Practical Guide
Last Update: October 2026
Primary Topic: Education, Learning, and Careers
Reading time: 9–11 minutes
Audience: Educators · College faculty · Academic leaders · Instructional designers · Students
Focus: Authentic learning · Assessment design · Academic integrity · AI literacy · Human judgment
Source and Research Basis
This article is a summary of the conference paper presented by Alex Evangelista, Ed.D., From Papers to Performance: Designing Evidence-Rich Assessments in the Age of AI, at the Northeastern Educational Research Association Annual Conference, October 2026. The paper is a conceptual synthesis, not an original study of student outcomes. Its five design moves, disciplinary examples, workload considerations, and accessibility discussion provide the foundation for this article
At a glance
For educators trying to understand what students have actually learned, generative artificial intelligence presents a practical challenge: a well-written assignment may no longer tell the whole story. This article explores how instructors can gather more meaningful evidence of understanding without abandoning writing, turning every assignment into an oral examination, or making teaching an exercise in surveillance. Students and academic leaders can also use these ideas to understand why assessment is changing and how AI can support learning without replacing it.
The central question is not simply whether a student used AI. It is whether the assessment gives the instructor credible evidence that the student understands and can use the material appropriately. A useful response is an evidence-rich assessment: combining a finished product with selected opportunities for students to apply knowledge, explain decisions, show part of their process, and reflect on what they learned. The purpose is not to create an assignment that AI can never complete. It is to make student learning credible and more visible.
Why a polished assignment is no longer enough
An essay, presentation, spreadsheet, or case response can demonstrate important abilities. But when AI can produce much of that work, the finished product becomes a less certain indicator of who performed the underlying reasoning. Consider a student who submits an organized paper about a psychological theory. The paper may describe the theory accurately and include persuasive examples. Yet the instructor still needs to know whether the student can recognize when the theory applies, identify its limitations, and explain why one interpretation fits better than another.
This is an assessment-validity problem. In plain language, validity concerns whether the evidence supports the conclusion we want to draw. If a grade is meant to represent a student’s analytical ability, the assessment needs to reveal that student’s analysis, not merely an acceptable answer.
Writing still matters. When writing is the learning goal, students need opportunities to develop arguments, use evidence, revise language, and make rhetorical choices. The change is that a final paper may need to be accompanied by an additional amount of relevant evidence. A brief conversation about a source, a draft checkpoint, or an explanation of a revision can help connect the submitted work to the learner.
Authentic work helps, but does not establish authorship
Authentic assessment asks students to do work that resembles meaningful disciplinary, professional, or community activity. Examples include designing a health resource, interpreting laboratory findings, preparing a client recommendation, or developing a lesson.
These tasks can make learning more relevant. However, a realistic assignment is not automatically protected against outsourcing. AI can also generate professional-looking reports, teaching materials, and recommendations. The distinction is between an authentic-looking product and evidence of the student’s own understanding.
A health-education handout becomes more informative when its creator explains why the language suits the intended audience. A business recommendation becomes more revealing when the student discusses an assumption that could change the decision. A mathematics solution becomes stronger evidence when the student explains why the method works. The artifact matters, but so do the judgments behind and supporting it.
Five ways to make learning visible
The suggested approach proposes five adaptable design moves. They are not in a mandatory sequence, and every assignment does not need to include all five. Their value depends on how well they match the course’s learning goals.
1. Contextualize: Give the work a meaningful setting
A general prompt such as “Write about healthy eating” invites a broad summary. Asking students to develop a nutrition message for first-year college students with limited time and cooking facilities requires more specific choices. Context can come from an audience, case, dataset, professional role, local issue, or disciplinary problem. It gives students something to interpret rather than simply a subject to describe.
Personal experience can sometimes provide useful context, but it should remain optional when disclosure could be sensitive. A fictional case, public example, or instructor-provided scenario can serve the same educational purpose. Personalization is also not proof of authorship. The stronger question is why the selected context matters and how it influenced the student’s decisions.
2. Apply: Ask students to use what they know
Knowing a definition is different from recognizing when and how to use it. A psychology student might apply a theory to an unfamiliar case and identify something it does not explain. A mathematics student might examine a flawed solution, locate the error, and show how to correct it. A science student might interpret an unexpected result rather than repeat a laboratory procedure.
The task does not need to be elaborate. Changing one variable or presenting a new example can reveal whether a student can transfer knowledge beyond a rehearsed answer. The key is to make application part of what receives credit, not an optional addition to information recall.
