Arizona State University Office of the Provost + EdPlus
← Back to overview

The full resource

Nine practical strategies for assessment design in the age of AI

Each one comes from real faculty who have published materials, working models, and examples you might consider adapting for your course and the ASU environment. You do not need all nine. One move from a few different clusters often produces the largest change with the least redesign effort.

Jump to a section
Frameworks: TILT & Two-Lane Where should I start? 1. Per-assignment AI policy 2. Student intros 3. Interaction per module 4. Micro-orals 5. Local context 6. Prior work 7. Process & feedback 8. Real audience 9. Assurance moment Get involved

Two foundational frameworks

The frameworks that anchor these strategies

The strategies fit into two frameworks that come from established faculty practice. Neither framework is required for adopting any single strategy, but holding both in mind might sharpen decisions about where to invest redesign effort. Adopting even one framework at the course level often makes the tactical strategies easier to prioritize.

Explicit expectations Presence & relationship Observed cognition Task design vs. AI substitution Process & feedback Lane 1 controlled assurance
TILT

Transparency in Learning and Teaching

Developed by Dr. Mary-Ann Winkelmes, TILT is a small, high-leverage change to how assignments are written. Many students, especially those first in their family to attend college, struggle not because the work is too hard but because the unwritten rules of the assignment are unclear. TILT fixes that by making three things explicit on every assignment. It is deliberately lightweight: transparency, not more work.

  • Purpose — why am I doing this? The skills students practice, the knowledge they gain, and how it connects to real life and course outcomes.
  • Task — what exactly do I do? The steps to follow, what to do and what to avoid, and any productive-struggle steps left open on purpose.
  • Criteria — how is it judged? A checklist or rubric, what a successful piece looks like, and annotated examples of real work.

An AAC&U pilot (2014–15) across 7 minority-serving institutions, 1,180 students and 35 faculty, found that applying TILT to just two assignments produced statistically significant gains of medium-to-large magnitude in academic confidence, sense of belonging, and skill mastery, with the largest gains for first-generation, low-income, and underserved students (Winkelmes et al., Peer Review, AAC&U, 2016).

How to apply it
  • Design assessments using the three headings. Structure each assignment as Purpose / Task / Criteria, often just reorganizing what already exists.
  • Name the real-world relevance. State the durable skill the task builds and how it links to a course outcome, so the work feels worth doing.
  • Add annotated exemplars. Show one or two samples of successful work with notes on why they succeed.
  • Start small. The research shows the effect from TILTing as few as two assignments; pick the highest-stakes ones first.
Quick self-check before publishing an assignment
  • Could a first-generation student, reading only this prompt, explain why the assignment matters?
  • Are the steps and the “what to avoid” spelled out?
  • Is there a rubric or checklist, and at least one example of success?
Pairs with the two-lane model. Once a task sits in the open (AI-permitted) lane, TILT is where you make the AI rules explicit: the Task section states exactly how students may use AI, and the Criteria reward their judgment over the raw AI output.
Where to look as you think about this

Sources: TILT Higher Ed (tilthighered.com). Winkelmes, M-A. et al., “A Teaching Intervention that Increases Underserved College Students’ Success,” Peer Review 18(1/2), AAC&U, 2016.

Two-Lane

Thinking about two purposes for assessment

Why we’re sharing it. We think this model offers a promising way to conceptualize assessments within a course and within a program. It is important to the workshop and will organize much of our work together. But we also want to know what you think of it: would you refine it? Revise it? Reject it? In short, we are asking that you work with this model, but not that you do so uncritically.

The two-lane model (University of Sydney; Danny Liu and Adam Bridgeman, 2024–2025) starts from a simple premise: assessment now has to do two jobs at once. It must assure, in the accreditation sense of providing trustworthy evidence, that students actually achieved the learning outcomes, and it must help students learn to work productively with AI. One assessment type cannot do both jobs well, so the model sorts every assessment into one of two lanes. Lane 1 assessments are secure and supervised; they may restrict or prohibit AI use, though a ban is not required. Lane 2 assessments are open, allowing instructors to assess a course’s learning outcomes as students will actually encounter them, in a world where AI tools are available.

