Student Resources

One thing is worth making clear at the outset: the students who get the most out of these tools tend to be the ones who already know the most. Skill with AI is not a substitute for knowing your subject. It rests on it.

You can see why from how the tools work. To get something useful out of a model, you have to know enough to ask a specific question, supply the context that matters, notice when an answer is a little “off,” and recognize what it left out. All four of those are subject knowledge. Without them you get fluent, confident, average output and no way to tell whether it is any good. Anthropic’s analysis of how people actually use these systems found that the sophistication of what people put in and the sophistication of what they get back track each other almost exactly — the tool does not close the gap between users, it reflects it.

The case for doing the work yourself, then, is not only that learning is worthwhile. It is that knowledge is what turns an ordinary AI result into a good one. That leads to three questions, which come in a fixed order. Most guidance addresses only the first.

Your instructor’s policy is the deciding factor.

Question #3 never overrides Question #1: If a course prohibits AI on an assignment, that settles it, whatever you conclude about your own learning. Nothing in this section is permission to substitute your judgment for a course policy — it is what you do with the space a policy leaves open, which in most courses is now considerable.

These systems are genuinely excellent at some things and genuinely bad at others, and the difference is not obvious from the inside. Most of what goes wrong for a student over a term traces back to handing over the wrong part of the work — sometimes as a grade penalty or an integrity case, more often as the quieter problem of a degree that did not build what it was supposed to build.

Feeling that you have learned something is not the same as having learned it

This is well established in research on learning, and it is seldom mentioned in guidance about AI.

Performance is not learning. How well you can do something right now, with the material in front of you, is a poor indicator of whether you will be able to do it next month in a different context. Durable learning — learning that lasts — requires friction:

• Spacing involves returning to material after you have partly forgotten it, making recall harder. The harder the recall the more likely the material finds a place in your noggin where it sticks. Thus, giving something a second reading is good, but spacing out that second reading — maybe a day or two after the first — is what makes it last.
• Retrieval involves pulling an answer out of your own memory rather than looking it up. Coming up empty for a second is uncomfortable, and that reaching around is exactly what strengthens the memory. Thus, rereading your notes feels productive, but closing them and asking yourself what was in them is what does the work.
• Interleaving involves mixing problem types together instead of doing twenty of the same kind in a row. When every problem on the page is a percentage problem, you stop asking what kind of problem it is and just start multiplying; when they are mixed — a percentage, then a ratio, then an average — you have to figure out what you are looking at before you can start. Thus, a jumbled problem set feels worse and teaches you the harder half of the skill: knowing which tool the situation calls for.
• Generating an answer before you see one means committing to your best guess before the solution is in front of you. Being wrong on purpose marks the exact spot where your understanding ran out, so the right answer lands in that gap instead of sliding past. Thus, guess first and check second. Even a bad guess makes the correction stick.
• Varying the conditions of practice means not always studying the same way in the same spot — different rooms, different formats, different order. Practice in one setting quietly ties the knowledge to that setting, and you will need it somewhere else. Thus, switching up where and how you study makes the material easier to find later, in a room that looks nothing like your desk.

Every one of these approaches slows you down and feels worse in the moment. But every one of them beats the easier alternative when it comes to fostering actually learning.

The same tool, two different outcomes

The variable is not which model you use or how good your prompt is. It is whether you do the effortful part or hand it over. Let’s consider some common student “tasks” and how AI can be used to bolster learning or to shortcut it.

Three questions worth asking yourself

1. Could I do this again next week without the tool? If not, you have produced an artifact, not a capability.
2. Can I explain why, and not only what? Being able to state the answer is performance. Being able to defend it is learning.
3. Did I have the idea, or did I recognize the idea? Recognizing a good answer feels almost identical to generating one, and it is a completely different skill — and it is the one that does not transfer.

Why this matters for your career

Set aside academic integrity for a moment. There is a practical case as well.

The tasks most exposed to automation right now are disproportionately the junior tasks — the summarizing, the first drafts, the routine analysis, the basic code. Those are exactly the tasks people have always learned on. Multiple independent datasets show employment and hiring effects concentrated among workers in their early twenties in the most AI-exposed occupations, while employment for experienced workers in the same occupations has held steady. The causal role of AI is still contested among economists. The pattern is not.

Read those two facts together and the implication is uncomfortable. The rungs of the ladder are thinning at the same moment the tool exists that lets you skip them in school. If you skip them in both places, the judgment that separates a person who can supervise AI output from a person who can only generate it never gets built — and supervising the output is the part of the job that is not going away.

The graduates who do well will not be the ones who used AI the most, or the least. They will be the ones who can tell when the answer on the screen is wrong.

That capability is not a personality trait. It is domain knowledge plus practice, and both of them are built by doing the work at the exact moments when a tool could have done it for you. Build the knowledge and the tool multiplies it. Skip the knowledge and the tool is all you have — which is not enough, and everyone else has it too.

Eight rules we can all agree on

The Decision Flow: “Can I use AI on this?”

The One-Sentence Version

If you would be uncomfortable explaining exactly what you did to your instructor, that discomfort is the answer. And when uncertain, ask your instructor.

Disclosures and Citations

Two different things, frequently confused:

• Citation is a formatted reference entry, used when you quote or reproduce AI output.
• Disclosure is a plain-language statement of what you did, and it is what most instructors want.

Below are some examples of AI disclosure statements. However, if AI usage is allowed, check — then double check — the disclosure guidance provided by your instructor.

State what you used it for. One or two sentences, specific. “Generated eight candidate counterarguments; used two. Checked APA formatting on my reference list.”

Clarify what you verified. Which claims you checked independently, and against what. “Verified both statistics against the original agency reports.”

Make clear what is entirely yours. “The argument, the structure, and all of the writing in sections 2 through 4.”

What never goes into a chatbot

This is the rule with actual legal consequences attached, and it is the one students receive the least instruction on. But here’s something clear and tangible:

Confidential data must never be entered into AI tools and information shared with these tools using default settings is not private.

There are versions of tools and models that add data loss safeguards. Unless you are 100% certain you are using one of these versions, assume you’re not and act accordingly.