3. Create: Give students something meaningful to produce
A useful product might be an essay, model, annotated source collection, spreadsheet, lesson, patient-education resource, or demonstration. Changing the format alone does not solve the assessment problem. An infographic is not inherently better evidence than a paper, and a slide presentation can be generated as easily as prose. What matters is whether producing the work requires disciplinary decisions and whether students can support and explain those decisions.
For example, a history student could select several primary sources for a small exhibit and discuss how the sources support competing interpretations. The exhibit provides a concrete basis for evaluating evidence, credibility, and uncertainty.
4. Explain and defend: Ask about an important choice
An explanation does not have to become a lengthy oral examination. A short exchange can focus on one decision:
“Why did you select this source?”
“What alternative did you consider?”
“What would change your conclusion?”
“How would your answer change if this condition were different?”
These questions help instructors explore understanding that the finished submission may not reveal. The exchange should be presented as a normal part of learning, not a surprise interrogation. Students should know what they may be asked to explain and how the explanation will be assessed.
A recorded report can be useful, although it is not conclusive evidence of independent understanding. When appropriate, an individualized follow-up question adds an opportunity to explore the reasoning more directly.
5. Show process and reflect: Make selected development visible
Instructors do not need a complete record of every keystroke. One well-chosen checkpoint may provide useful information about how the work developed. A student might submit an early thesis with a source note, an initial calculation, a draft excerpt, or a brief explanation of a changed decision.
Reflection becomes more informative when it concerns something specific. Instead of asking only “What did you learn?” an instructor might ask what changed between the first and final approach, which evidence most influenced the conclusion, or what limitation remains.
Where AI use is permitted, students can explain what the tool contributed, what they rejected, and what they independently checked. Such disclosures support transparency, but should not be treated as proof that every stage of the process has been documented.
What redesign can look like across subjects
The framework does not mean every course should use the same assignment format. Different disciplines require discrete types and supporting evidence.
Writing and composition
Keep the essay, but add an early thesis or source checkpoint. After submission, ask students to explain one interpretive or rhetorical choice and identify a revision that improved the argument. This preserves writing as the central activity while making the writer’s decisions more visible.
Mathematics and science
Use selected problems that require interpretation, comparison, or error analysis rather than relying entirely on repeated procedures. A student might compare two methods, explain why an answer is unreasonable, or discuss what an unexpected laboratory result does and does not support. Essential calculations and laboratory writing remain part of the work.
Psychology and health education
Ask students to apply concepts to a provided case or develop an educational resource for a defined audience. Require an explanation of the supporting evidence and the limits of the recommendation.
Students should not have to disclose personal psychological or health experiences to demonstrate understanding. Where practical skills are the goal, an observed demonstration can be paired with questions about the reasoning behind the action.
Business, history, and education
A business student might prepare a recommendation and explain a consequential assumption. A history student might build an interpretation from primary sources and discuss their limitations. An education student might design a short lesson and explain how they would determine whether learning occurred. The common element is not the presentation format. It is a clear connection between the intended capability and the evidence collected.
Match AI rules to the learning goal
A single rule for every assignment can obscure important differences. When the goal is independent competence in a foundational skill, restricting AI may be appropriate. When the goal is evaluating AI output, comparing alternatives, or practicing a permitted professional workflow, AI may be deliberately included.
Students need instructions that explain both the boundary and its purpose. “AI is allowed” is often too vague. Does that include brainstorming, outlining, editing, solving, generating prose, or analyzing data? An assignment can clarify what assistance is permitted, what work must remain the student’s own, and what use should be disclosed. It should also state that students remain responsible for checking factual claims, sources, and conclusions.
AI literacy involves more than obtaining a useful response. It includes deciding when the tool is appropriate, recognizing weak output, and knowing when not to rely on it.
Why detection should not lead the response
AI-text detection may appear to offer a simpler solution than assignment redesign. However, research has documented accuracy problems and false classifications, including concerns involving writers for whom English is not their primary language.
These findings require careful interpretation. Studies typically evaluate particular tools, versions, and writing samples. They do not establish that every current detector performs identically. They cannot provide an unassailable reason to treat a detection score as conclusive evidence of misconduct.
A question about the student’s reasoning, a source check, or an appropriate follow-up task can furnish more educationally relevant information. If misconduct is suspected, institutions still need fair procedures and evidence. A hesitant oral response, like a detector score, should not, in and of itself, be regarded as final evidence.
