Inside Israel’s AI targeting system: How data from a phone become a death sentence

BEIRUT — The buzz of the Israeli drone had never changed that day, and every time Ahmad Turmus looked up, it seemed to circle above him like an overly patient bird of prey.

So when the phone rang while visiting family on a Monday afternoon in February, Turmus wasn’t exactly surprised that the accented Arabic speaker was an Israeli officer.

What surprised him was this question.

“Ahmet, do you want to die with the people around you or alone?”

According to family members interviewed, Turmuş answered with one word before hanging up the phone: “Alone.”

The targeting of Turmush, which Israel acknowledged, shows how the Israeli army has mastered an intelligence war for which Hezbollah has no answers.

Since the spectacular pager attacks of September 2024, when Israel remotely detonated explosives stored in pagers carried by Hezbollah members, infantrymen, support personnel, field commanders, chiefs of staff and even a respected general secretary have been killed by an AI-powered targeting system.

Combining data from smartphones, security and traffic cameras, Wi-Fi signals, drones, government databases and social media, the system has given Israel a near-omniscient ability to track Hezbollah cadres’ every move.

Turmus, 62, served as a liaison between Hezbollah and residents of Talloosah, a small village less than three miles from the Israeli border that became a battlefield during Israel’s 2024 campaign against Hezbollah.

During the 15-month ceasefire that followed, he spent his time coordinating with repair personnel and civil defense teams to make the village operational, even as Israeli attacks continued in southern Lebanon.

His family described him as a former fighter of the militant Islamist group, but he had taken on an administrative role in his older years. Israel said he was working “on military and financial matters to improve Hezbollah’s terrorist infrastructure.”

Whatever his role, he too was now trapped in Israel’s chain of death; This was the culmination of an intelligence gathering process that had begun years before.

There is more than one way Turmus could fall into the army’s crosshairs; None of this is a smoking gun, per se, but they are all a potential grind for the algorithm that ultimately selected him to be killed that February day.

First, he lived in Talloosah, a predominantly Shiite village that supports Hezbollah; This means that the movements of Turmus and other residents are under constant surveillance by Israeli drones.

The drones’ cameras likely filmed and recorded his face, as well as the make and license plate number of his car and house, according to an artificial intelligence expert who worked with defense firms until he raised concerns about the use of such systems in Gaza.

The drones could use cell site simulators known as “stingrays” to pose as cell phone towers and trick the smartphone into making a connection, giving them access to not only Turmus’ data but also its movements in real time.

The artificial intelligence expert, who was granted anonymity to discuss his work, said that even if Turmus changed SIM cards, he was still tracked.

“This is a huge pipeline of data: phone metadata, location pings, SIM card swaps, app usage, social media behavior, sometimes even banking or facial recognition logins. So much is ‘collected’ from commercial platforms, mobile networks, joint intelligence agencies or spies in the field,” the AI ​​expert said.

Once collected, platforms like Palantir’s Maven standardize, tag and score all data, linking it to identities across devices and accounts. Palantir has spoken openly about its work with the Israeli military.

AI can then create a timeline of a person’s activity and map the network of their relationships.

Turmush may have been marked there, too: One of his sons was a Hezbollah fighter killed in early 2024; Another was injured in pager attacks.

Retired General Mounir Shehadeh, who served as the Lebanese government’s coordinator of the United Nations peacekeeping mission in Lebanon, said tracking Turmus would become easier thanks to Israel’s deep and accumulated intelligence infiltration into Lebanon.

Much of the country’s data infrastructure, including databases containing information on mobile phone subscribers or vehicle registrations, has been accessible to Israelis for two decades; He said they also infiltrated Hezbollah’s terrestrial network and signals units. Hezbollah’s involvement in the civil war in Syria from 2011 to 2024 further jeopardized the group’s security.

“These factors allowed Israel to establish a definitive bank of targets that included both field commanders and senior leadership figures,” Shehadeh said.

