Unit 8200 collected all the spoken Arabic text they had and created a database of about a hundred billion words from groups that were hostile to Israel. Israelis also trained the model to understand specific terms used by anti-Israel groups. This massive collection of training data included large volumes of communications between Palestinians, which was just what the model needed to succeed.
Earlier machine learning models used by Unit 8200 made wide-scale surveillance of Palestinians effective as a form of control, particularly in the West Bank where they said it has contributed to a greater number of arrests. The model enabled Israeli intelligence specialists to automatically analyze intercepted phone conversations and identify Palestinians planning to attack soldiers or Israelis living in illegal settlements. When Israeli soldiers entered West Bank communities, the AI system detected people using words indicating hostile activity like throwing rocks or using firearms against soldiers.
When used to select targets for airstrikes, the AI sometimes makes mistakes when pilots are sent to attack innocent civilians instead of militants.