This study investigates how teachers understand and position AI tools in middle school writing instruction, drawing on 27 teacher interviews collected during a study called Project Topeka that used an interactive argumentative writing platform with AI-generated scores and feedback. Based on the interviews, we generate an initial theoretical framework of how teachers position AI tools -- and therefore themselves -- in their teaching. We found that some teachers leveraged AI as a "teaching partner" that provided insights to help enhance teaching and learning while remaining central to instruction themselves and interacting with students in numerous ways. Others delegated aspects of assessment and learning to AI as a "grading assistant" to save time and increase efficiency, interacting with students with a slight emphasis on score attainment over skill development. Another group turned instruction over to the AI tool as if it were a "substitute teacher," interacting minimally with students and placing themselves on the instructional periphery. We describe each approach in detail and discuss implications for teaching practices, teachers' roles, the profession, and students' experiences and opportunities. [This work was supported by Gates Ventures.]