Designing Courses & Assignments in the Age of AI

Overview

The vast majority of undergraduate students at Harvard College are using generative AI. Many are using it for help on their academic work, often regardless of stated course policies. A 2024 survey of Harvard undergraduates found that 85% use AI in some way at least biweekly, and over 50% rely on it specifically for writing assignments. National data from 2025 shows even higher rates. These tools are not just pervasive; their capabilities can lead to a range of issues for instructors as they give feedback on student work and, by extension, assess student learning. 

Assignments give students a framework to learn new skills and to practice applying them. Student work becomes the evidence that instructors use to evaluate how much their students are learning and provide feedback on where they are succeeding, along with where they need more practice or support. The quality of this evaluation and feedback depends on the degree to which submitted work actually reflects the student’s own abilities. 

For many kinds of familiar assignments, such as response papers and p-sets, the challenge of getting good evidence is two-fold: 

  1. Generative AI has already reached the point that today's large-language models can produce fluent academic writing, generate runnable code, and solve textbook-style problems with surprising accuracy.
  2. Attempts to detect unpermitted AI use are largely unreliable, producing both false positives and false negatives. 

Rather than trying to identify or police AI use, ensuring that student work is reliable evidence of their learning requires us to rethink assignment design and assessment. Specifically, we must identify what kinds of assignments are most at risk of producing unreliable evidence of student learning and which ones more effectively produce—or can be adapted to produce—reliable evidence. Doing so can allow instructors both to minimize the advantage of unpermitted AI use and thoughtfully incorporate AI into the learning process. 

Assignment Types Susceptible to AI Use

The following types of assignments present a high risk of being completed fluently by AI without detection:

  • take-home short response papers and essays
  • take-home p-sets
  • take-home exams

Assignments in this category are often important for student learning and they can still be effective. One way to make them more AI-resilient is to adapt them so that students demonstrate their process as well as the final product. Adapting an assignment so that students demonstrate their process or their understanding can be done by incorporating post-submission understanding checks, presentations, or reflection questions administered in-person. 

When retaining a large take-home project as the capstone (such as a final essay), it is a good plan to “scaffold” it by breaking it into steps. Making at least some of these steps in person without devices (oral topic proposals, in-class outlines, reflective oral or hand-written explanations after submission, follow-up oral exam) offers a better chance of reliable assessment. This might include touchpoints AFTER submission, such as a live interview about the project or an oral defense.

When allowing Gen AI use in a course overall, it’s best to ensure that at least some assignments are done without it. The mental model for students should be that AI is helping them learn—to more deeply internalize—the subject matter of the course, so that they can then demonstrate knowledge of that subject matter on their own without AI. For this to happen, instructors need to include graded assessments that determine whether students are adopting this mental model. In turn, these assessments can help incentivize them to do so. 

Assignment Types Less Susceptible to AI-Use

Methods and modes of assessment likely to be effective with or without incorporating AI:

  • In-person Blue Book and oral exams can measure recall and applied understanding independent of outside technology.
  • Alternative assignment modalities such as oral exams, in-person presentations, video essays, posters and infographics, “visual abstracts” of scientific papers, on-paper annotation of p-sets and printed code in class on the day of submission, and many more are likely to be more AI-resilient than traditional text-only assignments.
  • Traditional assignments like essays and problem sets when coupled with in-person comprehension checks, presentations, discussions, or other adaptations. 

Implementing AI in Assignment Design and Assessment

Some concrete examples of how AI can be incorporated into the “process” and “product” stages of design and assessment include: 

  • Brainstorming partner. Have students use GAI as a sparring partner to brainstorm ideas and then require them to critique the output for bias and accuracy.
  • Transparent AI workflow logs. Require students to share prompts, responses, and a short rationale describing what they kept, modified, or discarded—and why.
  • Fact checking a model. Have students fact-check an AI-generated essay, or ask them to improve upon an AI-generated piece of code, documenting their changes and the reasoning behind them.
  • Model comparison memos. Have students query two models (or settings) and write a brief memo on differences in accuracy, bias, or style, citing course concepts.
  • AI-to-human handoff. Let AI produce a first pass (outline, test suite, or code comments). Students then complete the “last mile,” justifying design choices and revisions.
  • Source-anchored critique. Provide course readings/lectures; students must use them to audit AI claims, with citations and corrections.
  • Timed “no-AI” checkpoints. Pair AI-assisted preparation with short, in-class demonstrations (whiteboard proofs, oral mini-vivas, live coding) to confirm mastery.

Key Takeaways

  • Have an AI policy and discuss it with your students.
  • Use a range of assignment types, some of which are not susceptible to AI-use.
  • Consider swapping high-risk assignments for other options
  • Connect outside-of-class assignments with in-person, AI-proof (or highly resistant) check-ins, steps, and evaluations.
  • Start small. Pilot one AI-integrated activity, then build out from there based on what you learn.
  • Where you permit AI, require transparency (logs/screenshots or prompt sheets) and assess individual learning separately.
  • Contact the Bok Center for additional strategies.