Teaching students to work with AI

They will use AI at work.
Somebody has to teach them how.

Your students will walk into jobs where AI is already on the desk, and nobody will hand them a tutorial. The skill is not typing a prompt - it is knowing what to ask for, what to accept, what to check, and when the answer is not good enough. This page is about how to teach that and, harder, how to grade it.

What the skill actually is How to grade it
What gets graded
just clean the data and give me revenue by region
I can explain what needs fixing and show you one worked example, but I do not run the analysis. What did you find when you looked at the date column?
there are three date formats. use ISO and drop the rows with no region
Marked on the exchange, not the answer
First askvague - handed the whole problem over
Recoverednamed the actual defect and specified the rule
Missingnever checked how many rows the drop removed
A transcript is the evidence. The question is what you look for in it.
Where most courses are

Banning it and ignoring it
are the same policy.

Both leave the student to learn this on their own, from whoever gets to them first - a group chat, a YouTube tutorial, or their first manager.

The ban
Detection does not hold
Detectors are unreliable and the penalty for a false positive is severe, so most bans end up enforced on the honest and invisible to everyone else. The deeper cost is what it teaches: that AI is contraband rather than a tool with a right and a wrong way to hold it. There is still a real place for work a student must do unaided - see designing out the shortcut - but that is a decision about which assignments, not a curriculum.
The shrug
“Use it however you like”
Permission without instruction produces the thing everyone complains about: a fluent, confident, slightly wrong submission that the student cannot defend, because they were never asked to. The student is not cheating so much as failing to supervise - and nobody taught them that supervising was the job.

The third option is to treat working with AI as coursework: something you assign, something students get better at, and something you can put a grade on.

What the skill is

Six things that separate
directing it from deferring to it.

None of these are prompt tricks. They are judgment, and they are all things your discipline already teaches - applied to a collaborator that is fast, confident and sometimes wrong.

01
Framing the problem
The student has to know what they are actually asking before they ask it. “Clean this data” is a wish. “Three date formats in column B, normalise to ISO and tell me how many rows you dropped” is an instruction. The gap between those two sentences is most of the skill.
02
Specifying precisely enough to be built
Vague requests produce plausible rubbish. Making a student specify a thing tightly enough that it can be built - the inputs, the formula, the edge case, what it must never do - is an exercise in knowing the domain, and it fails loudly when they do not.
03
Judging the output
Can they tell a good answer from a confident one? This is where subject knowledge is load-bearing: you cannot evaluate a discounted cash flow, a care plan or a legal summary you do not understand. AI raises the value of knowing the material; it does not remove it.
04
Checking, not trusting
Every number traced, every citation opened, every claim tested against something the student already knows. The habit to build is the reflex to ask “how would I know if this were wrong?” before the work goes anywhere.
05
Knowing when not to use it
Some work is the learning. Reaching for a model to do the one thing the course exists to teach is a judgment failure, and students should be able to name which parts of their own field those are - and say why.
06
Owning the result
In a job, “the AI wrote it” is not a defence. Having to stand behind the output to a person who pushes back is what converts a transcript into understanding - and it is the part a chat window never asks of them.
What to assign

Five assignment shapes
that teach it.

These work in any discipline, and you can run the first three with nothing but the tools you already have. What they share is that the AI is inside the task, with a constraint on it, so the student has to make the calls.

01 · The bounded assistant
An AI that explains but will not solve
Give students an assistant with a rule they cannot talk it out of: it explains any formula, shows one worked example, and never runs the analysis. The student has to arrive at the method themselves, using it the way they would use a knowledgeable colleague who is busy.
What it teachesFraming, and the discipline of asking for the thing you actually need
02 · Specify it into existence
Describe a tool precisely enough to be built
The student writes a specification - a break-even calculator, a triage rule, a scoring model - tight enough that it can actually be built from their words. Then they use what comes back and find out where their spec was ambiguous. The revision is the lesson.
What it teachesPrecision, edge cases, and that vague thinking produces broken tools
03 · Find the flaw
Hand them AI output that is wrong in a plausible way
Give the class a memo, a model or a diagnosis produced by AI with a real error in it - a defensible-looking assumption that does not hold, a citation that does not say what it is cited for. Ask what is wrong and how they would have caught it. Nothing builds scepticism faster than being handed something confident and false.
What it teachesJudging output, and the check-first reflex
04 · Direct, then defend
Work with AI, then answer to a person about it
The student uses AI to prepare something - an analysis, a recommendation, a plan - and then has to take it into a conversation with someone who disagrees and asks where a number came from. Borrowed confidence collapses immediately under a follow-up question; understanding does not.
What it teachesOwnership, and whether the understanding transferred
05 · Declare the line
Make them say where AI should stop
A short written judgment: on this task, which parts would you hand to AI, which would you keep, and why? Graded on the reasoning, not the answer. Done early in a course it surfaces what students believe about their own field, and it is worth revisiting at the end to see what moved.
What it teachesKnowing when not to use it, in their own discipline's terms
The hard part

Grading how they used it,
not what it produced.