Our context: online and asynchronous. Lane 1 is defined by controlled conditions, not by a physical room, so the model fits async online directly. Online, we achieve “controlled” through course and platform design. The specific methods are an area our team should continue to refine to meet our learners’ needs and our model.
Lane 1 — Controlled (Assessment of Learning). The conditions are controlled enough to assure the work reflects this student’s own capability. The distinction is about conditions and purpose, not location.
  • AI may be banned, restricted, or deliberately permitted in specified ways, depending on the learning outcome being assured.
  • Options include identity-verified time-boxed tasks, authenticated performance work, and other methods, an area our team should continue to refine.
  • Purpose: a defensible signal of genuine achievement.
Lane 2 — Open (Assessment for and as Learning). Uncontrolled conditions, where AI use is permitted and even encouraged. This lane still assesses the course’s own learning outcomes. Students demonstrate their skills and knowledge in the environment where they will actually use them, while also developing AI literacy: evaluating, refining, and taking responsibility for AI output rather than simply submitting it.
  • Authentic, process-focused, real-world tasks, the natural home of most async coursework.
  • Guide AI use as a menu, not a traffic light: don’t micro-restrict; steer students toward the choices better for learning.
  • Process artifacts (drafts, AI chat logs, reflections) make the thinking visible.
  • Purpose: develop and assess students’ skills and knowledge in an AI-rich environment, building AI-capable graduates along the way.

The key move: assign lanes at the assessment level. Lane assignment is a course-level, assessment-by-assessment decision guided by each task’s purpose (program-level coordination is optional). Put controlled assessment only where an outcome genuinely needs to be assured, and let most of the course live in the open lane. A single task can combine both, for example an open project capped by a short controlled check. Lane 1 options are an area our team should continue to refine to meet our learners’ needs and our model.

Pairs with TILT. The two-lane model decides which lane a task lives in; TILT makes the rules of that lane explicit to students, especially the open lane, where “how may I use AI here, and how will it be graded?” must be stated plainly.
How to apply it in an async course
  • Consider each assessment on its own. Does the learning outcome call for controlled conditions, or would students benefit from working with AI? Label the lane for students.
  • Choose the control method deliberately. Where a task is controlled, work with the design team to select a method that fits; the method is deliberately left open here and should be refined over time.
  • Make the open lane AI-authentic. If AI can complete the task in one shot with no learning, redesign it around the student’s judgment and critique of AI, and collect the process, not just the product.
  • Focus on developing AI literacy rather than relying primarily on detection. Lead with clear expectations and AI-integrated design, and state AI-use rules plainly in every brief.
Common misreads to avoid
  • Lane 1 is not “in person” and does not have to be “no AI.” It means controlled conditions to assure authorship; AI may be restricted or purposefully permitted within them.
  • Lane 2 is not “unsecured” or a free-for-all. Open tasks are deliberately designed for critical, responsible AI use.
  • Consider where controlled conditions are truly necessary. Controlling every task raises cost and anxiety and crowds out authentic learning; aim for proportion, not maximum surveillance.
Where to look as you think about this

Sources: Liu & Bridgeman, “The two-lane approach to assessment in the age of AI,” Teaching@Sydney (FAQ July 2024; explainer March 2025); University of Bath Two-Lane FAQ (2026); University of Auckland TeachWell “Two-Lane Approach to assessment” (2026). Online/async adaptation by EDL for the Assessment Design Accelerator.

How TILT and Two-Lane work together

Two-Lane decides which lane a task lives in, controlled (Lane 1) or open (Lane 2). TILT makes the rules of that lane explicit to students, especially the open lane, where “how may I use AI here, and how will it be graded?” must be stated plainly. With those decisions in place, the nine strategies below offer practical ways to redesign individual assessments.

Where should I start?

Start with the teaching challenge that fits your course

You do not need to use all nine strategies. Pick the challenge that feels most relevant and jump to a strategy that helps:

Not sure where to begin? Choose one strategy and try it with one assessment you already use. Combining one move from a few different clusters often produces the largest change with the least redesign effort.

