Why most declaration forms do nothing
A declaration is a statement the student signs about how the work was made. Most of them ask one question: is this your own work, and was AI used appropriately? The student ticks yes. That answer cannot be checked, because "appropriately" was never defined and "own work" now has a dozen meanings.
The problem is not that students lie. The problem is that the question has no fixed answer. Australian researchers interviewed 19 students and 12 staff at a large university about where the acceptable line sits. Students had each built their own private rulebook for AI use. The rulebooks were complicated, sincere and different from one another. One participant called the line "an absurd line". A tick box asks every one of those students the same vague question and gets back a yes that means something different each time.
A second Australian paper takes the argument further. Its title is "Talk is cheap". The more detailed an institution's rules about acceptable AI use become, the more they expose a gap. That gap sits between what a rule can specify and what a marker can verify. A breach of a written rule only comes to light through leftovers in the text, direct observation, or the student saying so. A declaration that asks only for a promise adds nothing to any of those three. The authors call this an enforcement illusion: the paperwork looks like control and delivers none.
Ask what the form would let you check that you could not check without it. If the honest answer is nothing, the form is theatre. It may still have a place as a reminder of the rules, but it is not evidence of anything.
What a declaration can be as evidence
A declaration cannot prove that a piece of text was written by a person. Nothing on paper can do that, and the two detection sections explain why no software can either. What a declaration can do is narrower and more useful. It turns a vague suspicion into a specific, checkable statement.
CDU's Academic Integrity Policy makes this the working position. Students are responsible for "declaring and acknowledging use of generative artificial intelligence when creating academic content". This matters for how a concern is framed. The alleged breach is not "this text was generated by AI", which no one can establish. The alleged breach is "the student said one thing about how the work was made, and the record shows another". That is a question about a statement, and statements can be tested in ways that text cannot.
The same policy says what a test can look like. Students "may be required to present, explain, or defend their ideas verbally in real-time (face-to-face or online) as part of demonstrating the authenticity of their work". They may also be asked for process evidence, which the policy lists as "research notes, annotated bibliographies, drafts, or version histories". A declaration that names the tools used, the tasks they were used for, and what the student kept, gives those conversations something to start from. A tick box gives them nothing.
One limit has to be stated now and kept in view. A declaration is only as checkable as the facts it asks for. "I used AI appropriately" can never be shown false. "I used Copilot to check grammar on the final draft and for nothing else" can be, in conversation with the student and against whatever they kept.
The range of things "used AI" can mean
Before a declaration can ask a precise question, it needs a vocabulary for what AI use actually is. The useful picture is a range, not a switch. At one end the tool fixes spelling. At the other end the tool produces the reasoning the task was meant to assess, and the student submits it. Everything interesting sits in between, and the middle is where the arguments happen.
Two things follow from the diagram. The first is that "did you use AI" is the wrong question, because yes covers the entire bar. The second is that the line between acceptable and unacceptable is not a property of the tool. It is a property of the task. Rewriting a paragraph is help in a laboratory report where the chemistry is being assessed. The same rewriting is substitution in a unit where writing is the thing being assessed. A declaration has to be written against the learning outcome of the task it sits under, or it is asking about nothing in particular.
Several frameworks try to turn this bar into named bands. The best known is the AI Assessment Scale, an academic framework with a Deakin University co-author. It has five levels: No AI, AI Planning, AI Collaboration, Full AI and AI Exploration, and it ties the permitted level to the learning outcome. It is worth knowing about for two reasons. First, its own designers found the original version hard to communicate. They dropped its traffic-light colours in 2024 because red and green were being read as a ranking. Second, it is not a regulator instrument. TEQSA has not adopted a tier scale. Its 2025 guidance offers three options for any task instead: permit AI use within defined parameters, design the task so AI use is irrelevant, or restrict use through direct supervision. CDU's Generative AI Policy draws categorical lines, not tiers, and leaves it to teaching staff to say which activities and assessments may use gen AI, and how far.
