A first-time guide to teaching with AI

For instructors new to allowing AI in a STEM course

This guide is for an instructor who is about to teach a STEM course in which students have access to AI, and who wants them to come out of it having actually learned. It does not assume you have a policy yet, or that you have decided how permissive to be. It assumes only that pretending the tools are not there is no longer an option.

The one shift to make first

Students can reach an AI assistant whether or not your syllabus permits it. So the useful question is not how to keep AI out of your course, but how to teach so that students still build the competence the course is for, with these tools in the room.

One principle runs through everything below. The verification you teach is the verification you owe. If you want students to trust their own work rather than the AI’s, you have to teach them how to check it, and you have to hold your own course to that same standard: assess what they can actually do, not what they can produce with help they cannot yet evaluate.

It helps to separate three things a student needs, which this book calls core competence, AI-assisted practice, and trustworthiness:

• Core competence is what a student must be able to do unaided, because it is the foundation everything else stands on.
• AI-assisted practice is the work they can reasonably do with the tools, once that foundation is there.
• Trustworthiness is the habit of checking AI output rather than taking it on faith, so that “the AI said so” never stands in for understanding.

Redesigning a course is mostly a matter of being clear about which is which, and making sure your assignments and your assessments line up with that split.

Start small, not all at once

You do not have to redesign everything before the term starts, and you should not try to. Two common opening moves both fail. Banning AI outright drives use underground and teaches students nothing about using it well, and the evidence in this book is that bans are hard to enforce and easy to rationalize around. Going fully permissive with no change to how you assess lets students hand in work they cannot reproduce or explain, and you find out too late.

A better path is to make a few deliberate changes in your first term and add more as you see what happens.

Four concrete first moves

1. Decide your policy and state it plainly. Students behave better when the rules are explicit and they understand the reason. Decide how permissive your course will be, write it into the syllabus in plain language, and say why. You do not have to draft this from scratch: instructors/briefings.md has ready-to-adapt syllabus paragraphs for a restrictive, a default, and a permissive stance, each with reasoning a student can follow.

2. Redesign one assignment to assume AI access. Pick a single assignment and rewrite it on the assumption that students can use AI, so that doing the work still requires understanding. The usual move is to ask not just for an answer but for the reasoning, a check that the answer is right, and a short note on where AI was used. An assignment that can be finished by pasting a prompt and copying the reply is one the tools have already made obsolete; an assignment that requires the student to verify and explain is not.

3. Add a low-stakes check of unaided competence. You need some way to see what students can do without the tools, or you are assessing the AI rather than the student. This does not require a return to high-pressure closed exams. A brief in-class problem, a short conversation about a submitted assignment, or a few minutes spent explaining their own work all reveal quickly whether the understanding is there. Keep the stakes low and the frequency steady.

4. Model the disclosure you ask for. If you want students to be honest about how they used AI, be honest about how you use it. When you use a tool to draft a problem set or check an example, say so. Disclosure becomes normal when it is shared rather than confessed.

What to expect

The first term will not be tidy. Some students will lean on the tools too heavily, and their unaided checks will show it, which is exactly the signal you added those checks to get. Others will barely use AI and may need encouragement to practice with it, because using these tools well is itself a skill they will need. Treat the first run as calibration. You are learning where your course’s real core competence lies and where AI genuinely helps, and you will adjust.

Where to go next

You can point your students at the materials in this repository directly. students/start-here.md orients a student new to using AI for coursework, and students/from-answers-to-learning.md is written for the student who has been using AI to get answers and wants to switch to using it to learn. The students/study-sessions/ folder has four worked examples, in statistics, coding, physics, and mathematics, each showing a student using AI to study while keeping the thinking and the verification their own. These make good in-class illustrations and good assigned reading.

The rest of the instructor materials are here now. instructors/what-to-assign.md is a one-page schedule of what to hand students and when, and it is the fastest way in. instructors/briefings.md has three ready-to-drop syllabus AI policies, at a restrictive, a default, and a permissive setting. instructors/semester-redesign.md works a single course through a full term. instructors/assignment-templates.md gives reusable assignment patterns that build in verification and disclosure. instructors/assessment-models.md covers assessing genuine competence when class size is the binding constraint, and instructors/rubrics/ has the worked rubric. instructors/faculty-faq.md answers the questions instructors raise most often.