The first AI war: US and Israel test autonomous tech

Thursday, 12 March 2026

Larisa Brown is Defence Editor of The Times, London

In the first 24 hours of Operation Epic Fury, the US military struck more than 1000 targets in Iran with the help of artificial intelligence.

Given the rate of 42 suggested targets per hour, experts have asked whether machines are now in charge on the battlefield because the human brain cannot keep up.

They have raised the possibility that artificial intelligence may have misidentified the primary school in Minab on the first day of the war. Growing evidence suggests that the US fired what appears to have been Tomahawk cruise missiles at the site, killing 110 children and dozens of others.

Artificial intelligence systems can crunch huge amounts of data from sources including satellite images, surveillance aircraft data and human testimony to select or dismiss targets.

Historical satellite images have shown that the Shajareh Tayyebeh primary school was once part of an Islamic Revolutionary Guard Corps complex, home to a naval barracks and support buildings. However, for at least the past nine years it has been walled off from the complex. Colourful murals on the walls and a small playing field have been visible in some satellite imagery.

Questions have been raised about whether an AI system relied on the old images for targeting purposes.

Noah Sylvia, a research analyst for the Royal United Services Institute, said: “If the school bombing was in error, was it human error or the speed of automation of the system? Was it based on old data? Was it a machine that did it automatically? The number of strikes we are seeing lends some credence to the idea that targets are largely autonomously created.”

Dr Craig Jones, senior lecturer in political geography at Newcastle University, said: “At this point we can't rule out that AI may have … failed to identify the school as a school and instead identified it as a military target.”

He said the key issue was the extent to which AI was used in the collection and analysis of intelligence data to present the school as a military target, pointing out that any human decision to strike was based on data and analysis that AI had helped to construct. “Whatever ends up being the case - and the facts will remain obfuscated for some time - the strike was a catastrophic intelligence failure, whether AI-driven or human-driven with some AI component,” he said.

Asked by The Times whether artificial intelligence systems had provided outdated information on the school, the Pentagon said an investigation was ongoing. President Trump suggested, without evidence, that Iran or “somebody else” might have carried out the strike, but analysis of the site has suggested that American weapons were used.

The US and Israel have a multitude of AI systems at their disposal in the war on Iran. The Pentagon has been developing Project Maven with the help of the CIA-backed data mining company Palantir since 2018 and the software is embedded in all US combatant commands.

In the Iran war, the US has been using AI and machine learning to help with military targeting and surveillance. Experts have likened it to a military version of Uber.

Embedded into the Maven Smart System is Anthropic’s Claude chatbot. It has been used in Operation Epic Fury for intelligence analysis. In theory, it could pull together open-source data on what is going on inside the Iranian government.

It has strict usage policies and is not used for targeting. Anthropic believes humans should be responsible for identifying targets and making decisions on strikes. The Trump administration has now blacklisted the company, shocking officials in Washington.

Although Palantir’s software has already been used in the Ukraine war, Sylvia said the scale of operations in the Middle East was “far more dramatic”.

He said the data sources would be different and the US and Israel were using different systems than those used in Ukraine - long-range strike weapons, rather than artillery and first-person view drones. “It is fundamentally different because of the geographical scope of operations and the orders of magnitude,” Sylvia said.

“Eastern Ukraine is tiny compared with Iran. The ability to do strikes over a large proportion of Iran requires a lot more data and far more significant systems to punch through the data.”

With that came risks, he said, adding that commercial companies were churning out more advanced models and the service personnel operating them needed to be up to speed. He raised concerns about a “rapid deployment of technology and a lack of technical expertise among personnel”.

“When you have a feeling of urgency, humans can only help so much. As war increases in tempo of operations, increasingly on an unprecedented scale, we are going to see increases over time in the level of autonomy in some of the systems. This raises the question, how comfortable people are about this?”

Elke Schwarz, professor of political theory at Queen Mary University of London, said systems that suggested thousands of targets a day produced an “automation bias”, meaning they became the authority. “The human has limited cognitive abilities to override this capacity,” she said.

There was also an “action bias”, she said, referring to the rapid machine decision creating an urgency to act. “Humans need time for deliberation, and this deliberation is crucial for legal and ethical evaluation.”

She said that 42 suggested targets per hour was “a lot, and it becomes difficult to evaluate the validity of all of these targets when speed and scale is prioritised”.

Experts in the field fear what the future of conflict may hold and warn of “mission creep”. Schwarz said future AI systems might identify not only possible targets at speed and scale, but also “suspicious” behaviour or movements, which could then prompt pre-emptive strikes.

She said: “We must assume that AI will come to play an ever-growing role in the decision to use force - the decision to initiate conflict - and that is terrifying.”

- The Times