We often discuss in education the importance of keeping the “Human-in-the-Loop” (HITL) when using generative AI. It’s the golden rule of AI literacy. Don’t let the robot do the thinking. Use it to amplify your cognition, not replace it.
But what does that actually look like in practice? When a student sits down with an AI tool, how can we distinguish between collaboration and delegation?
This week at the Literacy Research Association (LRA) conference, I presented new research on this exact question. Instead of just looking at the final essays students produced, we opened the “black box” of their interaction logs. We examined the timestamps, edit history, and the specific prompts they used to negotiate with the machine.
What we found was a tale of two distinct interaction styles, and a new definition of what it means to have agency in the age of AI.
The Digital Trace: The Loop vs. The Line
In our analysis of students using NotebookLM, two distinct profiles emerged. I call them The Orchestrator and The Outsourcer.
1. The Orchestrator (High Agency) For this student, the interaction log looked like a messy, circular loop.
- They prompted the AI.
- The AI generated a summary.
- The student read it, paused, and refused it.
- They re-prompted with constraints: ” Don’t just list facts. Capture the tone.”
The Orchestrator engaged in what I refer to as a Correction Loop. They treated the AI not as an oracle, but as a junior analyst whose work needed to be checked, challenged, and often rejected. Their agency was defined by friction. They worked harder than the machine.
2. The Outsourcer (High Delegation) For this student, the interaction log looked like a straight, vertical line.
- They prompted: ” Help me study.”
- The AI offered a structure: ” Here is a quiz and a study guide.”
- The student accepted it immediately.
There were no deletions, no re-prompts, and zero friction. Gravity took over. The student delegated the structure of the knowledge to the AI. While efficient, this represented a form of cognitive offloading. The machine did the heavy lifting of organizing the concepts.
Moving from Verification to Valuation
In the era of Google Search, we taught students to check for Credibility (Is this true?) and Relevance (Does this answer my question?).
In the era of Generative AI, that isn’t enough. RAG (Retrieval-Augmented Generation) models, such as NotebookLM, usually pass the credibility check because they are grounded in your sources.
The new literacy skill is Valuation. The Orchestrator didn’t reject the AI because it was wrong ; they rejected it because it was shallow.
- The Outsourcer asks: ” Is this acceptable?” (Yes).
- The Orchestrator asks: ” Is this nuanced?” (No).
The Takeaway: Agency is the Refusal of the Generic
The big lesson from this research is that meaningful AI integration requires us to teach students to raise their “Threshold of Satisfaction.”
If a student accepts the first, safe, generic draft the AI spits out, they have outsourced their agency. To be an Orchestrator, they must be willing to embrace the friction. Stop the line, reject the default, and force the tool to submit to their specific voice and epistemic standards.
Agency isn’t just typing a prompt. Agency is the refusal of the generic.
Dig Deeper
I have uploaded the full slide deck, including diagrams that show the “Correction Loop” versus the “Straight Line of Delegation,” to my digital garden. You can also find my full notes on the methodology and the coding schema we used to trace these interactions.
👉 Read the full breakdown here in my digital garden
Note: Images in the slide deck were created using Google’s Nano Banana or NapkinAI.