AI ASSESSMENT SCALE · AIAS IN PRACTICE

AI Assessment Scale (AIAS) in Practice: How to Make Each Level Hold

The AI Assessment Scale gives every assessment a level of permitted AI use, from No AI to AI Exploration. This guide explains the original and the revised scale from the source papers, then shows the evidence and checks that make each level hold.

By Eduface · October 2026 · 15 min read

Your programme team has adopted the AI Assessment Scale. Every assessment brief now carries a box that says Level 2: AI Planning. On the March intake, a take-home report comes in fluent, well structured and strangely generic, and the associate lecturer marking it asks you what Level 2 allows her to do about it. For an unsupervised report, the honest answer is very little, unless the assessment was designed to show its own working.

What is the AI Assessment Scale?

The AI Assessment Scale (AIAS) is a five-level framework by Mike Perkins, Leon Furze, Jasper Roe and Jason MacVaugh, published in the Journal of University Teaching and Learning Practice in 2024, that tells students how much generative AI they may use on a specific assessment. The revised scale (Perkins, Roe and Furze, December 2024) names the levels No AI, AI Planning, AI Collaboration, Full AI and AI Exploration. A level only holds when the assessment design can show that it held.

What is the AI Assessment Scale, and where does it come from?

The AI Assessment Scale is a framework that gives each assessment task one of five levels of permitted generative AI use, so students and staff share the same expectation before work starts.¹ It was published open access in the Journal of University Teaching and Learning Practice, volume 21, issue 6, in 2024, by authors at British University Vietnam, Deakin University and James Cook University.¹

It began as a traffic light with three settings, “No AI”, “Some AI” and “Full AI”, which the authors refined into five levels.¹ Its purpose, in their words in the revised paper, is to support transparent conversations between educators and students about appropriate AI use and to push assessment redesign, “rather than attempting to control AI use”.² By December 2024 they reported use in over a dozen countries, translations into at least 12 languages, and recognition of the original version by Australia’s higher education regulator, TEQSA, as a potential tool for delineating appropriate AI use in assessment.²

So the AIAS is a communication and design tool; on its own it stops nobody from using AI. For the wider design question, see the hub article on designing assessment for academic integrity in the age of AI.

What are the five levels of the original and the revised AIAS?

The revised AIAS keeps five levels, renames three, moves “Full AI” from Level 5 to Level 4, and adds AI Exploration at the top.² The level names below are exactly as the authors print them.¹ ²

Level

Original scale (2024)

Revised scale (December 2024)

What changed

No AI

Now explicitly a controlled, supervised environment

AI-Assisted Idea Generation and Structuring

AI Planning

Adds initial research; the submission must show how the student developed and refined the ideas

AI-Assisted Editing

AI Collaboration

Drafting with AI allowed; the appendix of AI-free work is dropped

AI Task Completion, Human Evaluation

Full AI

Students direct AI throughout and are assessed on how they do it

Full AI

AI Exploration

New: students and educators co-design new uses of AI

Three revisions matter in practice.

Level 1 now means supervision. Level 1 assessments “should occur in supervised settings where the absence of AI can be assured, rather than relying on honour systems or detection tools for take-home work”.² The original paper had already recommended supervision or low-stakes, formative use for Level 1.¹

The Level 3 appendix is gone. The original Level 3 asked for the student’s AI-free work in an appendix. The revision dropped it, “acknowledging that the current detection methods cannot reliably verify such submissions”.²

The colours changed. Red and green read as “bad” and “good”, so the revision uses a neutral palette and a circular version to show that no level is better than another.²

The authors call the AIAS a starting point that “is not a prescriptive tool”.² That is permission to adapt it, and a reminder that an adapted scale still needs a check behind every level. If a colleague shares a graphic labelled “version 2” or “revisited”, check the names: Level 3 AI Collaboration and Level 5 AI Exploration mean it is the revised scale.

Why does an AIAS level only count if you can check it?

A level on a brief states what is permitted; whether it held depends on the conditions in which the work was done. The revised paper says so. For Level 2: “we cannot technically restrict AI use to planning stages”. For Level 3, enforcing limits on AI-assisted drafting “is neither practical nor pedagogically sound”.²

That gives a head of quality one test per level: if a validating partner or an appeal panel asked how we know this level held, what would we show them?

Level 1, No AI

Enforceable only under supervision: exam room, in-class task, live oral

On take-home work it is a request, not a rule

Levels 2 and 3

AI Planning and AI Collaboration

The boundary cannot be enforced unsupervised

Only evidenced, through process artefacts and a conversation

Levels 4 and 5

Full AI and AI Exploration

No AI limit left to enforce

The check is whether the student directed the AI and judged its output

The scale tells students what is permitted. Only supervision enforces it; everything else is evidence.

