Learn With Jesse
Menu

Feedback, assignments, and assessment

Process-based assessment when students can use AI

Look at how a student got to the work, not only at the finished file. Drafts, an AI-use log, and one short live check give you evidence that survives whether your course allows AI or bans it.

Emerging evidenceWorks with or without AIIn personLive onlineSelf-paced onlineAny class sizeA few hours to set up

What it is

Process-based assessment collects evidence of the thinking behind a piece of work. That can be a proposal and outline, a draft with revisions, a short log of what the student asked an AI tool and what they kept, and a brief live check such as a few questions about their own work or a short in-class task.

Detection software is not a substitute. In a test of 12 publicly available tools and two commercial systems, the authors concluded that none was accurate or reliable enough, and the errors leaned toward labeling human writing as AI-generated.1 A separate study found that GPT detectors frequently misclassify writing by non-native English speakers as AI-generated.2

Evidence and fit at a glance

Emerging evidence

The case for redesigning assessment rests on detector tests, a real-world test of AI submissions in one university's exam system, and conceptual analyses. I did not find controlled trials showing that process-based designs improve learning or reduce misuse, so the evidence is labeled emerging.

How evidence is labeled

  • Set-upA few hours to redesign one major assignment with checkpoints and a live check. Class size drives the ongoing time cost.
  • Class timeLight touches at each checkpoint, plus 2 to 5 minutes per student for a live check, or a short in-class task.
  • DeliveryIn person, Live online, Self-paced online
  • Class sizeAny class size. Works best with under 25 students or 25 to 60 students
  • AI roleWorks with or without AI

When it works, and when it doesn’t

Best when

  • Students can use AI outside class, as in take-home work and online courses, and you still need evidence of what each student can do.
  • The assignment is large enough to have natural checkpoints, such as a proposal, an outline, a draft, and a revision.
  • You can add a short live check, such as a conversation about the student's own work or an in-class problem, even if only for a sample or for flagged cases.
  • Your course allows AI and you want students to show what they asked, what they kept or rejected, and why.

Weak when

  • The class is very large and you have no help. Individual live checks do not scale, so rely on short in-class tasks and sampling.
  • The process artifacts are submitted only at the end, with no live check. Drafts and logs can themselves be produced with AI, so they are weak evidence on their own.
  • The only evidence is a detector score. Detectors were not accurate or reliable in testing and were biased toward flagging human writing.1,2
  • The only change is a rule in the syllabus. The authors of one analysis argue that rules students may ignore create an "enforcement illusion" and that constraints need to be built into the assessment itself.3
  • Students who find live questioning very stressful, or who need accommodations, have no alternative route.

How to do it

  1. Name what the grade must show.Write the learning outcome, then ask: what evidence would show that this student can do this without help?
  2. Choose the level of AI use for this assignment.A tiered scale, the AI Assessment Scale, lets you choose how much generative AI use suits each assessment, based on its learning outcomes.4 Different assignments in one course can sit at different levels.
  3. Add at least one structural element.Build in something the student cannot complete without being there: in-class writing or problem solving, an oral question, or a supervised step. Corbin and colleagues call changes of this kind structural, in contrast with discursive changes that rely on instructions students may ignore.3
  4. Stage the work.Collect a proposal, an outline, and a draft with brief feedback at each step. Keep each checkpoint light, and give credit for completion.
  5. If AI is allowed, ask for a short log.What did you ask, what did you keep or reject, and why? Treat the log as evidence to discuss, not proof.
  6. Hold a short live check.Two or three questions about the student's own work, scored with a rubric. If a submission looks inconsistent with a student's in-class work, talk to the student before concluding anything.

With AI and without

Without AI

If AI is banned, the only enforceable parts of the design are structural ones: supervised in-class work and live checks. A ban written in the syllabus, with nothing built into the assessment, depends on students following instructions.3

With AI

If AI is allowed, say at what level and for which steps, and ask students to show how they used it. AI can also help you draft rubrics and question lists for live checks, which you then review and own. Do not run student work through unapproved tools, and do not rely on AI detectors.

What the research says

Why a finished file is a weak signal

In a real-world test, researchers submitted fully AI-written work into a UK university's assessment system across five undergraduate psychology modules. The authors report that most of the submissions went undetected, that the AI work received somewhat higher grades than real students' work, and that AI submissions outperformed a random sample of genuine student work in most modules. They raise concerns about unsupervised take-home assessment.5

Why detectors are not the answer

A study that tested 12 publicly available detection tools and two commercial systems concluded that none of them was accurate or reliable. The authors report a bias toward labeling human-written text as AI-generated, and found that obfuscating AI text further degraded detection.1

A separate study reported that GPT detectors frequently misclassify non-native English writing as AI-generated, which raises fairness concerns for evaluation in educational settings.2

What the redesign literature argues

A conceptual analysis argues that generative AI undermines assessment validity because students can complete tasks without showing their own capability. The authors note that common institutional responses, such as traffic-light frameworks that mark what AI use is permitted, mainly communicate rules rather than redesign how assessments work. They distinguish discursive changes, which rely on instructions students may ignore, from structural changes, which build constraints into the assessment itself, and they argue that discursive approaches create an enforcement illusion.3

A separate framework proposes the AI Assessment Scale, a flexible tiered scale that lets educators decide how much generative AI use suits each assessment, based on its learning outcomes. Its authors present it as a way to bring clarity to students and staff and to shift the conversation toward how these tools can support learning.4

Pedagogical justification, ready to paste

Use this in a lesson plan, syllabus, or course-design form. Replace the [bracketed] parts with your own course details. Check the claims against the references before you submit.

