Stop Policing AI Use. Observe Learning.

I have been puzzled about artificial intelligence (AI) use in education. I have been reading, watching talks, and talking with students, teachers, and people who hire. This post is my attempt to summarize what I think I have learned so far. These are only my thoughts and not a recommendation.

Students do not start on the same line¶

University students do not start on the same line. Some already use advanced AI tools. Others can use a computer but do not know about those tools or how to use them. Some still struggle with basic computer skills and do not know how a computer’s file and folder system works.

Teachers are uneven too. Some adopt AI quickly in teaching and research. Others lag, partly because the tools change fast and partly because of the courses they teach.

A take-home or unsupervised online assignment often hides whether learning occurred. People I talk with worry that AI use will replace thinking. I separate two cases. Cognitive offload is handing a task to a tool while you continue to think. That can be acceptable. Cognitive stunting is when engagement with problem-solving drops away. That is not acceptable. I also hear that employers want graduates who can use these tools at work.

Access and literacy come first¶

The tools help only if students know they exist. I want access and literacy visible. Many students do not know that access is already in place, or which courses teach how to use the tools.

Access and tool-use literacy then raise harder questions. What counts as AI use? What counts as AI abuse? How should a course teach responsible use?

What I keep hearing¶

In those conversations, people who use AI often treat it as a productivity tool. They want clear, consistent rules on what counts as allowed AI use versus cheating, so they know the line before they submit work. They want fair access to paid tools when free ones fall short.

Teachers I talk with treat AI as an integrity and critical-thinking issue. They want to know whether students learned the course material. They do not want false positives or false negatives when they judge whether a student learned the material. They are stuck when they cannot validate authentic student learning.

Both groups want literacy about access and use. Both also want a clear definition of what counts as AI use. Is it only running a generative AI tool end to end? Or is a single click on a PowerPoint design button also AI use?

Some of the teachers I talk with ask whether naming every AI tool still matters. AI use is now so common that many treat it as ordinary.

Students also say a single line on a syllabus cannot enforce integrity. When students who keep integrity watch friends cheat, get away without being caught, and earn better grades, those students feel punished for holding to integrity.

How I use these words¶

AI use means any help from an AI tool on teaching, learning, and research. That help can be a full generative run. It can also be a small built-in feature, such as a design button in presentation software. I count that button as AI use.

Transparency is not a ritual of saying you used AI. It is an ethical duty and right to be accountable for the final output. I do not focus on how the project was made. I focus on whether the author will stand behind its accuracy, ethics, and original value. I think, this standard does not weaken productivity, quality, or innovation.

Responsible AI use is explainability: the author’s readiness and capacity to articulate, validate, and defend the work. Privacy, attribution, and care for data still matter. They are not the same as proving that learning occurred. The burden of proof is the author’s defense of outputs and methods, not a detector’s guess about which tool was used.

Integrity means holding to that standard of transparency and explainability when no one is watching. I cannot see that directly. So I do not try to infer which tools a student used in the dark. I want to observe whether the student can defend the work.

Make learning observable¶

Asking a student to explain a proof at the board is ordinary teaching. A timed exam in the room is an exam condition. I do not pretend those are the same as watching a student through a camera.

Access to AI tools without observation makes take-home and unsupervised online work easy to fake. If an easy path exists and no one can see the work being done, many take it. Take-home and online assessments then hide whether learning occurred.

Teaching responsible AI use is necessary but not sufficient. A better next step is to build a responsible AI product for learning that solves the logistical nightmare. This step sits closer to sufficiency. Large firms that sell AI tools have a reason to lock people into their tools. So their idea of a responsible AI product is not the same as a teacher’s.

A responsible AI product deters abuse. It stays friction-free. It solves the logistical nightmare. It makes learning observable so originality shows. It tests whether the student can explain and defend the work. It holds integrity by validating authentic learning. It does not require a ritual declaration of AI use.

How I would make learning visible¶

A responsible AI platform should evaluate understanding of the course material at least through three systems. AI may support logistics and scoring. The object of measurement is the student’s defense of the work, not a detector’s guess about tool use. These systems can be oral, written, or peer-then-oral.

Oral examination¶

Students speak their answers under a time limit. The exam stays tied to what I taught, or to the assigned chapter. A student who never read the assigned material can still produce a fluent answer, or even a correct one. That answer still fails if it is ungrounded in the assigned material. AI can help me score. I still lead the rubric, feedback, grading, and accommodations. The intent is to see whether the student learned, not to detect AI use.

Written examination¶

A written exam can be timed, in the room, and handwritten. The integrity gain is the exam setting. A computer-vision model can read images of the page for scoring. Computer vision here means a model that reads images of the page. The work can include math, figures, graphs, and other written material. The AI model can be a grading aid.

Peer review, then oral defense¶

A student writes a narrative or essay. Peers review it. The student revises the draft using those peer review comments and my rubric. Then the student submits a final version. I would then run an oral exam on that student’s own submission. The first draft, the peer review, the revision, and the spoken defense can all be scored with AI support. Peer review can be faked the same way an essay can, so the spoken defense of the student’s own final draft is the test that matters to detect authentic learning.

These modes fit recap-and-defense courses. Studio art, coding, and group projects often do not. There the work is a made object and a shared process, not a recap of a chapter. Those courses still need an observable artifact and a defense the student can stand behind. I do not have a second design for them here.

Conditions I would require¶

The aim is not to replace the professor. I want grading logistics off the desk, faster feedback, and time to stay engaged with students and to teach better.

I would only use these formats if feedback is timely, the grade is fair, and accent, disability, or weak internet do not decide the mark. The freed time should go to teaching and to hearing defenses.

Match AI use width to the course goal¶

I would set the allowed width of AI by the learning goal of a course.

Domain-knowledge courses build core knowledge that must live in the student’s head. Examples include grammar, pre-algebra, statistics, accounting basics, and core linguistics. In those courses, we should forbid generative substitution on graded work, or allow only narrow, named uses. Students need that knowledge before they can guide and check a tool.

Systems-thinking courses join that knowledge to solve end-to-end problems. Examples include econometrics, upper-level marketing, supply chain, intermediate or advanced accounting and finance, and risk. In those courses, we should allow wide use. Use may even be full, once the student has passed the matching domain course. AI can help with data analysis, idea generation, and workflow design. The student must still be able to critique the tool.

This is the third step of the sequence. Narrow use in domain courses protects the knowledge faculty want in the student’s head. Wider use in later systems courses is how graduates become able to use the tools at work. The two course types have to be ordered. Width by course type without that order will not serve students, faculty, and industry at once.