Figure 1: Only Level 1 can be enforced, and only under supervision. For Levels 2 and 3 you collect evidence; for Levels 4 and 5 you assess judgement.

Three kinds of check do three different jobs.

Supervision enforces. Only an invigilated exam, an in-class task or a live oral can make “no AI” true.

Process evidence makes a level plausible. Drafts show how the work developed, but the authors warn that tracking tools “are fallible and can be circumvented by determined students”.² Draft history supports a judgement; it does not settle one.

A conversation checks the outcome directly. The original paper lists “ad-hoc or planned viva-voce examinations, question and answer sessions” as Level 1 activities.¹ A short oral on the student’s own work is a secured moment inside an open task.

For heads of quality:

Approve the level and the check together. A level with no check behind it treats the student who follows it and the student who ignores it identically, and gives your associate lecturers nothing to act on.

How do you make each AIAS level hold?

You make a level hold by pairing it with evidence that it was respected and a check your staff can run in the time available. The “may do” column follows the authors’ wording for the revised scale;² the evidence and checks are our working method, not part of the AIAS.

Level

What the student may do with AI

Evidence the level held

How to check it

1 No AI

Nothing, at any point

Work produced where AI was not available

Invigilated exam, in-class task or live oral; online, a live video oral or remote invigilation

2 AI Planning

Brainstorming, outlining, initial research; final work shows how ideas were developed

An outline, then a draft that develops it; a short AI-use declaration

Checkpoint submissions; a 5 to 10 minute conversation on the student’s choices

3 AI Collaboration

Help with drafting, feedback and refinement; AI output critically evaluated and modified

Visible revision between drafts; the student can defend choices in the text

Draft checkpoints with feedback, then an oral on two or three passages

4 Full AI

AI throughout, with the student directing it to the task goals

Prompts, choices, rejected outputs, judgement on quality

Assess the process record; ask the student to critique an AI output live

5 AI Exploration

Creative use, possibly co-designing approaches with the instructor

A documented design and the student’s evaluation of the AI’s contribution

Supervision meetings and a final presentation or viva

Level 1 is the only level you can enforce outright, and only in the room. The authors note that the supervision requirement “presents challenges for online and distance-learning environments”.² If you teach online, decide which outcomes need a live, supervised moment and budget for it.

Levels 2 and 3 are where most private-college coursework sits. The line between planning and drafting is crossed on the student’s laptop, out of sight. What you can see is whether the thinking moves between checkpoints and whether the student can talk about the work. The authors also warn about the “illusion of finality”, the tendency to accept AI-generated text as complete and authoritative.² A conversation about the student’s own passages shows whether that happened.

Levels 4 and 5 move the check from the product to the judgement. The risk is a student who directs nothing and accepts everything. The authors add that these assessments must “remain valid, regardless of students’ access to advanced AI models”, which may mean providing tools institutionally.²

Equity note:

An oral check tests understanding, not accent or fluency. Mark the content, let students refer to their own work, and offer the usual adjustments. The original paper flagged the reverse risk: students with English as a first language or paid AI tools are better placed to use AI undetectably in “no AI” take-home work.¹

What is the main criticism of the AIAS?

The main criticism is that the levels between No AI and Full AI cannot be enforced in unsupervised work, so five levels suggest a precision that does not exist. The sharpest version comes from Danny Liu and Adam Bridgeman at the University of Sydney. Their “two-lane” approach separates secured lane 1 assessment (in-class assessment, viva voces and interactive orals, supervised exams) from open lane 2 assessment in which students use AI.³ They write: “We do not foresee a viable middle ground between the two lanes. It needs to be assumed that any assessment outside lane 1 (i.e. that is un-secured) may (and likely will) involve the use of AI.”³

The AIAS authors answer this in the revised paper. They acknowledge “this reality in unsupervised environments” but argue the AIAS “serves a different purpose”: transparent conversation about appropriate use and redesign for the chosen level, not control.² They add that five levels give students more clarity than “the broad and potentially ambiguous dichotomy of the two lanes”.²

They record two more weaknesses.

Superficial adoption. The scale has been used for superficial changes to assessment, and Pratschke (2024) argued that full transformation will eventually be required.² In practice: a level printed on an unchanged essay brief.

Declarations are unreliable. Research shows students are reluctant to declare AI use even when it is permitted (Gonsalves, 2024).² An AI-use statement is context, not evidence.

Our reading: both sides agree on the facts. The AIAS authors accept that unsupervised levels cannot be policed, rebuilt Level 1 around supervision, and lean on Dawson and colleagues’ argument that assessment validity must take precedence over traditional notions of academic integrity.² The disagreement is about everything outside the exam room, and you do not have to pick a side to act on it.

