What Can AI Change in a Conflict With Iran—and What Can’t It Decide?
Sources checked September 22, 2026 • National Security
A faster recommendation can shorten a decision cycle. It cannot tell a commander whether the information is still true—or tell political leaders whether the action will bring the conflict closer to an end.
That is the question the conflict with Iran puts to AI: what does speed change, and which decisions still belong to people?
That question should not depend on a running tally of strikes or the assumption that the conflict has stayed in one phase. Reporting on September 22 described renewed diplomatic efforts and competing positions on maritime access and military pressure. These are developments, not a settled outcome. The durable distinction is between tactical tempo, operational endurance, and political resolution. [1]
Speed is a capability—not an end state
The public case for AI-enabled speed is not solely a vendor narrative. In March, CENTCOM commander Adm. Brad Cooper publicly described the military’s use of AI to accelerate information processing, while emphasizing that humans retained final strike decisions. Those statements are evidence of the command’s stated use and assessment of the technology. They are not an independently reproducible measurement of its effect on targeting accuracy, civilian protection, or strategic success. [2]
That distinction matters. A system that helps analysts find relevant information sooner can have operational value. It may help identify a fleeting threat, reduce repetitive work, or make additional information available before a decision. But a faster recommendation is not automatically a better recommendation. Its value depends on the quality and freshness of the inputs, the questions the system is designed to answer, and the scrutiny applied before action.
Nor should the absence of a publicly available performance audit be treated as proof that AI contributed nothing. Public evidence supports a narrower conclusion: claims about speed and adoption do not, by themselves, establish a causal link to a durable political outcome.
Verification must keep pace with recommendation
The Minab school strike illustrates why this discussion must remain anchored to dated evidence. A March account cannot be treated as the last word on the investigation.
On September 21, Senators Jack Reed, Mark Warner, and Chris Coons sought inspector-general investigations into reported AI-related failures. Their letter cited public reports involving outdated geospatial information in the Minab strike and alleged AI-generated false intelligence in a separate operation. The letter establishes a formal request for investigation; it does not itself establish the technical cause of either incident. [6]
Separately, a UN-mandated fact-finding mission reported reasonable grounds to believe that U.S. strikes, including the Minab school attack, constituted war crimes. That is an investigative finding, not a final judicial judgment, and it is not proof that a particular AI model or vendor caused the attack. [7]
These developments strengthen the case for scrutiny while making careful attribution essential. It is not responsible to say simply that AI caused the strike. It is equally inadequate to dismiss the issue by relying on earlier reporting that had not established AI involvement. Investigators need to distinguish data failures, software behavior, institutional assumptions, review procedures, and command decisions.
The practical questions are concrete. How old was the source information? What changed at the location? Which assumptions were carried forward? What contradictory information was available? Who could challenge the recommendation, and what happened when they did?
Human accountability needs a usable process
Keeping a human involved is necessary, but the phrase alone says little about the quality of oversight. A reviewer needs enough time, context, training, and authority to recognize uncertainty and reject an unsupported recommendation.
The Defense Department’s January 2023 update to Directive 3000.09 required appropriate human judgment over the use of force and set expectations for care and compliance in operating autonomous and semi-autonomous weapon systems. This is a policy benchmark with a defined scope. It is not evidence that any particular strike complied with the policy, and it should not be used as a blanket description of every intelligence-support tool. [8]
The management challenge is to preserve provenance, contradictory reporting, confidence limits, and a record of human review as information moves faster. Performance measures should reward validated findings and justified decisions, not just the number of recommendations processed.
The physical constraints remain
Software does not eliminate the constraints of supply, maintenance, and infrastructure. A March 24 analysis published by RUSI estimated that coalition forces expended more than 11,000 munitions during the first 16 days of the conflict. That is a dated analytical estimate covering a specific period—not an official current expenditure total. [3]
A July 27 CSIS assessment likewise estimated substantial pressure on U.S. interceptor inventories. AP’s account of that assessment reported an estimated 759–827 Patriot interceptors remaining at that point, compared with a prewar estimate of 2,330. Those numbers should not be presented as today’s confirmed inventory. The relevant strategic issue is the relationship between expenditure, replenishment, and the demands of other contingencies. [4]
Infrastructure offers a related lesson. Air & Space Forces Magazine reported on June 5 that Iranian strikes had rendered the Combined Air Operations Center at Al Udeid inoperable. It also reported that the United States had directed the campaign from Shaw Air Force Base from the outset, anticipating the threat. This was evidence of both physical vulnerability and the value of advance continuity planning—not evidence that the air campaign stopped or that repairs had restored the facility. [5]
AI may help people allocate resources and evaluate options. It cannot make a damaged facility usable merely by producing a faster assessment, and it cannot erase the time needed to replenish an inventory. Operational advantage depends on how information systems work with those physical realities.
Strategic outcomes require political choices
The effects of conflict extend beyond the targeting cycle. In an April assessment, the World Food Programme projected that continued conflict and its economic effects could increase acute hunger. Its estimate of 45 million additional people was conditional and model-based, not a measured count that can be carried forward unchanged. The broader lesson is that food access, shipping, prices, and humanitarian capacity belong in the strategic assessment. [9]
AI can support analysis of these pressures. It cannot decide which political objectives are legitimate, which risks leaders should accept, or what terms will make an agreement durable. Those choices require human judgment and accountability.
The strongest case for AI is therefore neither technological triumphalism nor wholesale rejection. It is disciplined use: exploit genuine gains in information processing, test them against operational outcomes, preserve meaningful review, and invest in the logistics and infrastructure that make action sustainable.
The decision is not automation or strategy. It is whether automation serves a strategy whose objectives, constraints, and measures of success have been made explicit.
AI can compress the time between information and action. It cannot determine what the political end state should be.
Sources
1. Reuters, September 22, 2026: Iran’s proposed conditions for reopening Hormuz and renewed diplomacy. Statements attributed to an Iranian official; not confirmation of an agreement.
2. DefenseScoop, March 11, 2026: CENTCOM commander describes AI use during Operation Epic Fury; AP, March 11, 2026: Cooper’s statement on human strike decisions.
3. RUSI, March 24, 2026: Over 11,000 munitions in 16 days.
4. CSIS, July 27, 2026: Renewed Iran War Would Test Diminished Interceptor Inventories; AP, July 28, 2026: reporting of the CSIS estimates.
5. Air & Space Forces Magazine, June 5, 2026: damage at Al Udeid and operations at Shaw.
7. AP, September 17, 2026: UN-mandated mission’s findings.
8. Department of Defense, January 25, 2023: update to Directive 3000.09.
9. World Food Programme, April 2, 2026: conditional food-insecurity projection.
Research cutoff: September 22, 2026. This article is analysis, not a live conflict tracker. Recheck newly released investigation findings and conflict developments before publication on a later date.