One-sentence version

Major assignments are assessed through staged checkpoints and one live check, with AI use stated for each assignment, because AI-written submissions mostly went undetected in a university assessment system,5 detection tools were not accurate or reliable and were biased toward flagging human writing,1 and the authors of a conceptual analysis argue that rules alone create an enforcement illusion unless constraints are built into the assessment.3

Full rationale

Because students can use generative AI outside class, major assignments are assessed through evidence of process as well as product. Each assignment states the level of AI use allowed, chosen from a tiered scale according to the learning outcomes the assignment assesses.4 Students submit a proposal, an outline, and a draft with brief feedback, keep a short log of any AI use, and answer two or three questions about their own work in a short live check. In a real-world test at a UK university, AI-written submissions mostly went undetected and received somewhat higher grades than real student work, which raises concerns about unsupervised take-home assessment.5 Detection tools were found not accurate or reliable, with a bias toward labeling human writing as AI-generated,1 and they misclassify non-native English writing,2 so no detector scores are used as evidence. The design follows an analysis arguing that changes built into the assessment itself are needed, because rules students may ignore create an enforcement illusion.3 Controlled evidence on this approach is limited, so the course compares live-check results with written work to judge whether the design is working.

Learning outcome
Students will be able to [verb + skill] on [assignment] and explain their own reasoning in a short oral or in-class check.
Evidence of learning
[Rubric scores] on the final work, on each checkpoint, and on a short live check, plus agreement between the written work and the live check.
Learning activity
[Assignment] staged as a proposal, outline, and draft, with an AI-use log at level [N] of the course's AI scale, and a live check of [2 to 5] minutes.

The copy buttons paste plain text. The pasted version renumbers the references from 1 and includes them with a source line, so it works on its own. The numbers shown here match the reference list below.

Limits of the evidence

  • I did not find controlled trials of process-based assessment on learning outcomes or on misuse. The sources here are detector tests, a real-world test of AI submissions, and conceptual analyses.
  • The test of AI submissions covers one university and one discipline (psychology), and the detector tests are from 2023. Tools change, so treat the detector findings as a reason for caution, not a permanent ranking.
  • The AI Assessment Scale and the structural-versus-discursive distinction are frameworks and arguments, not effects measured in experiments.
  • Drafts and logs can be faked with AI if they are only collected at the end. They help most when combined with a live check.
  • This design takes instructor time, and the cost rises with class size. Live checks need accommodations and alternatives for students who need them.

Questions faculty ask

Can I use an AI detector to catch students who cheat with ChatGPT?

Not reliably. A test of 12 publicly available tools and two commercial systems concluded that none was accurate or reliable, with a bias toward labeling human text as AI-generated.1 Another study found that detectors frequently misclassify non-native English writing as AI-generated.2 A score is not proof of anything.

What is the difference between structural and discursive assessment changes?

Discursive changes rely on instructions, such as an AI policy statement or a traffic-light scheme. Structural changes build constraints into the assessment itself, such as supervised in-class work or live questioning. One analysis argues that discursive changes alone create an enforcement illusion.3

How do I decide how much AI to allow on an assignment?

Start from the learning outcome. The AI Assessment Scale is a tiered framework that lets educators choose the level of generative AI use that suits each assessment.4 Different assignments in the same course can use different levels.

Does an AI-use log prove a student did the work?

No. A log is evidence to discuss, and it can itself be written with AI. It is most useful alongside a short live check. This is a practical judgment, not a research finding.

References

  1. Weber-Wulff D, Anohina-Naumeca A, Bjelobaba S, et al. Testing of detection tools for AI-generated text. Int J Educ Integr. 2023;19(1):26. doi:10.1007/s40979-023-00146-z
  2. Liang W, Yuksekgonul M, Mao Y, Wu E, Zou J. GPT detectors are biased against non-native English writers. Patterns. 2023;4(7):100779. doi:10.1016/j.patter.2023.100779
  3. Corbin T, Dawson P, Liu D. Talk is cheap: why structural assessment changes are needed for a time of GenAI. Assess Eval High Educ. 2025;50(7):1087-1097. doi:10.1080/02602938.2025.2503964
  4. Perkins M, Furze L, Roe J, MacVaugh J. The Artificial Intelligence Assessment Scale (AIAS): a framework for ethical integration of generative AI in educational assessment. J Univ Teach Learn Pract. 2024;21(6). doi:10.53761/q3azde36
  5. Scarfe P, Watcham K, Clarke ADF, Roesch E. A real-world test of artificial intelligence infiltration of a university examinations system: a "Turing test" case study. PLoS One. 2024;19(6):e0305354. doi:10.1371/journal.pone.0305354

About this page

Written by Jesse Rogers, Learning Lab Coordinator, Palm Beach State College. Last updated October 15, 2026.

AI disclosure. Model used for research, drafting, editing, and verification: claude-sonnet-5-5 (Claude, by Anthropic). The author is responsible for every claim and citation.

How it was checked. Bibliographic details were checked against DOI records. Numbers and findings were checked against abstracts or open full texts. Where that was not possible, the page says so. See the method page.

What I could not confirm.

  • The studies of detectors, AI submissions, and assessment redesign were checked against their abstracts and publisher metadata. Figures such as false-positive rates and the share of undetected submissions appear in the full texts and are not quoted here. The description of the AI Assessment Scale is limited to what the abstract says, that it is a tiered, flexible scale. Its level names are not reproduced.

License. CC BY 4.0. Share and adapt it with credit. This is an independent project and is not an official resource of Palm Beach State College.

Corrections. If you find an error, message me on LinkedIn and I will check it and fix it.

Cite this page (AMA). Rogers J. Process-based assessment when students can use AI. Learn With Jesse. Published October 15, 2026. Updated October 15, 2026. Accessed [date]. https://learnwithjesse.com/techniques/process-based-assessment/