AIAS or the two-lane approach: which fits a private college?

Most private colleges will get more from combining them: AIAS levels for what is allowed on each task, and a lane 1 moment for the outcomes that must be assured.

Question

AI Assessment Scale (AIAS)

Two-lane approach (University of Sydney)

What it is for

Telling students what AI use is permitted, and prompting redesign²

Deciding which assessments are secured and which assume AI use³

Worked example: mapping a marketing plan to an AIAS level

This example is illustrative. All numbers are example numbers; replace them with your own.

Learning outcome (what must be shown without AI?)

AIAS level (what AI use still leaves the outcome visible?)

Evidence (what shows the level held?)

Check (supervision, process trail or conversation)

Cost (minutes × students × intakes)

Brief (level, check and decision rule in writing)

Moderation (same check on every site)

Start from the outcome, choose the level, then name and cost the check before the brief goes out.

The setting. A private business school runs a two-year programme with three intakes a year, on two campuses plus an online cohort. Its second-year module Marketing Planning has 150 students per intake and six associate lecturers. The assessment is a 3,000-word marketing plan for a real small business, worth 100% of the module. Two outcomes matter: analyse a market with appropriate frameworks and data, and justify a marketing mix to a client.

Step 1: decide what must be shown without AI. Graduates will research and draft with AI at work, so the plan need not be AI-free. The outcome to assure is the second: can the student justify the recommendations in their own words?

Step 2: choose the level. The written plan goes to Level 3, AI Collaboration. A Level 1 oral is added for the justification outcome, and the weighting becomes 70% plan, 30% oral.

Step 3: name the evidence and checks, and cost them.

Week

Component and AIAS level

Evidence it produces

Staff time (example numbers)

Outline: business, frameworks, data sources (Level 2)

The student’s starting choices, on record

About 5 minutes per student

Full draft (Level 3)

Revision visible from outline to draft to final

One formative feedback round

Final plan plus a 150-word AI-use statement (Level 3)

Final product and declared AI use

Normal marking

10-minute client briefing on the student’s own plan (Level 1, live)

Student defends recommendations under questioning

150 × 15 minutes = 37.5 hours per intake

The oral is costed at 15 minutes per student: ten of conversation, five of notes and marking. Over six associate lecturers that is about 6.25 hours each per intake, or 112.5 staff hours a year. That number belongs in the workload model before the brief is approved.

Step 4: write the decision rule. The oral is marked on three criteria: explains the analysis, justifies the recommendation with evidence, handles an unprepared challenge. A weak oral lowers the oral mark; on its own it is not a misconduct finding. Where the gap is large, for example a plan at 70 and an oral below 40, the module leader reviews the case under the normal academic integrity procedure.

Step 5: moderate across sites. Every site uses the same structure: two questions on passages the student picks, one on a passage the marker picks, one challenge. The module leader moderates a sample from each site. Questions anchored in each student’s own plan are much harder to pass on to the next intake than a fixed exam question.

The Level 3 label did not change. What changed is that the outcome the programme cares about most now has a check the team can defend. Our guide to AI oral examinations in higher education covers how to structure and moderate that kind of oral.

A checklist before you approve an AIAS level

Use this when an assessment brief reaches your quality committee or programme board.

Outcome first. Which outcome must be shown without AI, and which can be shown with it?

Level matches conditions. Is Level 1 used only where the task is supervised?

Evidence named. For Levels 2 and 3, which artefacts show the thinking developing?

Conversation planned. Is there an oral or in-class moment for the outcomes that must be assured, and who runs it?

Cost per intake. Minutes × students × intakes, inside the associate lecturers’ workload model?

Decision rule written. What happens when the written work and the oral disagree, and who reviews it?

Equity checked. Is assistive technology allowed at Level 1? Does Level 4 depend on paid tools?

Students told. Does the brief use the authors’ level name and state the check in plain language?

Detectors left out. Both the AIAS authors and the Sydney team describe AI detectors as unreliable.²,³ See why AI plagiarism detection is failing higher education.

Where does Eduface fit in an AIAS-based assessment?

Eduface can produce part of the process evidence and run part of the conversation for Levels 2 and 3; it does not prove authorship, and it is not an AI detector. We are Eduface, an AI platform for assessment and feedback in higher education, so we have a position here.

Draft progression through the Paper Grader’s formative feedback. The Paper Grader gives feedback in several rounds, from first draft through second draft to final version, following the lecturer’s instructions. Student progress is followed across successive drafts, so the lecturer sees how the work developed. The lecturer chooses whether feedback reaches the student directly or after review and approval. For a Level 2 or 3 task, that is the checkpoint evidence in the table above.

