How to Use AI for Homework: The Version Where You Still Do the Work

"Use AI for homework" is a phrase that hides two completely different
activities. One is: I read the problem, I work out the answer myself,
and I use the tool to catch things I missed. The other is: I read the
problem, I paste it in, and I copy what comes back. Both are "using AI
for homework." Only one of them is doing homework.

You can start with the free tier, no card required, on your own notes and see whether the loop holds for your course before you pay. Get started on the free plan
This page is about the second one, because that is the version most
students actually drift into without noticing, and the grade eventually
reflects it even if the single assignment looks fine. The fix is not
to stop using the tool. The fix is to change when in your workflow
you hand anything to it.

The actual point of a homework problem

A homework problem in a STEM course is doing four things at once:
it is testing whether you picked the right concept, whether you
connected that concept to this particular situation, whether you
carried out the manipulation correctly, and whether you can say
what the result means. Most students treat all four of those as
one single task: "get the answer." That framing is what makes a
tool feel like a shortcut, because a shortcut is defined relative
to a goal, and if the goal is "get the answer," then a tool that
gives you the answer is, by definition, the fastest route to
the goal.

But the goal was never the answer. The goal was the four things
above. If you have been teaching your brain which of the four to
reach for, in which order, and what it means when one of them
fails, then the answer is a byproduct. If you have not been
teaching your brain that, the answer is the whole point, and
you have just outsourced the whole point.

The three moments in a homework session, and which ones to protect

A homework session has three distinct moments, and each of them
deserves different treatment from a tool.

The diagnosis moment. This is before you do anything. You
read the problem, and you decide what kind of problem it is.
This is the hardest moment in the session, and it is the one
that carries the most learning. If you skip this and go straight
to "how do I solve this," you have skipped the part where you
demonstrate to yourself that you know what you are looking at.
This is the moment you should protect most aggressively, because
it is the one that is the hardest to do on your own when you are
tired, and it is the one a tool is happiest to just do for
you.

The work moment. This is the manipulation. The algebra,
the integration, the derivation, the arithmetic of a
statistical test. This is the moment where a tool is genuinely,
usefully helpful, because it is the mechanical part, and
mechanical parts are where humans make careless mistakes that
have nothing to do with whether they understood the material.
Using a tool here is not cheating, it is just being smart
about where your finite attention goes.

The judgment moment. This is after you have an answer.
Does it make sense? Is the sign right? Does the magnitude fit
the situation? If I doubled the input, would the output do what
I think it would? This is the moment most students skip
entirely, because they have an answer and they move to the
next problem. Skipping this moment is the single most
expensive habit a homework student can have, because it is
the habit that quietly decouples "I did the work" from
"I can do the work again next week."

You do not have to do all three of these moments on your own,
but you need to be clear about which ones you are doing on
your own and which ones you are handing to a tool, and
whether that split is intentional or accidental.

The split that works for most people

The split that works in practice is this: do the diagnosis on
your own, use a tool in the work moment as much as you want,
and do the judgment on your own. It is a simple rule, and it
happens to protect exactly the two moments that carry
knowledge and leave the moment that is just mechanical
open to whatever tool is most convenient.

The diagnosis on your own does not mean you do it with no
help. It means that when you are stuck, your first move is
not "here is the problem, solve it" but "here is what I
think this problem is about, is that right." The second
form is a check, and a check requires you to have said
something first. If you have said nothing, you cannot be
checked, and the tool has nothing to check you against,
so it just answers.

The judgment on your own is a different kind of pressure. By
the time you are in the judgment moment, you have an answer,
and it is tempting to treat the answer as settled and move on.
The question to ask, in that moment, is not "is my answer
right" but "is there a version of this problem, close to the
one I was given, where my answer would be wrong, and I
haven't noticed?" A tool can be very useful in that
conversation, because it can look for the version of the
problem you have not considered, but the version of the
problem has to come from you, not from the tool,
or you have not done the judgment, you have
done another check.

What the loop looks like on a real problem, end to end

Let me walk through a real intermediate mechanics problem with
the split above, because the abstract version of the rule is
easier to state than it is to actually hold:

The problem. A car of mass m travels along a straight
road. It applies a constant braking force for a time t,
comes to rest, and the brake pads heat up. Estimate the
fraction of the car's initial kinetic energy that ends up
as thermal energy in the pads, and state your assumptions.

