A video has emerged of a senior official from Israel’s cyber intelligence agency, Unit 8200, speaking last year about using machine learning “magic dust” to help identify Hamas targets in Gaza.
The images raise questions about the accuracy of a recent statement regarding the use of artificial intelligence (AI) by the Israel Defense Forces (IDF), which said that it “does not use an artificial intelligence system that identifies terrorist agents or try to predict if a person is a terrorist.”
However, in the video, the head of data science and artificial intelligence for Unit 8200, named only as “Colonel Yoav,” said he would reveal an “example of one of the tools we use” before describing how the division of intelligence used machine learning techniques. in the May 2021 Israeli offensive in Gaza to “find new terrorists.”
“Let’s say we have some terrorists who form a group and we only know some of them,” he said. “By practicing our data science magic dust, we can find the rest.”
The descriptions in the video of the technology used by Unit 8200 have similarities to recent testimonies from six IDF members about their use of an artificial intelligence tool called “Lavender” during their offensive against Hamas. They said the AI-generated database had been used to assist intelligence officers involved in the Gaza bombing campaign, helping to identify tens of thousands of potential human targets.
In its rebuttal, the IDF said some of the accounts were “baseless.” However, the accounts are consistent with Yoav’s comments during a conference on AI at Tel Aviv University in February last year. The video, in which Yoav can be heard speaking but not seen, was hosted on the university’s YouTube channel and until recently had fewer than 100 views.
When he came on stage wearing a military uniform, the audience was ordered not to take any photos of Yoav or record his performance. “It’s good because it’s a bad day for me,” she joked.
In the 10-minute presentation, titled “Digital Transformation and Artificial Intelligence in Intelligence,” the colonel offered a rare glimpse into how secret military and intelligence agencies are using opaque AI systems.
When AI is used to predict whether someone is a terrorist, he explained, Unit 8200 takes the information it has about people it believes are members of terrorist groups and aims to “find the rest of the group.”
Referring to a specific example, the official said that in the May 2021 IDF military operation in Gaza, his department applied this principle to “find missile commanders of Hamas squads and anti-tank missile terrorists in Gaza to operate against they”.
He explained that using a form of machine learning, known as “unlabeled positive learning,” “we take the original subgroup, calculate its close circles, then calculate the relevant features, and finally, classify the results and determine the threshold. .”
The colonel stated that the feedback from intelligence agents is used “to enrich and improve our algorithm” and stressed that “flesh and blood people” make decisions. “From an ethical standpoint, we put a lot of emphasis on this,” he said, adding that “these tools are meant to help break down your barriers.”
According to Yoav, Unit 8200 was able to break “the human barrier” during the May 2021 offensive when it managed to produce more than 200 new targets. “There were times when this amount took us almost a year,” he said.
Contacted for comment on the video, the IDF said the colonel’s participation in the conference had been approved by the military. However, a spokesperson denied that his comments conflict with the IDF’s recent denials about the use of AI. In a subtle change in wording that was not used in its original statement, the IDF told The Guardian that its artificial intelligence systems do not “select targets” for the attack.
“The IDF never denied the existence of a database of agents of terrorist organizations, which collates existing information on such agents,” he said. “At the same time, the IDF fully maintains its claim that it does not use artificial intelligence systems that choose targets for attack and that the database in question is not a list of agents eligible to attack. There is no contradiction.”
In their testimony, the six intelligence officials who spoke last week said the Lavender tool had been used to help identify potential targets at an unprecedented scale and pace. The IDF was not accused of using systems that automatically select targets for attack.
Four of the sources said that, early in the war, Lavender listed up to 37,000 men in the enclave who had been linked by the artificial intelligence system to Hamas or Palestinian Islamic Jihad. None of the sources denied that humans were involved in the attack authorization process, although some confessed to minimal human supervision.
“At this stage, I would spend 20 seconds on each target and do dozens of them each day,” one intelligence officer said. “I had no added value as a human being, other than being a seal of approval. He saved a lot of time.”
His stories were published by the Israeli-Palestinian publication magazine +972its Hebrew media Local Call and The Guardian.
Colonel Yoav’s description of the partnership between AI and intelligence personnel echoes a model for targeting operations envisioned by his commander, Unit 8200 chief Yossi Sariel, who, as revealed by The Guardian, left his identity exposed online when he secretly wrote a book published in 2021, The Human. Machine equipment.
At one point, the colonel explained that the artificial intelligence and data science department he heads at Unit 8200 is also known internally as “the human-machine integration center.”
Speaking eight months before the IDF began operations in Gaza following the Hamas-led October 7 attacks, the colonel spoke optimistically about how the IDF is moving “from the era of postcards to the digital age” where “ suddenly you can react during battle with applied data.” -Solutions driven by science.”
Looking ahead, he added: “I’m curious to know what the next operation will be like, digitally speaking.”