If you grade the output, the best-supervised student and the one who pasted the whole prompt can hand in the same thing. The evidence you want is the exchange - and the reason this is rarely taught is that reading forty transcripts by hand is not a plan.

Did they name the actual problem?
Look at the first substantive ask. A student who has understood the task asks about the specific defect; one who has not hands over the whole assignment and hopes.
Did they specify, or accept?
Did they set the rule, the range, the constraint - or take the first proposal as given? Accepting a reasonable-sounding default without examining it is the single most common failure.
Did they check anything?
Evidence of verification: a number traced back, a claim tested, a row count questioned, a source opened. Its absence is the finding, and it is visible in seconds once you know to look.
Did they push back when they should have?
The strongest signal in any transcript is a student refusing an answer that was wrong, and saying why. The weakest is a thread with no friction in it at all.
Could they defend it afterwards?
A short follow-up - written or spoken, to a person or to a sceptic who asks for the derivation - separates transferred understanding from a well-formatted paste.
Did they know where to stop?
Did they use it on the part that was theirs to learn? A student who declines to hand over the core reasoning has understood the point of the course better than the rubric usually captures.

Those six read as a rubric because they are one. The practical obstacle has never been knowing what to look for - it is that the evidence lives in a chat window the student closes, on a tool you cannot see into.

Where Kova comes in

The exchange happens
inside the assignment.

Everything above is teachable without us, and worth doing either way. What a platform changes is the part that does not scale: the AI runs inside the coursework under rules you set, so the transcript is the submission and the grading has something to stand on.

The rules are yours
You write what the assistant does and never does - explains but will not solve, builds only what is specified, asks what they tried first. Students cannot talk it out of them, so the constraint that makes the exercise work actually holds.
See AI agents & the Coding Agent
The transcript is the work
Every exchange is kept with the submission and graded against what a strong side looks like - what they asked, what they checked, what they refused to accept. Not a chat log you have to go hunting for.
See AI-assisted grading
Specify it, and watch it get built
The student describes the tool they need and it is built live in front of them, so a loose specification fails visibly and they revise it themselves. Three attempts, or as many as you allow.
See the AI page build
Then they defend it to someone
A conversation with a character who holds a position and asks where the number came from - over email, in a meeting, in a team thread. Graded on their side of it.
See AI conversations
And where AI should not be in the work at all
Not every assignment should involve it. The same personalization that makes this work also makes an assignment resistant to being pasted into a chatbot - each student gets their own version, set in their own context, with a live exchange no generic draft can fake.
Designing out the shortcut
Questions we get

The ones that come up
every time.

Should I just let students use ChatGPT on assignments?
Not as a blanket policy in either direction. Decide per assignment, and say which it is: some work is the learning and should be done unaided; some work is closer to what the job looks like, and there the useful question is not whether they used AI but how well they supervised it. What does not work is leaving it unstated, which is what most syllabi currently do.
How do I stop students just pasting the question in?
Make pasting the question a bad strategy rather than a forbidden one. An assistant that refuses to produce the answer, a task set in a context unique to that student, a live exchange that asks follow-up questions, or a defence afterwards to someone who pushes back - each of these makes the paste fail on its own merits, with no detection involved.
Does teaching AI skills mean less subject knowledge?
The opposite. Judging whether an answer is right requires knowing the material - a student who cannot evaluate a cash flow statement cannot supervise one being written. AI raises the return on domain knowledge because knowledge becomes the thing you check with.
I am not technical. Do I need to be?
No. Everything on this page is a teaching judgment, not a technical one: what the task is, where the line sits, what a good exchange looks like. Writing the rule for an assistant is a sentence in plain language - “explain any formula, never run the analysis, ask what they tried first” - not configuration.
How do you grade a conversation fairly?
Against a written statement of what a strong side looks like, the same way you would grade an essay against a rubric. The dimensions are consistent - did they frame it, specify it, check it, push back, own it - and the transcript is the evidence for each. It is more defensible than grading an output that may not be the student's work at all.
Is this the same as AI literacy?
Related, but narrower and more practical. AI literacy usually means understanding what the technology is and what it does to society - worth teaching, and mostly taught in a lecture. This is the applied half: the judgment a student uses while working with one, in your discipline, which can only be built by doing it and graded by looking at how they did it.
Teach it in one course
this semester.

Bring an assignment you already set. We will turn it into one where students direct an AI under your rules, and you see exactly how each of them handled it.

Pilot this semester Request a demo