Protect accessibility, privacy, and fairness
Assessment redesign can create barriers if instructors unintentionally measure confidence, technology access, or oral fluency instead of the intended knowledge. A student who understands the material may find a public presentation difficult. A multilingual student may need more time to formulate an explanation. A multimedia assignment may reward access to sophisticated software more than exhibiting disciplinary understanding.
Where the learning outcome permits, alternatives can include an instructor-only conversation, a short recording, a written explanation, or a lower-technology artifact. Accommodations should preserve the capability being assessed rather than require identical conditions for everyone.
Process evidence also needs boundaries. Selected drafts can be useful without requiring unrestricted access to personal accounts or exhaustive version histories. Students should know what will be collected and why. Better assessment should increase confidence in learning, not expand surveillance.
Start with one assignment, not an entire course
A manageable redesign begins with a question: What should students be able to do, and what evidence would convince me they can do it?
From there, an instructor can follow five practical steps:
Identify the core learning outcome.
Consider which parts of the existing task AI could complete without student understanding.
Add one meaningful application or contextual requirement.
Add one focused opportunity to show reasoning or development.
Clarify the AI boundary and align the grading criteria with the evidence.
The new component should replace lower-value work where possible. A brief explanation might replace part of a longer submission. An error-analysis task might replace several repetitive problems.
Even short oral checks take time across a large class. Group demonstrations with individual questions, one checkpoint instead of several drafts, and carefully planned sampling can help manage workload. These arrangements should be transparent and should not leave important learning outcomes dependent on chance.
What the evidence does and does not establish
The five-move framework comes from a practice-oriented conceptual paper that combines existing scholarship with teaching experience. It does not report an experiment showing that these moves improve grades, reduce inappropriate AI use, or lower faculty workload.
That distinction matters. The framework offers a research-informed approach to design, not a validated guarantee. Its practical value is that it directs attention toward the evidence an assignment collects. Instructors can try a small redesign, examine the quality of student explanations, consider accessibility and time demands, and revise accordingly.
The question that guides better assessment
Rather than asking, “How can I make this assignment impossible for AI?” ask: “What would students need to do to show that they understand and can apply what I am teaching?”
That question keeps the focus on learning. It leaves room for writing, technology, creativity, independent practice, and professional judgment while recognizing that no single polished product tells the whole story.
Explore Related Articles
Using AI: Start Thinking, Not Stop Thinking
A closely related Brief Insight on using AI to develop understanding rather than bypass the thinking needed to learn. It appears on the Education, Learning, and Careers page.
Preparing for College, Careers, and Work in an AI-Driven Economy
Explores why subject knowledge, communication, practical experience, and ethical judgment remain important as educational and workplace tools change. It appears on the same topic page.
How to Choose AI Tools That Actually Help
Provides practical questions about purpose, privacy, accessibility, and whether a tool supports rather than replaces the user’s thinking.
Further Resources
Tertiary Education Quality and Standards Agency: Assessment Reform for the Age of Artificial Intelligence
Expert-developed guidance for higher education on adapting assessment to the opportunities and risks of generative AI. The accessible landing page includes the downloadable discussion paper, Assessment reform for the age of artificial intelligence (2023).
UNESCO: Guidance for Generative AI in Education and Research
Miao, F., and Holmes, W. (2023). Guidance on human agency, privacy, inclusion, equity, and the educational purposes that should shape AI adoption.
Developing Evaluative Judgement for a Time of Generative Artificial Intelligence
Bearman, M., Tai, J., Dawson, P., Boud, D., and Ajjawi, R. (2024). An open-access conceptual paper explaining why students need to judge the quality of AI outputs, their own work, and the processes used to produce it. Assessment & Evaluation in Higher Education, 49(6), 893–905.
Testing of Detection Tools for AI-Generated Text
Weber-Wulff, D., and colleagues (2023). An open-access evaluation of 14 detection systems that documents accuracy limitations and the effects of text modification. Its results concern the systems and conditions tested, rather than every subsequent product version. International Journal for Educational Integrity, 19, Article 26.
GPT Detectors Are Biased Against Non-Native English Writers
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). A freely accessible article reporting substantial misclassification of non-native English writing in the detectors studied and discussing the implications for fairness. Patterns, 4(7), 100779.
A Scoping Review on How Generative Artificial Intelligence Transforms Assessment in Higher Education
Xia, Q., Weng, X., Ouyang, F., Lin, T. J., and Chiu, T. K. F. (2024). A review of 32 empirical studies examining implications for students, educators, and institutions, including assessment diversity, responsible learning, and self-regulation. International Journal of Educational Technology in Higher Education, 21, Article 40.