In the last week, I've spoken with 23 instructors about assessment, and a trend emerged. Assessments are now taking much more time to manage than they used to; instructors are shifting back to in-class quizzing, grading manually, redesigning assessments, building new ones, and setting up and managing proctoring tools.
So, AI has made managing assessments more time-consuming, but I firmly believe the same technology can also help instructors:
Redesign assessments so they hold up when students use AI
Cut the time it takes to build and update quiz questions
Save time on grading
Why am I so confident about this? Well, because it’s already happening. In a short poll on LinkedIn last week, I asked instructors if they use AI to build or grade assessments. 80% of people replied 'Yes', and around ⅔ of the instructors I spoke with in the last week confirmed that they’re doing so. Take that number with a pinch of salt: the BT audience leans more towards tech-savvy instructors. I expect the real number is a little lower.
So, how is AI being used?
Here's the pattern that stood out: the most common uses are also the safest. So think of what follows as a swimming pool, running from the shallow end, where anyone can paddle in safely, to the deep end, where you need to know what you're doing before you dive in.
The shallow end: AI as a design partner for updating assessments. Take an existing assignment and use Claude or ChatGPT to rework the assignment so that it better fits how students might use AI.
Our First Week Assessment Checklist gives you a framework to work from. It sorts assessments into three lanes: Visible (you assess the process, so AI use shows), Resistant (tasks AI still fails at reliably), and Required (students must use AI, and you grade their critique). To put it to work:
1. Upload the checklist to Claude or ChatGPT so it understands the framework.
2. Add the assessment you want to update and ask which lane it fits best.
3. Tell it what you're trying to achieve, then iterate.
With some thoughtful iteration, an outdated assessment can become fit for purpose in a fraction of the time. When Bellevue University brought AI into its course design process, the time to build a course dropped from 16 weeks to 8, and that included creating the assessments. One thing to watch: AI doesn't know your course or your students, so treat its redesign as a first draft and check the assessment still measures what you need it to.
Wading in: Creating or updating quiz question banks is the one that came up second most frequently, though there’s more to consider. Upload a set of questions, ask Claude to change the numbers, create dynamic versions in the format your LMS uses, or ask it to create more questions in the same style. The time savings can be dramatic. In one study, ChatGPT wrote 50 medical exam questions in about 20 minutes, while two experienced examiners needed around three and a half hours between them, and blinded reviewers rated the two sets similarly overall. Even with time set aside to audit every question, that's a big head start.
We use skills that run to over 2,000 words, and a multi-step process for creating questions in the BT platform. You can create a solid set of questions with less prompt engineering (this UMich guide is a good starting point), but be sure to check for errors. In another study of ChatGPT-generated practice questions, 49% of questions had item-writing flaws, and 22% had factual or conceptual errors, so even good prompt writing doesn’t remove the need for a human expert to audit the output!
The deep end: Assisting with grading is significantly less commonly used, but shows a lot of potential for saving time. Upload students’ work to an LLM, along with a rubric, and use the AI tool to grade the work. One Northeastern University professor reported cutting his grading time by 60% after setting up an AI grading tool, and estimated it could save him eight hours per exam.
Before giving this one a try, a word of caution: AI is increasingly being used to grade students’ work, but FERPA rules are strict, and at times, hard to navigate. Anthropic’s August 2025 educator report stated that about 7% of educator chats were about assessing students. Of the grading chats, 48.9% handed the work over to the AI in a way the researchers called “concerning”. If you’re doing this, or planning on doing it, check your university policies, and never upload any student data into an LLM without being absolutely sure that it’s being done in a way that’s FERPA compliant.
So, if these are the common ways AI is used today, what’s next? Following the Financial Education Association conference last month, my bet is on creating end-to-end assessment tools that deliver the assessment experience and the grading. We’re already seeing study tools built with AI, for example, Frank Strobel at The University of Birmingham used Claude to build two in-class games for his Advanced Financial Markets & Institutions class, but stopped short of using AI to assess students.
So, why aren’t more instructors taking the leap, and building assessment tools yet? The big reason is data privacy. Uploading any student data into an LLM is risky. Unless you’re completely clear with your university policies on data protection, and have access to an approved tool, I recommend holding off on building anything that incorporates input from students that will be graded.
So, can AI give back the time it took? Add up the numbers in these examples: half the time to build a course, 20 minutes instead of three and a half hours to write a set of exam questions, and one professor cutting his grading time by 60%. Even with time set aside for checking, that's a lot of hours back!
If you're not yet using AI to make managing assessments easier, start in the shallow end: pick one assessment and redesign it this week. When you're ready to wade a little deeper, try updating a quiz bank, but plan in time to audit the questions. If you're thinking about grading, check your university's data policies first, and proceed with caution.
Sources
Bellevue University course build: Anthology case study, 2025. https://www.anthology.com/sites/default/files/2025-04/BellevueUniversity_CaseStudy_EN_v3.pdf
ChatGPT vs human exam questions: Cheung et al., PLOS ONE, 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC10464959/
Item-writing flaws in AI-generated questions: Camarata et al., Advances in Physiology Education, 2025. https://doi.org/10.1152/advan.00106.2024
Northeastern grading time: "AI meets the grade: Faculty reflections on Essaybot in the classroom", D'Amore-McKim School of Business, 2025. https://damore-mckim.northeastern.edu/news/three-damore-mckim-faculty-pilot-essaybot-to-rethink-grading-with-ai
Educator grading conversations: Anthropic Education Report: How educators use Claude, August 2025. https://www.anthropic.com/news/anthropic-education-report-how-educators-use-claude