The powerful new AI model is designed to analyze intercepted communications – but experts say such systems can exacerbate biases and are prone to making mistakes