VET uses a different word for the same problem. Under the Standards for RTOs, assessment evidence must be authentic. The ASQA practice guide defines that as the assessor being "assured that a VET student's assessment evidence is the original and genuine work of that VET student". The guide names AI directly, as a question of validating that evidence "has not been plagiarised or generated with artificial intelligence (AI) tools". There is no tier scale in VET either. There is one line, and the assessor has to be satisfied the work sits on the right side of it. CDU's Generative AI Policy adds a hard rule for the marking side: "VET competency judgements must be made by a qualified assessor", and gen AI platforms are not qualified assessors.
What a workable declaration asks for
A workable declaration asks for facts that can be discussed, not a promise that cannot. It asks which tools were used, by name. It asks what each tool was used for, in terms that map onto the bar above rather than onto the word "appropriately". It asks what the student kept, if the task told them in advance to keep anything. And it tells the student that they may be asked to talk through the work, which CDU policy already allows.
Three design rules come out of the research. The first is specificity: every question on the form should have an answer that could be wrong. The second is advance notice. Australian researchers writing on detector evidence put it plainly. Requirements to produce drafts or logs "must be in writing and they must be provided to students in advance as part of the assessment brief". The University of Sydney's teaching guidance says the same thing from the other direction: "silence or lack of drafts cannot be interpreted as guilt". A form cannot ask after the fact for something the student was never told to keep. The third rule is fit to the task. The form should name the learning outcome. It should say which stages of the bar are in and which are out for this piece of work. That is the only thing that makes "out" mean anything.
It is worth being honest about what even a good declaration does not achieve. It does not stop a determined student. It does not make the text itself provable either way. What it does is remove the ambiguity that a vague form creates. An honest student then knows exactly what they are signing. A dishonest one has made a specific statement that can be examined. That is a modest gain. It is also the only gain a declaration can offer, and the tick box does not offer it.
Activity: rewrite this declaration
The declaration below is a composite of the kind that circulates. It is deliberately weak. Click or tap any underlined phrase to see what is wrong with it. Then rewrite it for an assessment task you actually set, using the annotated version further down as a starting point rather than a template to copy. The aim is a form whose every line could, in principle, be shown false.
I declare that this assignment is my own work.
I have not used AI tools inappropriately in completing it.
I understand that the University uses AI detection software.
I will provide drafts if asked to do so.
The version below asks the same student for facts. Notice that it is longer, and that every field either has a checkable answer or tells the student in advance what will be expected. It is a course draft for discussion, not a CDU instrument; any form CDU adopts has to sit inside its own policies.
Fill the first three fields for a task you set this semester. If you cannot say which stages of the bar are out, the task itself is not ready for a declaration, and no form will rescue it. That is worth knowing before the work comes in, and it is where the assessment redesign section picks up.
When a declaration is plainly false
Sometimes the form says "none" and the submission says otherwise. It contains a fabricated citation, a paragraph the student cannot explain, or a reference to data that was never supplied. The temptation is to reach for a detector score to confirm the suspicion. The evidence in the two detection sections explains why that score adds little. TEQSA's own toolkit says an AI score "alone is insufficient to bring an allegation of misconduct". A false declaration is a better footing than a percentage, and it should be handled as what it is.
The question to put is not "was this written by AI". The question is "is the statement on the form accurate". That is answered by the things a declaration was designed to produce. One is a conversation with the student about the work, which CDU policy authorises. The other is the process material the brief told them to keep. The TEQSA toolkit lists a student's inability to answer questions about their own assignment as a clear signal, in contrast to a score. The same toolkit asks markers to look for evidence that disconfirms AI use as well as evidence that confirms it. A student who explains the work fluently has answered the question, whatever the software said.