Artificial intelligence comes into play at this stage. It quickly examines terabytes of data to detect patterns and compare them to the movements of someone who is a known threat or has appeared near marked areas. It also analyzes deviations from a subject’s routine. All of this is used to create the so-called threat profile.

The result, according to an Israeli colonel interviewed in a February 2023 Israeli military article about artificial intelligence in warfare, is a system that can find targets quickly.

“The system performs this process in seconds, whereas in the past it would have taken hundreds of researchers several weeks to do this,” said the head of the Israeli army’s Artificial Intelligence Center, identified only as Colonel Yoav.

But one concern, according to the AI ​​expert, is that these systems use data, not logic, to determine whether someone is dangerous. And if this information is erroneous, then he will continue to repeat the same mistakes, but “faster and safer.”

“This creates the illusion of certainty, which is dangerous because it always translates correlation into action without context,” the expert said.

“This isn’t like a laboratory,” he added. “So how does the system know who is who? And when it flags someone, is it a human decision or is it just an algorithm flipping a switch?”

Vasji Badalic, a professor at the Institute of Criminology in Slovenia who wrote a 2023 research paper on the rise of metadata and big data-driven targeting processes, said such systems rely on tracking mundane, routine activities (such as who talks to whom, where and when they travel) to calculate the likelihood of someone becoming a combatant, potentially leading to false positives.

“Relatives or people involved in propaganda or finance are not warriors, but the machine recognizes them as such because they have similar communication patterns,” Badalic said.

“Where do they put the threshold separating combatants and civilians?”

The effort to use machine learning to identify targets or predict events in a war zone is not new. During the Iraq war under President George W. Bush, the U.S. military deleted phone metadata and processed it to look for what it deemed suspicious activity.

The National Security Agency also developed a behavioral profiling program called SKYNET to identify Al Qaeda couriers in Afghanistan.

By 2019, companies like Amazon and Microsoft had developed enough “computing” (computing power) to run math on more complex scenarios to improve predictions.

The US military in Afghanistan used these advances to develop Raven Sentry, an artificial intelligence trained on reports of insurgent attacks dating back to the ’80s, along with helpful information such as the amount of street lighting in various areas.

By the time the United States withdrew from Afghanistan in 2021, the model’s predictions of upcoming attack locations achieved a 70% success rate, roughly on par with human analysts, according to Col. Thomas W. Spahr, who has written about Raven Sentry at the U.S. Army War College.

Despite Israel’s success in Lebanon, there are signs that Hezbollah is adapting to being in Israel’s AI-enabled crosshairs.

During the current conflict, which began after the group struck Israel in response to the killing of Iran’s Supreme Leader Ayatollah Ali Khamenei and repeated violations of a 2024 ceasefire, Hezbollah has returned to its guerrilla warfare roots by adopting smaller units with a decentralized structure. It also relied on more secure, though less convenient, forms of communication, according to retired general Shehadeh.

It’s unclear what action triggered the algorithm to move Turmus from surveillance to the kill list. His liaison was a non-combatant member of Hezbollah, and family members said he didn’t even bother to switch phones. (“The Israelis already know me, so what does it matter?” he would say.)

On February 15, the day before he was killed, he turned off his smartphone and left it at home while he went to a town hall meeting in a nearby village the next day. The call from the Israelis came shortly after he went to his home in Talloosah and turned on his smartphone.

His face changed when he hung up the phone, family members told The Times. He told them that the Israelis were after him and that they should leave the house and let him die alone. They begged him to try to escape, to disguise himself so they could leave.

However, Turmus refused. He went to the door. “They know my face. There’s nothing we can do against it,” he said. Family members said his wife walked in as he was leaving, but she did not acknowledge him, so he did not try to stop her.

He got into his car, started it and drove off. Less than 30 seconds later, the scream of two missiles hitting Turmuş’s car was heard.