Restrict AI, or embrace it?

Why not both?

When the goal is assessing foundational knowledge — the concepts, methods, and fluency students need to have internalized — allowing AI use will mask whether learning happened at all. When the goal is applied skill, AI can circumvent learning, but the result is often uninspired boilerplate. Capable students can bring high-level thinking to bear on AI output to create work that truly stands out. They need to exercise judgment about what to ask, polish the material they choose to keep, and have enough grasp of the material to know what to throw out. 

Think of assessment like a barbell. Load one end with secured, in-room work where foundational knowledge gets demonstrated without AI. Load the other with ambitious, AI-permitted work held to a higher standard than you would have set three years ago. Keep the middle mostly empty.

Some of the recommendations below will be familiar and continue to work well with this model. Others will be new ideas you'll explore and test, as well as some previously reliable options which, sadly, you'll need to avoid.

1. Closed-book, in-person exams

In-person exams that exclude the use of electronic devices will not need to change as students cannot use AI tools to take them. However, students might use generative AI as a study partner to prepare for an in-person exam in the same way they might study with a classmate.

2. Proctoring when a computer is required

Proctoring software is the most practical way to minimize unauthorized use of AI during exams that require a computer. Proctoring cannot completely eliminate the use of AI on exams, but it can limit it substantially. Two options are available at NYU.

An AI-powered assignment is one where a student produces better work by pairing their knowledge with AI's advanced capabilities, rather than thoughtlessly handing off work.

For students to be able to successfully do this, they need enough command of the material to direct and evaluate any AI-assisted work, and the task has to be ambitious enough that a simple AI prompt won't generate high-quality results without a student providing context, judgment, and revision.

2. Raised expectations

Here's an opportunity worth putting to students directly: everyone has the same tools now, which means the work that stands out is the work someone made better. Employers are hiring the judgment to direct powerful tools, and the expertise to catch what they get wrong. Students who develop that have something genuinely scarce in the marketplace.

Students will calibrate their efforts to the bar you set. Here are some practical ways to set the bar high.

• Ask for the work you couldn't ask for before. The investment analysis that once took three weeks of data gathering can now support comparison across several firms, sensitivity testing against different assumptions, or a counterargument the student has to answer. The effort hasn't disappeared — it's moved from collection to judgment.

• Evaluate judgment and quality of reasoning over polish and volume of output.

• Show good examples. While exemplars may have been held back in the past to discourage copying, every student already has a head start on a pre-fabricated first draft. When students can clearly see a specific, high standard, they aim at it, and the focus can shift from what's required to what's possible.

• Talk about the work as professional practice. Framing the deliverable as something a real analyst would be expected to produce sets expectations more effectively than any point deduction.

• Be willing to grade to the standard you set. This part is hard, but a high bar communicated in the syllabus and not reflected in grades isn't really a high bar. Reserving excellent marks for truly excellent work over and above what AI can achieve on its own is the key here. Generic, uncritical analysis that does not directly grapple with the key concepts outlined in the course is the telltale sign that AI did the heavy lifting and students were just along for the ride.

Make sure to have a conversation with students during class about why AI is permitted for these assignments and the higher expectations that come with it so the standard is visible from the beginning.

3. Live presentations and Q&A for AI-assisted work

Even when students are able to leverage AI for complex and professional-quality work, a live presentation that includes Q&A requires students to understand the material well enough to defend it without reading from a script. A robust Q&A session helps to ensure students have correctly processed and applied their judgment to any work they have done in collaboration with AI tools.

• Ask students to apply the material rather than restate it: what would change if a key assumption were different, how this framework handles a case they haven't seen.

• Follow up on the first answer. The second question is usually where preparation ends and understanding starts.

• Bring in classmates. Peer questions are harder to script for, and answering them is its own skill.

Keep in mind: Presentation time scales with enrollment. For larger sections, consider team presentations, shorter formats with a strict clock, or a rotation where a subset presents each week while the rest serve as the questioning panel.

1. Take-home exams

An unproctored take-home exam doesn't test whether students know something. It tests how efficiently they can look things up, or whether they can hand the whole thing to a model. Either way, the score you get back doesn't answer the question you asked it.

What to do instead. If you want to test foundational knowledge, secure the conditions — proctor the exam, or run it in class on paper. If you want to see what students can do with AI, don't half-permit it: design an AI-amplified assignment and raise the bar to match.

2. Dictating AI use outside of class

Use AI for research but not drafting. Use this model, not that one. Outline with it, then write on your own. These feel like reasonable middle paths, but they don't hold up outside the classroom. Students will typically take the most efficient route to finishing an assignment, and loose guidance on how to use AI mostly goes unfollowed, which means the rule ends up applying only to the students who were most conscientious to begin with.

If you want students to use AI in a specific way, make it an in-class activity where you can actually facilitate it. If you want AI out of the picture, use a proctored or device-free environment.

3. Simple, generic essay prompts

Standard essays, short case write-ups, and simple reflection assignments are, for a large share of students, now done by AI. The prompts aren't specific enough to require anything the student uniquely brings, and the output is competent enough to pass.

What to do instead. Essays still work in the room — in-class writing gives you the same thinking under conditions you control. They can also work as take-home assignments when the prompt is genuinely personal and the course carries high intrinsic motivation, which is more often true in electives than in required courses. Outside those cases, the fix is specificity: a real firm, a decision with real tradeoffs, an audience who will respond.