Two cautions travel with this. The first is that the burden sits with the institution. The Australian authors quoted above state that "the burden of proof lies with the institution to establish misconduct, not with the student to disprove the allegation". A declaration does not reverse that; it gives the institution something specific to establish. The second is that process evidence is not self-proving in either direction. Current tools can "generate, edit, and iteratively refine a document over time, mimicking the human drafting process". The Sydney guidance notes that drafts and revision histories are easy to fabricate. A drafts folder supports a student's account; it does not settle it, and its absence settles nothing unless it was required in advance.
For VET the path is shorter and the standard is different. The assessor has to be assured the evidence is authentic before a competency judgement is made, and that judgement must be made by a qualified person. A declaration the assessor has reason to doubt means the assessor is not yet assured. Oral questioning and practical demonstration are the standard ways to become assured, and the decision stays with the assessor, not with a tool.
It does not prove which sentences a tool wrote. It does not prove intent, and a student who misunderstood a vague permitted-use rule has not lied. The cleaner the form, the smaller that second defence becomes, which is one more reason to write the form properly.
Where this leaves the form on your desk
A declaration is worth having when it asks for facts and names the task's learning outcome. It must say in advance what is to be kept, and warn that a conversation may follow. It is theatre when it asks for a promise. The difference is not length or legal tone; it is whether any line on the page could be wrong. The form does not catch anyone. It makes the honest student's position clear and the dishonest student's statement specific. That is the whole job, and it is enough.
The form also exposes the task underneath it. If the permitted and not-permitted fields cannot be filled, the problem is the assessment design, not the declaration. That is the subject of the assessment redesign section, where the same bar is used to decide what a task can still measure.
Sources
Claims on this page rest on Australian peer-reviewed research, regulator guidance and CDU's own policies. Quotations were checked against the primary text or an open repository copy on 22 August 2026. Commentary is identified where used.
- 'Where's the line? It's an absurd line': towards a framework for acceptable uses of AI in assessment, Corbin, Dawson, Nicola-Richmond and Partridge, Assessment & Evaluation in Higher Education 50(5), 2025 (open record at Macquarie University). The 19-student, 12-staff interview study and the finding that students build their own ethical frameworks.
- Talk is cheap: why structural assessment changes are needed for a time of GenAI, Corbin, Dawson and Liu, Assessment & Evaluation in Higher Education 50(7), 2025. The "enforcement illusion" and the specify-versus-verify gap. Full text was behind a publisher block at the time of checking; wording confirmed from the abstract and from two secondary summaries quoting the paper.
- The AI Assessment Scale Revisited, Perkins, Roe and Furze, arXiv, December 2024. The five levels and the decision to drop traffic-light colours. An academic framework, not a regulator instrument.
- Enacting assessment reform in a time of artificial intelligence, TEQSA, September 2025. The three task-level options and the list of secure assessment formats.
- CDU Academic Integrity Policy, effective 31 October 2025. The declaration responsibility, the verbal explanation provision and the process-evidence list, quoted.
- CDU Generative Artificial Intelligence Policy, date not shown on the page. The VET qualified-assessor rule, quoted.
- Practice guide: Assessment, ASQA, version 1.0, published 17 June 2025. The authenticity definition and the AI wording, quoted.
- Heads we win, tails you lose: AI detectors in education, Bassett and colleagues, Journal of Higher Education Policy and Management, online 29 January 2026. Advance notice of documentation requirements, the burden of proof, and the drafting-mimicry caveat, quoted.
- False flags and broken trust: can we tell if AI has been used?, Bridgeman, Liu, Brockley and Bassett, Teaching@Sydney, 30 October 2025. Institutional guidance: lack of drafts is not guilt; drafts can be fabricated.
- Detecting plagiarism of AI-generated text in student assessments, Guy Curtis, TEQSA Academic Integrity Toolkit, date unavailable. The "AI score alone is insufficient" statement, the disconfirming-evidence instruction and the clear-signals list.
- No TEQSA or ASQA instrument prescribing the wording of a student AI-use declaration was located at the time of checking. The rewritten sample above is a course draft and carries no regulatory standing.