tl;dr: Before you redesign your assessments, update your learning outcomes.

Since at least Mager’s work in the early 1960s, instructional designers have recognized that learning outcomes should include three kinds of information. They need details about the performance (the thing the student will do), conditions (the tools, constraints, and environment in which they’ll do it), and criteria (how well they need to do it). For example:

Learning outcome: Using a digital multimeter, an unpowered breadboard circuit, and a schematic diagram, the learner will locate and identify two intentional open-circuit faults within 8 minutes.

Performance: Locate and identify two intentional open-circuit faults

Conditions: Using a digital multimeter, an unpowered breadboard circuit, and a schematic diagram

Criterion: Both faults identified within 8 minutes

Pause for a moment and imagine how you might assess this learning outcome. It doesn’t take a lot of imagination, right? When learning outcomes are well written, assessment strategies flow directly and easily from them.

Over the years, many people have gotten away from explicitly stating conditions and criteria in their learning outcomes. If you saw the outcome above in a modern class, it would probably say “the learner will locate and identify open-circuit faults”, leaving the conditions and criteria implied. But the emergence of generative AI has brought “conditions” back to popularity. Suddenly, the condition “without using generative AI” is everywhere.

While people think “without using generative AI” is a constraint imposed on assessments to prevent cheating, it’s actually a condition added to a learning outcome. If you specifically want your students to be able to complete a task without relying on generative AI, that’s a change in your learning outcome. And since you want them to learn to complete that task without AI, you’ll naturally test to see if they can. Changes we make to learning outcomes (explicitly adding the “without using generative AI” condition) flow through to become changes in our assessments.

(I deleted a lengthy sidebar here about backward design. You’re welcome.)

Work on the Right Problem

Part of the reason why working to make our assessments “AI-proof” is such frustrating work is because it’s often the wrong work. Assessments flow from learning outcomes. Trying to change our assessments without changing our learning outcomes is like trying to correct defects in part after part as they come off the manufacturing line, instead of fixing the production process further up the line to just make the parts correctly.

Imagine (or remember!) a course from the early 1990s in which students are assigned write a research paper. The learning outcomes include performances like “go to the library,” “use the card catalog,” “locate books in the stacks,” “organize notes on note cards,” etc. Then suddenly the internet arrives on campus. In order to prevent “cheating” a new condition is added to all these old performances - “without using the internet.” Then, after a painful intermediate period, instructors realize that rather than needing a new condition what they really need are “internet-aware” performances. Things like “search the library’s online database” and “organize notes using an online reference manager.” Once the performances are internet-aware, the “without using the internet” condition becomes irrelevant.

Or you might further exercise your imagination (or memory) thinking about an accounting class taught before spreadsheets were invented. Or a drafting class taught before CAD existed. Or a photography class taught before photo editing software existed. Etc. Please notice that in each of these cases people still need deep disciplinary knowledge to accomplish the relevant tasks. And the underlying tasks-to-be-accomplished haven’t changed (e.g., accurately track and categorize all income and spending). But the ways the tasks are performed have changed. And that’s ok.

When it comes to cheating with AI, instead of bolting a condition onto a performance that seemed appropriate ten or twenty or thirty years ago, why not get to the root of the problem and re-evaluate the relevance of the performance itself?

AI-aware learning outcomes naturally lead to AI-aware assessments.

This is what I mean when I say we need to put first things first. When we try use conditions to pretend we still live in the old, familiar performance context (e.g., “do accounting without using a spreadsheet!”), we’re not addressing the real issue. We’re choosing to be the factory worker doomed to spend their time forever reworking each malformed part as it comes off the line. Or to be more specific, we’re choosing to be the instructor doomed to forever suspect and inspect every piece of student work we receive, searching for evidence our students used AI. If that doesn’t sound like the way you want to spend the rest of your career, let’s focus our time and efforts upstream - fixing our learning outcomes.

It’s true that some learning outcomes are completely unaffected by the existence of generative AI. (For example, I was a vocal performance major as an undergrad. There’s not really a way I can use AI to cheat when it’s time for me to stand up and sing my pieces for the end-of-term jury.) But many of our learning outcomes are affected by generative AI, and they’re now out of sync with the world around them. Researchers use the internet. Accountants use spreadsheets. Photographers use photo editing software. And knowledge workers use generative AI.

Taking a hard, honest look at our learning outcomes and updating them as appropriate is likely the most important work we can do both (1) to stop students from “cheating” with generative AI and (2) to effectively prepare them to thrive in the world after school.