An oral check on the student’s own paper with Academic Integrity (beta). In the Academic Integrity module, which is in beta, the student gets critical questions about their own paper and answers them orally, as in a thesis defence. The Oral Examination module, also in beta, covers formal oral assessment: the lecturer configures the examiner’s character and situation, the conversation adapts to the student’s answers, and it is transcribed and assessed in real time.

The lecturer decides. The AI assists. The academic decides. Grades go back to the gradebook in Moodle, Canvas, Blackboard or Brightspace after the lecturer approves them. Student work is processed in the EU, never used to train AI models, and not sent to third-party AI APIs such as OpenAI.

Where another option fits better, use it. For a cohort of twenty, a lecturer can run live vivas without any software. For a high-stakes Level 1 outcome, an invigilated room is stronger than any platform.

What Eduface doesn’t do:

Eduface is not an AI detector and does not score text for the likelihood that AI wrote it. It does not prove who wrote a paper; no tool does, and the AIAS authors warn that process tracking can be circumvented.² It cannot make take-home work Level 1; only supervision does. Academic Integrity and Oral Examination are in beta, so pilot them on one module before a programme relies on them.

What this method leaves out

Programme-level mapping. This guide works task by task; Sydney’s advice points to mapping which programme outcomes need a secured check.³

Staff use of AI in marking. The AIAS governs students; how markers use AI is a separate policy.

Misconduct procedure. A check produces evidence; your academic integrity policy decides what happens next.

Frequently asked questions

Is the AI Assessment Scale free to use?

Yes, with conditions. The original AIAS article is open access under a Creative Commons BY-NC-SA 4.0 licence, and the revised scale graphic carries the same mark.¹ ² That allows non-commercial adaptation with attribution; a for-profit provider should ask its legal team how that applies.

Can a student use assistive technology at AIAS Level 1?

Yes. The revised AIAS paper says Level 1 tasks must accommodate students who require assistive technology, as distinct from AI tools that compromise validity.² Its own Level 1 example provides a word processor or a scribe. Name the permitted tools in the brief so every campus applies the same rule.

Should we use AI detectors to enforce the AIAS levels?

No. The AIAS authors describe AI detectors as unreliable and argue against using them to catch students,² and the University of Sydney team advises against them too, noting possible bias against non-native English writers.³ Use process evidence and a conversation instead.

Can an online programme use AIAS Level 1?

Yes, but only for tasks you can supervise. The revised paper notes that AI-free conditions are hard to guarantee in online and distance learning without supervision.² In practice: a live video oral or remote invigilation for the outcomes that must be assured, and Levels 2 to 4 with process evidence for the rest.

Does the AIAS replace our academic misconduct policy?

No. The authors describe the AIAS as a starting point for reforming assessment, explicitly not a prescriptive tool.² It tells students what AI use is permitted; your academic integrity policy still defines a breach, who reviews a case, and what evidence is needed.

References

1. Perkins, M., Furze, L., Roe, J., & MacVaugh, J. (2024). The Artificial Intelligence Assessment Scale (AIAS): A framework for ethical integration of generative AI in educational assessment. Journal of University Teaching and Learning Practice, 21(6). https://doi.org/10.53761/q3azde36 [Key finding: the original five levels; the scale began as a “No AI / Some AI / Full AI” traffic light; Level 1 recommended under supervision or for low-stakes formative tasks; viva-voce examinations listed as Level 1 activities; equity risk of “no AI” take-home work; open access under CC BY-NC-SA 4.0.]

2. Perkins, M., Roe, J., & Furze, L. (2024). The AI Assessment Scale revisited: A framework for educational assessment (preprint, December 2024). arXiv:2412.09029. [Key finding: revised levels No AI, AI Planning, AI Collaboration, Full AI, AI Exploration; Level 1 in supervised settings; Level 3 appendix removed because detection cannot verify it; AI use cannot technically be restricted to planning; tracking tools can be circumvented; neutral colours replace the traffic light; response to the two-lane critique; used in over a dozen countries and at least 12 languages; original recognised by TEQSA.]

3. Liu, D., & Bridgeman, A. (2023, 12 July). What to do about assessments if we can’t out-design or out-run AI? Teaching@Sydney, The University of Sydney. https://educational-innovation.sydney.edu.au/teaching@sydney/what-to-do-about-assessments-if-we-cant-out-design-or-out-run-ai/ [Key finding: “We do not foresee a viable middle ground between the two lanes”; any un-secured assessment may and likely will involve AI; lane 1 includes in-class assessment, viva voces and supervised exams, used sparingly for programme-level outcomes; AI detectors are not reliable.]

See it on your own assignments

Eduface drafts criterion-level grades and feedback for your instructors to approve. Book a demo or start free.

Book a demo

Create free account