Making Assignments AI-Resilient

You cannot control whether students use AI on work done outside of class, but you can design your course and assessments so that students know when and what kind of AI use will be detrimental to their learning and succeeding in your course.  They’ll then have incentives to eschew AI where you deem it inappropriate  and be rewarded for going through the productive struggle of learning.

Students want to learn. What often pulls them toward substituting AI use for learning is not indifference, but instead time pressure, a desire to optimize performance, and concerns about fairness. You can support your students in three ways.

• Guidance. Make your learning objectives clear and, for each assignment, help students see how and where substituting AI use for individual work will interfere with success in your course.
• Fairness. Ensure that a student who follows the policy is not disadvantaged relative to one who does not. If an AI-generated submission can earn the same or better grade as a flawed but individually-produced submission, students will feel pressure to use AI where you have told them not to.
• Reward. If you have significant, graded in-class assessments that depend on students productively struggling with work done outside of class, that will reward students who follow the AI policy and incentivize all students to do so.

Your AI policy should be easy to understand, implementable, and enforceable. A course-level policy can state the overall policy with respect to AI use, but for the course policy to be clear and implementable, you will also need specific guidelines for each assignment.

Although it might be tempting to enact a total AI ban and enforce this policy by using AI-detection software, we urge caution. This software can provide a useful indicator when student AI use is naïve, but it is not sufficiently accurate to be relied upon to detect all cases. It produces both false positives and false negatives. Informing students that you intend to use AI-detection software can also create a detection and evasion dynamic that takes the focus off of the learning goals of the course. If you use AI-detection software, we encourage you to also implement the resilience strategies suggested below.

Five AI-resilience strategies

We recommend five basic strategies for making assignments resilient. How these strategies are implemented can vary widely across courses and assignments. The Guidance by assignment type section breaks down how to apply these strategies to make the most common assignment types more AI-resilient. 

1. Move assessment of central skills into the classroom. Assess in class, on paper or orally, the skills you previously assessed only through take-home work.
2. Pair take-home work with impromptu in-person understanding checks. The checks can be short and low-stakes, and occur after submission of take-home work. Ask students to explain, extend, or reflect on what they handed in as a check of the depth of their understanding of their own submitted work.
3. Scaffold large assignments with staged, in-person checkpoints. Employ a workshopping model in which students produce some combination of artifacts (such as proposals, outlines, source reviews, drafts, revisions) each with a live touchpoint (for more on touchpoints, see below).
4. Anchor assignments in material not found in LLM training data. This might include students working with objects from the Harvard collections (libraries, museums, etc.), conducting interviews, or drawing on materials produced or collected during class/section discussions.
5. Provide structure and social accountability options for device-free or AI-free work. Offer places and times, ideally with instructional support resources (and snacks!) for students to work without recourse to AI. Providing this structure can help students enact the behaviors they aspire to. If they are not required, these times can’t be used for verification, but they can still be useful. 

Although we have broken these out into five strategies, they all rest on in-person touchpoints that relate to work done outside of class. They incentivize students to be prepared to demonstrate their mastery of the learning objectives without assistance from AI. A student may or may not use AI during their preparation for an in-person touchpoint, but, either way, their out-of-class work will be most effective if it is directed toward performing well in person. For long-form work that is scaffolded, the frequency of touchpoints makes AI shortcuts less productive relative to engaging directly in the learning activity. Similarly, anchoring assignments in students’ personal experience of Harvard collections, interviews, or other materials helps make AI shortcuts less productive. 

These strategies require time and work to apply. The Proctoring, collecting, and grading in-class work section provides some advice on how to manage it without undue burden on you and your teaching team. 

What to ask at a touchpoint

At your in-person touchpoints, ask for a meaningful demonstration of the learning objectives in pedagogically valuable activities. Do not ask the student to recite what they submitted or quiz them in ways meant to catch AI use rather than measure learning. Good questions at a touchpoint tend to fall into a few categories.

• Reflection: identify the assumption, source, or definition your work depends on most. What breaks without it?
• Defense of a choice: identify a choice you made that a reasonable person could have made differently. What was the alternative, and why did you reject it?
• Extension: how does your argument change if we remove or alter this assumption or add this new piece of evidence?
• Transfer: apply the approach you used in your work to a new scenario or problem.

Use prompts that cannot be answered well without genuine understanding of the submitted work as well as prompts that contribute to student learning. Prompts that ask only for recall can be prepared for in advance, sometimes with the help of AI, and they emphasize verification over learning. The prospect of a reflection or extension prompt changes how students approach the original assignment, because the best preparation is to do the thinking the assignment was designed to elicit.