behavioral researchers call automation bias. The Patriot fratricide incidents during Operation Iraqi Freedom are the textbook case: nominal authority to override, operational tempo and prior reliability producing confirmation rather than supervision, friendly aircraft engaged.

The officer pressing the button isn't exercising the human judgment the WPR's decisional architecture presupposes. The officer is ratifying a decision the system has already made.

A revised statute that treats human confirmation as curing the attribution vacuum will be satisfied by confirmation procedures that don't, in fact, restore meaningful human decision-making. The drafting question isn't whether the system has a human in the loop; it's whether the human's role is capable of being more than confirmation.

The revised WPR's new § 1547(g) addresses this by treating systems where the human role is limited to confirming algorithmic selection at operational tempo as functionally autonomous for WPR purposes. Sentinel with a confirmation step is harder than Sentinel without one, because the confirmation step gives the executive a textual handhold the unmodified Sentinel lacks. The repaired statute denies the handhold.

Four problems, organized.

The lessons cluster around four problems autonomous AI poses for the WPR, plus a fifth set of compoundings that cut across them.

The definitional problem. The WPR uses hostilities eleven times but never defines it. Autonomous AI completes the executive's fifty-year trajectory of definitional erosion. Every Koh four-factor prong (limited objectives, limited exposure, limited escalation risk, limited means) is satisfiable by design when the platform is autonomous. Whatever the executive had to argue for in Libya 2011 was a stretch; the same move for Sentinel is a layup.

The attribution problem. The WPR's operative provisions all assume a human decision-maker at the moment of force. Sentinel breaks the assumption: every actor in the chain has a plausible not me defense if the engagement produces an unlawful outcome. The repair is a clock that runs from deployment with engagement-authorizing ROE, not from individual engagements. The individual engagements are mechanical consequences of the deployment; the deployment is the moment of human decision, and that's where the WPR's temporal commands should fire.

The velocity problem. The 1973 Congress drafted against an operational environment in which escalation decisions took hours or days. Sentinel's escalation tier fires in minutes. By the time the Gang of Eight could be assembled, the bunker-busters have been delivered and the third-day strikes are being targeted. Whatever Congress wants to contribute must be contributed ex ante, at the deployment and ROE stage. Post-deployment intervention will arrive too late, not because Congress is slow but because the operational tempo has moved past the timescale on which the WPR's machinery operates.

The salience-absence problem. The approval gap is overcome, when it's overcome, through public salience. Lebanon's Marines died visibly. Soleimani's strike produced a direct Iranian missile retaliation. Even Kosovo, no American combat deaths, generated salience through duration. Sentinel zeroes the variable. No troops in theater, no pilots at risk, no footprint for the press to photograph. Every mechanism by which the approval gap has historically been overcome depends on public salience, and autonomous operations are designed to produce none.

The salience-absence problem isn't novel to autonomous AI; the Pakistani drone program ran for more than a decade with sharply limited public salience. What autonomous AI does is institutionalize the failure mode the drone program revealed and remove the residual salience-generating mechanisms (the operator who might leak, the contractor who might whistleblow). If text alone can't fix the problem for autonomous AI, it couldn't have fixed it for the drone program either. That doesn't mean we shouldn't fix the deficient statute. It means there's a limit to what text can do.

Technological compounding. Four additional features compound the four above. The evasions no longer take turns; they queue up at once. Sentinel deploys the full executive-side portfolio simultaneously: the consistent with formula, operational brevity, the no-hostilities theory supercharged by zero American exposure, the associated-forces stretch, the AUMF compounding, the appropriations-as-authorization theory, mission creep under standing ROE, the fait accompli of a pre-positioned platform, and a clock that never starts. The techniques that took half a century to assemble deploy in a single operation.

§ 7Conclusion: the WPR's twilight, and its sunrise

The War Powers Resolution failed. That's the verdict of our jury of sixteen, and it isn't close. Presidents don't consult, don't file the reports that start the clock, and don't withdraw when the clock runs, because the clock almost never runs. The Resolution's three mandatory shalls have degraded into optional courtesies the executive extends when convenient and ignores when not.

The verdict's second part matters more than the first. The drafters wrote into a political vacuum they couldn't fill. The approval gap, which the text doesn't mention, correlates with outcomes with uncomfortable consistency.

A jury's verdict doesn't end a litigation, and our verdict-of-sixteen doesn't end this analysis. Two reasons.

Textual repairs raise the price of political will's absence. Define hostilities comprehensively, and the executive can't redefine the operative term to exclude seventy-eight days of bombing or a campaign run by software. Replace open-ended consultation with a standing committee and a forty-eight-hour advance briefing, and notification hours before the first bomb stops counting as compliance. Run the clock from operational facts rather than from a report the President controls, and the fifty-year fiction that not-filing is not-triggering loses its predicate. Replace the concurrent-resolution mechanism with a Chadha-compliant joint resolution; eliminate consistent with; add sunsets, geographic limits, force ceilings, and capability limits to AUMFs; publish OLC opinions within ninety days; create a congressional analog to OLC. The interpretive ratchet meets institutional resistance on the legislative side for the first time.

The future arrives now. Autonomous AI breaks every assumption on which the 1973 statute was built: the exposure assumption, because autonomous systems face none; the attribution assumption, because the decision chain runs through an algorithm; the salience assumption, because an operation no American sees produces no public reaction for Congress either to channel or to resist. Failure modes that took fifty years to assemble compound when the combatant is software.

None of this creates political will. The approval gap will keep determining whether Congress enforces what the statute commands, and it widens when operations are conducted by systems the public can't see. No statute can make military action salient when its design makes it invisible.

What a statute can do is deny the executive the legal camouflage that lets evasion happen without political cost. A president who evades a comprehensively defined hostilities is evading, not interpreting. A president whose clock runs from operational facts can't stop it by withholding paperwork. A president who deploys an autonomous lethal system without AUMF capability authorization is acting without authority, not exploiting an ambiguity. Modest gains, but gains the current statute doesn't deliver.

The 1973 Congress assumed that if Americans were dying abroad, the political system would eventually produce a reckoning. Autonomous systems disable the assumption. The branch constitutionally assigned the war power is about to face a class of military operation for which its political machinery has no natural response.

The statute we have is utterly inadequate to that future. The statute proposed here is less inadequate, and less inadequate is the most any statute of this kind can credibly promise.

Whether the branch with the power to enact it is capable of the act is a different question. The answer may not be comfortable. But the draft statute helps to clarify what's at stake.

© Vernon M. Winters · New York City · Spring 2026