Blog article Education
Assessment with AI: where is the student’s work?
A polished answer can hide very different decisions. I propose looking at the questions, revisions and criteria a student can explain.
Vinicius Covas3 min read
The problem begins before submission
Imagine two students submitting equally clear essays. One questioned the assignment, checked sources and rejected a model suggestion because it lacked support. The other accepted the first answer and changed a few words. Their final products may look similar while their intellectual processes differ considerably. Assessing only the finished document loses the distinction that matters to teaching. Evidence of learning should therefore be designed before opening the tool: which decision do we want to observe, by which criterion and at what point in the process?
What LOGOS contributes
The LOGOS preprint proposes examining human cognitive agency in AI-assisted tasks. It is a five-level conceptual taxonomy, not evidence that a teaching strategy improves grades. Its question offers a starting point: what intellectual work is the person doing while using the system? For an assignment, I would translate that question into observable actions: framing a problem, comparing interpretations, justifying a choice and revising a conclusion. A level label should never replace examination of the actual work.
An exercise that makes decisions visible
I would ask for a short initial explanation without assistance, followed by an AI conversation aimed at finding an objection. Submission would include a revised version and a comment on one accepted suggestion and one rejected suggestion. Assessment would address the quality of reasons and source use, not the number of messages sent. If the model offered an overly broad definition, the student would identify a case it fails to distinguish and propose a better formulation. The revision matters because it can be defended, not because it sounds more academic.
Transparency without permanent surveillance
Documenting a process does not require recording every student action. A small set of justified decisions can be more useful than an endless chat history. Agree in advance which uses are permitted, which data should not be shared and how students without paid accounts will be assessed. A fair assignment needs comparable alternatives. If success depends on buying a better tool, the rubric may confuse technological access with learning. Explain these conditions before work begins, using examples every student can recognize.
What would need to be tested
The activity above is a teaching design proposal, not a validated intervention. Studying it would require defined outcomes, comparable work and checks that criteria are applied consistently. It would also be useful to see whether students can tackle a new case and explain its limitations without returning to the same conversation. My starting criterion is simple: a good AI-assisted assignment should make human reasoning more visible. If it merely polishes the product while hiding decisions, we still do not know what was learned.
Communication researcher and professor. I study how humans and machines build meaning and relationships.
About the author