Authentic Assessment in the Age of Generative AI

What changed for assessment#

For most of the last century, a submitted essay, report or problem set was reasonable evidence that the author understood the material. A language model can now produce a competent version of many of those artifacts in seconds. The submitted artifact still has value, but it no longer proves on its own that a particular student did the thinking.

Lodge, Howard, Bearman and Dawson prepared a discussion paper on this problem for TEQSA, the Australian higher education regulator. They state that many assessment tasks are at risk because of generative AI, and they also observe that assessment design always involves compromise and that there is no formulaic fix. That framing is helpful. It suggests that the work ahead is design work, not a search for a single tool that restores the old certainty.

Two kinds of response#

Corbin, Dawson and Liu distinguish between changing what you tell students and changing the assessment itself. Rules, policies and declarations of acceptable AI use are one kind of response. Changing the task so that it collects different evidence is the other. Their paper argues that structural changes are needed, and that instructions alone are a weak control because they depend on students following them and on instructors being able to tell when they did not.

That distinction is a useful test for any plan. If the plan is mostly a statement about what students may do, it is probably a discursive change. If the plan alters what students must produce or how they must show their reasoning, it is structural.

Ask a validity question first#

Dawson and colleagues argue that validity matters more than cheating. In practice this means starting from the claim you want to make about a student. If you want to say "this student can evaluate a clinical case," ask whether the current task still supports that claim when a model can write the analysis. If it does not, the problem is validity, and no amount of monitoring fixes a task that no longer measures what you say it measures.

Three questions help:

1. What do I want to be able to say about this student after the assessment?
2. What evidence would make me confident in that statement?
3. Could a student with access to a capable language model produce that evidence without the capability I care about?

If the answer to the third question is yes, change the evidence you collect.

Structural changes that tend to hold up#

• Add an explanation of the work. Ask the student to explain and defend what they submitted, in conversation, with follow-up questions. This is the approach behind oral defense and interactive oral assessment. Sotiriadou and colleagues studied authentic assessment, including interactive oral assessment, as a way to uphold academic integrity and build skills. See our guide to what an AI oral exam is.
• Assess the process. Collect proposals, annotated drafts, decision logs or short reflections that show how the work developed. These can be generated too, so they are most useful when combined with an explanation.
• Use tasks tied to specific context. A case discussed in class, a local dataset or a scenario with details that exist only in your course gives a general model less to work with, though it does not remove the risk.
• Use live performance. Presentations with questions, in-class problem solving and simulations in which the student must respond in real time to a counterpart.
• Make AI use part of the task. Ask students to use a model, record what they asked, critique the result and justify what they kept. Lodge and colleagues list appropriate, authentic engagement with AI as a proposition for assessment reform. This approach assesses evaluative judgment directly.

What does not work alone#

Detection software. Weber-Wulff and colleagues tested a set of detection tools and concluded that they are neither accurate nor reliable, with a tendency to classify text as human-written. Liang and colleagues found that detectors consistently misclassified writing by non-native English speakers as AI-generated. A grade or an integrity charge should never rest on a detector score. See AI oral assessment compared with AI detection.

Blanket bans. A ban is a discursive control. It may set a norm, but it does not give you evidence about any one submission.

Task-specific tricks. Hidden prompts, unusual formatting requirements and similar traps tend to age quickly as models change.

Before and after examples#

In each row the original artifact stays. The revision adds evidence of understanding.

For a worked example of reflection, interview and process-evidence redesigns, see our blog essay on redesigning assignments for the age of AI.

Equity, accessibility and fairness#

Oral and live formats introduce their own risks. Students who are anxious, who speak English as an additional language or who have a disability affecting speech may be disadvantaged by a format that rewards fluency. Plan alternatives before you launch: extra time, a text option for responses, and a practice attempt that lets students learn the format. Write rubric criteria around the quality of the reasoning, not the polish of the delivery. Treat recordings and transcripts as education records and follow your institution's retention rules. UNESCO's guidance on generative AI in education calls for a human-centred approach, and that applies here: a person should make the final judgment about a student.

A sequence you can try this term#

1. Choose one existing assignment that currently carries a lot of weight.
2. Write the claim you want to make about students who complete it.
3. Keep the artifact and add a short explanation of it, in a conversation with follow-up questions.
4. Write a four-criterion rubric that scores the reasoning.
5. Offer a practice attempt that does not count toward the grade.
6. Review a sample of recordings or transcripts yourself before you trust any automated summary.
7. After the term, compare the explanation scores with the artifact scores and look for gaps.

Where Rocketproof.ai fits#

Rocketproof.ai provides the explanation step as a structured oral defense after a submission (see the oral exams page) and as AI role-play scenarios for performance tasks. The instructor reviews the transcript and rubric evidence before a grade is released. Other tools and approaches exist, and our guide to AI oral assessment platforms compares several. If you are new to the idea, begin with what authentic assessment is.