1
Explicit expectations

Set an explicit AI-use policy for each assignment, not just the syllabus

Why it works

Blanket “AI is prohibited” or “AI is allowed” statements at the syllabus level don’t help students make good choices on specific tasks. Per-assignment guidance treats students as capable of ethical judgment, removes the ambiguity that leads to quiet decisions about AI use, and keeps the rules for one course from bleeding into another. Faculty are converging on a shared vocabulary: the AI Assessment Scale (AIAS), Level 1 (No AI) through Level 5 (Full AI), or a simpler Stoplight (Red / Yellow / Green).

How to do it
  • Choose a simple approach: the five-level AIAS, or the simpler Red / Yellow / Green Stoplight for entry-level adoption.
  • On each assignment, add one sentence: “AI use on this task: [level]. Reasoning: [1 sentence].”
  • Ask students to disclose AI use in a brief note at the end of the submission, regardless of level.
Where to look
2
Presence & relationship

Open the term with short, unscripted student video or audio introductions

Why it works

A 60–90 second unscripted intro establishes each student’s voice, identity, and presence for the instructor and their peers. Hearing students early gives you a reference point for their later work, and students who introduce themselves early often engage differently with the course going forward. It signals from the start that this is a course where the student is the primary voice.

How to do it
  • Use your LMS’s built-in recorder (Canvas Studio, Panopto, Kaltura all support this).
  • Prompt: “Introduce yourself in 60 to 90 seconds: your name, one thing you hope to learn, and one thing about your context that shapes why this course matters.” No slides, no reading.
  • Grade for completion, not polish.
Where to look
3
Presence & relationship

Build in a brief synchronous or semi-synchronous interaction each module

Why it works

A brief 1:1 or small-group interaction per module, even 10 minutes, builds relationship and belonging and gives you a regular chance to hear students think. That presence is what keeps students engaged, and it tends to reduce corner-cutting: students are less likely to cut corners in a course where they are known to their instructor and peers. When synchronous doesn’t scale, semi-synchronous alternatives (recorded video responses, threaded voice discussions, structured office-hour sign-ups) achieve much of the same effect.

How to do it
  • Small sections: offer 10-minute 1:1 video check-ins at one or more points in the term.
  • Larger sections: use peer groups of 3 or 4 with structured prompts and, if available, rotating facilitators.
  • Semi-synchronous: students record 2-minute reflections on a weekly prompt; peers respond via video.
Where to look
4
Observed cognition

Embed brief (1–2 minute) low-stakes oral concept explanations within quizzes or activities

Why it works

A short “micro-oral,” a one- to two-minute recorded explanation, asks students to talk through their reasoning on a concept or answer choice. It puts the focus on how a student thinks, not just the answer, and it builds the everyday skill of explaining your reasoning clearly. Students report that explaining aloud sharpens their own thinking, and because the reasoning is in the student’s own voice, it gives you authentic evidence of their understanding. It also scales more readily than a full oral exam.

How to do it
  • Add a “record your explanation” prompt to 3 to 5 quiz questions per module. Use Canvas Studio or a video-response feature.
  • Score for evidence of reasoning: did the student explain why, not just repeat the answer? Two or three criteria is enough.
  • Weight low (5–10% of the grade) so students engage without high stakes.
Where to look
5
Task design that resists AI substitution

Design assignments that require local, personal, or course-specific context

Why it works

Assignments anchored in a student’s own context (their workplace, their community, a real case they are close to) push students to apply course ideas to a situation they actually know, which deepens learning and makes their thinking visible. That same specificity is hard to outsource: AI can produce a generic answer, but not the details of a particular workplace situation or lived experience, so it shifts AI from a substitute into a supporting tool. Read alongside Strategy 6, which draws context from inside the course.

How to do it
  • Instead of “Analyze a company’s marketing strategy,” try: “Analyze the marketing of a company you have personally interacted with in the past month. Include specific details as evidence.”
  • Draw on a real setting: “Take a decision your organization made in the last 90 days and evaluate it against our framework.”
  • Give students choice of case: they identify the situation and defend its relevance.
Where to look
6
Task design that resists AI substitution

Personalize assignments by referencing students’ own prior work or contributions

Why it works

This builds a coherent arc across the term: students see their own thinking develop over time, which strengthens intrinsic motivation and ownership of their work. Where Strategy 5 pulls context from outside the course, this pulls it from inside. A prompt that references what this student wrote earlier, or what a named classmate contributed in Week 3, is one AI can help with only if the student actively brings that course-internal material in.