Diagnosis, on my own. This is an energy conservation
problem with a twist: the twist is the "estimate your
assumptions" clause. The core concept is kinetic energy,
the manipulation is the work-energy relationship, and the
judgment is what fraction actually goes where, and what
my assumptions are doing to that fraction. I can say that
much without any help, and saying that much is the point
of the diagnosis.

Work, with a tool. The actual algebra here is simple,
but I do want to check my sign conventions and my
assumptions list, so I show my setup and my assumptions
to a tool and ask it to flag anything inconsistent.
That is a legitimate use, because the mechanical part
is the stage where a mistake is a typo, not a conceptual
gap.

Judgment, on my own. Here is the piece that actually
tests whether I understand the problem. The "estimate"
clause means the answer depends on my assumptions, and
if I assume all the energy goes to the pads, I am
wrong, because some goes to the tires, some to the
brake fluid, some to the road. A tool can help me find
the versions of the problem I have not considered, but
the version I have to name is the one where my assumption
breaks, not the one the tool happens to find.

That is the whole session, and the split I used is
diagnosis-on-me, work-on-us, judgment-on-me. It is the
split that keeps the learning with me and uses the
tool where it is actually good.

Where the shortcut version quietly becomes the only version

There is a pattern here worth naming directly, because it is the one
most of us recognize in ourselves more than we are willing to admit:
the first week of a course you do the diagnosis yourself, because you
are still close to the material and the problem feels fresh. By the
third week, the diagnosis is the piece that feels slow, so you skip it.
By the fifth week, the diagnosis is the step you skip entirely, and
you have not noticed that the skip is happening, because every
single session still ends with a right answer.

The reason this happens quietly is that the shortcut version
does not feel like a shortcut. It feels like efficiency.
It feels like using the tool where it is good. And it is
true that the tool is doing the mechanical part well.
But the diagnosis is not the mechanical part, and if
you are not doing it, you are not building the skill
of knowing what you are looking at, which is the skill
that carries you through the next problem, and the
one after that, and the one on the exam.

The fix is not to resolve to never skip the diagnosis,
because you will, on the tired night, and on the night
you are behind. The fix is to make the skip visible
instead of invisible. If you can name the moment where
you stopped diagnosing and started asking for the
answer, you can put it back. If you cannot name
that moment, it is happening every night, and
you do not know it.

That is the honest version of the risk, and it is the
reason the diagnosis moment deserves the protection
I described above, and the reason a tool like Fennie,
which is built to ask you what you think before it
gives you anything, is the right kind of tool for
this particular use case.

The honest line between help and a shortcut

The line is not "is the tool smart enough to do the whole
problem" (most of them are), it is "did I do the part
where I was supposed to learn?" If the answer to that
question is no, the tool is a shortcut, even if it is a
helpful one. If the answer is yes, the tool is a
check, even if it is the same tool that could have
done the whole thing.

Fennie is on the check side of that line by design.
Its positioning is explicit about this: it asks you
what you are thinking before it offers more, and its
free tier is built around that loop rather than around
a "give me the answer" mode. If you want to see
what that looks like on your own problem set, run your own notes through a free, no-card loop.

Practical rules I would actually follow

Read the problem out loud to yourself once, and say
what it is testing before you look for a solution.
If you cannot say what it is testing, that is the
actual place you are stuck, and it is worth spending
ten minutes on before you spend an hour in the
solution.

Write your setup on paper before you type anything.
The physical commitment surfaces what you actually
know versus what you assume you know, and that
gap is where the tool is most useful, because
it is the gap you cannot see from inside your
own head.

Ask the tool to find the version of the problem
where your answer would be wrong. That prompt does
two things at once: it forces you to have an actually
stated answer, and it gives the tool something to
check against instead of a blank, so it cannot just
answer the question you were originally going to ask
it.

Treat "it looks right" as a trigger to ask for
the judgment pass, not as a signal to move on.
"Looks right" is doing the judgment pass for you, and
it is doing a poor job. The version that is useful is
the one you can defend in a single sentence, and that
sentence is where the tool earns its place.

Bottom line

Using AI for homework is not a yes or no. The question is which moments of the session you are doing on
your own, and which ones you are handing to a tool, and whether that split is one you chose or one that
happened to you. The split above, diagnosis on you, work with the tool, judgment on you, is the one
that keeps the learning with you, and it is the one that holds when the semester ends and the tool is
no longer in the room. Fennie is built for that split by design, and
you can run your own problem set through the free, no-card loop
to see whether the split holds for the course you are in.