- History
- Revitalizing History
This is adapted from The Selection Problem: Aligning Pacific Allies for Algorithmic Warfare, a new publication by Hoover’s Allied Coordination Working Group and the Hoover History Lab. Download a copy of the essay here.
The United States has entered the era of algorithmic warfare. Algorithmic targeting is the AI-enabled fusion of drone video, radar, signals intelligence, and friendly-force data to identify targets and recommend strike solutions. It was a key reason the United States could launch precision strikes against more than thirteen thousand targets in Iran in thirty-eight days. It is also crucial to deterring and, if necessary, defeating a Chinese attack on Taiwan.
But the promise of superior targeting conceals a trap. Algorithmic targeting cannot eliminate the risk of civilian harms. Algorithms can fail, and they can be deceived. When they fail, and civilians are harmed, the consequences can be politically explosive. The allies therefore need to discuss and align on protocol for using algorithmic targeting before a crisis comes.
In any near-term cross-strait conflict, US munitions stockpiles would enter hostilities perilously low. Stocks were already drawn down to supply Ukraine. Then, the Iran war consumed thousands of critical munitions in its initial five weeks. The Pentagon is pouring tens of billions of dollars into cheap munitions, including drones, to supplement exquisite precision missiles. But scaling up production of missiles and drones will take years. Because of these factors, the United States and its allies will need to depend even more than they may have hoped on algorithmic targeting in any Pacific war.
But China has also designed a force structure that presents hard problems for algorithmic targeting. The People’s Liberation Army (PLA) pioneered the use of containerized missiles, deployed off commercial ships. It also deploys the People’s Armed Forces Maritime Militia (PAFMM), a fleet of dual-use, nominally civilian vessels that blurs the line between legitimate military targets and noncombatants.
If war erupts, these paramilitaries create five separate problems for allied targeting.
- First, militia boats look just like civilian fishing boats, so there’s no clean way to train algorithms to distinguish combatants from noncombatants.
- Second, engaging them wastes scarce munitions on low-value targets.
- Third, any attempt to engage risks killing civilian fishermen or fleeing refugees, which could fracture the coalition.
- Fourth, even a strike that is accurate and lawful will still look like an attack on civilians.
- Fifth, if allied forces hold back instead, the militia becomes a shield for the PLA Navy and helps it target US carrier strike groups and allied ships.
Allied models must therefore be trained to screen thousands of China’s maritime assets and identify which ones pose significant threats. This is a real dilemma with no clean solution. Miss a real threat, and allied ships are at risk. Mis-flag a fishing boat as a threat, and civilians die. Allies must accept trade-offs no matter how the targeting system is built.
If the United States mistakenly strikes civilians at any point in a conflict, the political fallout likely lands on allies first—and strains the coalition. The Philippines alone supplies a quarter of the world’s mariners. If Filipino seafarers are killed or injured in US strikes, the result could be a political crisis in Manila and backlash against the United States. Strikes against civilian targets could also disrupt merchant shipping, breaking supply chains and triggering a global financial panic. Innocent shipping will need guidance about how to operate safely, or it may avoid the region altogether.
Why allied coordination is the answer—recommendations
To use algorithmic targeting effectively, the allies need a shared protocol for how algorithmic targeting can be used. Allied coordination creates friction, and if a fast-moving conflict breaks out, it will be too late to deliberate about the issue from first principles. The answer is to deliberate in advance, establishing a shared taxonomy for validating and prioritizing high-value targets.
A shared target-value taxonomy. This taxonomy needs three axes:
Value. What does a target contribute to the campaign plan, and how much cost and risk should the coalition accept to neutralize it?
Confidence. What level of certainty is required before each class of target may be engaged?
Escalation sensitivity. Which targets are politically permissible to strike from, or within, each ally’s territory?
Seek allied input on civilian harm mitigation during development and fielding. Securing allies’ and partners’ buy-in to US concepts for algorithmic targeting will also require deliberate, ongoing information sharing. There should also be allied personnel and civilian harm mitigation and response (CHMR) staff embedded in the Pentagon offices that oversee unmanned and autonomous weapons: the current Direct Reporting Portfolio Manager for Unmanned Systems and Defense Autonomous Warfare Group, and the proposed Robotics and Autonomous Systems Combatant Command.
Joint control measures. The United States and its allies and partners should also predesign control measures for whenever autonomous capabilities are activated: exclusion zones at sea, announced to mariners in advance, and no-fly zones in the air above them, announced to pilots. By declaring these jointly, allies can promote awareness of risk among shippers and local populations and make unfortunate incidents less likely and less politically complex.
These measures admittedly come with trade-offs. Declaring an exclusion zone whenever a system activates telegraphs allied operating areas to the PLA. One partial solution is to declare exclusion zones broadly and persistently rather than reactively. Another is to use overlapping or decoy declarations. These approaches are morally, legally, and politically preferable to using unannounced autonomous engagement zones in waters thick with civilian traffic.
Flag states and marine insurers also need to be brought into the conversation in advance. Many of the merchant ships that may inadvertently carry the PLA’s containerized missiles are owned and operated by companies in neutral countries and fly flags of convenience from neutral states such as Panama, Liberia, and the Marshall Islands. A strike on a Panamanian-flagged vessel is a diplomatic incident with Panama. If allied exclusion-zone declarations are coordinated with the insurance market, compliant civilian traffic may exit risky areas of their own accord, making it more likely that ships that remain in the area are not actually civilian.
These are hard questions that admit no perfect answers. But one thing is clear: the algorithmic selection problem is ultimately an alliance problem. These conversations must happen now, in peacetime, because there will be no time for them in war.
Read the full essay here.
