The massive popularity of IoT devices raises new challenges for user privacy. Hence, manufacturers are obliged to notify users about their privacy practices as well as give them choices to have control over their data. Privacy policies are long and full of legal jargon, thus not understandable by average users. The problem becomes worse with IoT devices due to the ability of these devices to access sensitive information about users. Previous research has addressed problems related to websites and mobile privacy policies. However, few works focus on analyzing IoT privacy policies. In this paper, we analyze and annotate 50 IoT privacy policies to determine whether the IoT manufacturers collect personal information about the user as well as the type of such information. To ensure that we extract the correct information, we study in-depth the complicated and ambiguous sentences that average users won't understand. With our method, we aim to mimic how an ordinary person reads and understands such policies sentence by sentence. We use supervised machine learning to label the collected personal information according to its sensitivity level to either sensitive personal information or non-sensitive personal information. The high accuracy achieved by the classifier (98.8%) proves its validity and reliability.