How to do it
  • Design a mid-term or final that requires building on, extending, or challenging the student’s own earlier submission.
  • Reference class discussions: “In Week 3 we distinguished X from Y. Apply that to a case from your own experience.”
  • Make peer review substantive: “Identify the strongest and weakest point in [named classmate’s] project, and defend your evaluation with specifics.”
  • On discussion boards, require students to cite and respond to two named classmates’ posts.
Where to look
7
Process & feedback

Make how students develop their work, and respond to feedback, part of what they submit

Why it works

Revising work in response to feedback is one of the most reliable ways to deepen learning, and it is something a student has to do for themselves. When you ask students to show how their work developed (early drafts, the decisions they made, how they responded to your comments), the assignment centers that development rather than only the final product, and you get formative evidence to act on before the final grade. AI can produce a polished final draft in seconds; it cannot easily produce a coherent account of how a particular student got there.

How to do it
  • Break one major assignment into milestones (proposal → draft → final), with specific feedback at each stage.
  • Require intermediate steps: an outline before the draft, a short decision log alongside a project.
  • Require a response memo: “Here’s how I addressed each piece of feedback,” one paragraph per point.
  • Ask for a 200-word process reflection on what changed between draft and final, and grade it alongside the artifact.
Where to look
8
Task design that resists AI substitution

Design an assignment for a real audience beyond the instructor

Why it works

When students write for a real audience (a community, a workplace, a client, a public forum), the definition of “good work” becomes concrete and the stakes become meaningful. Real-audience assignments also tend to require context and voice AI cannot easily fake. As Torrey Trust frames it: “AI can do this. Why am I asking them to do this?” If the answer is “because a real person needs it,” the assignment justifies itself.

How to do it
  • Partner with a campus or community organization: students produce something the partner will actually use.
  • Aim work at a specific public forum: a blog post, a policy brief, a community presentation, a portfolio piece.
  • Give students choice of audience: they identify the recipient and defend the fit.
Where to look
9
Lane 1 controlled assurance

Include a controlled (Lane 1) assessment where students demonstrate what they know and can do

Why it works

This comes directly from the Two-Lane model. If a course grants a credential, at least one assessment needs to happen under conditions controlled enough to assure the work reflects this student’s own capability. Controlled does not mean AI-banned, and it does not mean in-person. It means conditions designed to give you confidence that the work reflects what the student knows and can do. Once that assurance moment exists, the rest of the course can live comfortably in the open lane where real learning and AI fluency develop.

Why it matters at ASU. This provides assurance of learning in the accreditation sense without turning the course into a constant contest over AI use. It gives a defensible answer to “how do you know?” without asking every task to bear that weight. Because “controlled” is about conditions and not location, the model fits async online directly, and the specific methods are an area our design team is continuing to refine.

How to do it
  • Ask, for each assessment: which learning outcomes require students to demonstrate their own capability, and what conditions would give you confidence in that evidence? Consider at least one controlled (Lane 1) assessment; more is fine when the outcomes warrant it.
  • Choose the control method deliberately: identity-verified time-boxed tasks, authenticated performance work (oral defense, live demonstration, synchronous check-in), or another approach that fits your learning outcome, needs, and course context, in partnership with your instructional designer.
  • Decide what AI role, if any, is appropriate within the controlled conditions. Controlled and “AI-permitted-in-specified-ways” can coexist.
  • Communicate the Lane 1 assessment early: tell students why it matters, what it will look like, and how they can prepare.
Where to look

Get involved

The Office of the Provost and EdPlus are actively developing this work with faculty. If you want to join a workshop, contribute to resource development, sit in on a listening session, or simply hear more:

  • Explore the frameworks, strategies, and Integrity Through Design guide on this site.
  • Tell us who you are, what you teach, and what you want help with.
Express your interest →