What Students Are Actually Doing

The story that travels, and the number that doesn't

In July 2026 a Brown University economist told Inside Higher Ed that he suspected most of his class had used AI to earn near-perfect marks on a take-home midterm. The story moved quickly, because it fits a shape we already carry: the sense that undisclosed AI use is everywhere, that honest work is now the exception, and that a marker's suspicion is usually right. Stories like it shape how a paper is read before the reading starts.

That intuition is worth holding at arm's length, for a simple reason. A single class, a single marker's impression and a percentage on a detector report are all anecdote in the sense that matters here: none of them measures how common AI use actually is, and none can be checked. What can be checked is the survey evidence, and it tells a more modest and more honest story than the one that travels. This section is that evidence, and an account of exactly how far it reaches.

What the Australian surveys actually measured

The best-documented national picture of Australian student use comes from a survey by Newell and Dahlenburg, presented at the ASCILITE conference in 2024. Its headline findings are these.

Two things are worth noticing before the caveats. The first is that a third is not a majority; the confident sense that "everyone is using it" is not what the strongest Australian survey found, at least as at its 2024 fieldwork. The second is that the same students who use these tools are, in large numbers, uneasy about what they do to the value of a qualification. The cohort in this data is not a wave of enthusiastic cheats; it is a group of people using a new tool and worrying about it at the same time. That ambivalence is the part the anecdotes leave out.

Why the numbers are soft

Every figure above should be held loosely, and it helps to be precise about why. The point is not that the research is poor; it is that this is genuinely hard to measure, and the honest position is to know the shape of what we do not know.

The survey is preliminary and modest in size, with 399 respondents, so it establishes that a pattern exists rather than pinning its exact proportions for the sector. It relies on self-report about a behaviour that is, in many units, against the rules, and self-report understates behaviour people have reason not to admit. It carries no breakdown by international or domestic status, by discipline, or by English-language background, so it cannot be narrowed to any particular group. And it is aging: adoption of these tools has moved fast enough that a 2024 figure is a photograph of a receding moment, almost certainly an undercount of the position today.

The newer figures that circulate should be held more loosely still, not less. Australian commentary in 2026 has carried headline claims as high as "almost 80 per cent of students now use AI", but the primary survey behind that number is not publicly documented in a form a reader can open and inspect, which is exactly the weakness this section is about. A separate 2026 figure, that more than half of Australian assignments showed some AI, comes from a detection vendor's own data on submissions, not from a survey of students, and a detection score is not a finding of misuse. A bigger number is not automatically a better one; a number you cannot trace is one to quote with care, if at all.

Staff are mid-adoption too

It is easy to write this as a page about students and a new habit. The evidence does not support that framing. A survey of 3,421 staff across 17 Australian universities found that 71 per cent had used generative AI for their university work in 2024, with the remainder holding back, often citing low confidence and limited familiarity rather than principle. Adoption among academic staff ran higher than among students in the same period.

This matters for how the whole workshop is pitched. The sector is not divided into students who use AI and staff who police it; it is a whole population learning the same tools at once, at different rates and with different rules applied to each group. Keeping that in view guards against the reflex that reads student use as a moral failure and staff use as productivity, and it is a more accurate description of where CDU actually sits.

International students, and the inference you cannot make

International students are a large and valued part of CDU's cohort, and they are also the group most often at the centre of unspoken assumptions about AI use. The evidence here is worth stating carefully, because the honest reading of it is a set of limits rather than a finding.

CDU's own 2024 submission to the Review of Regional Migration Settings records international students as 17 per cent of the total student population and 27 per cent of degree students, with India, Nepal and Bangladesh the primary source countries. That figure establishes who is enrolled. It says nothing about how anyone uses AI, and no located source measures generative-AI use by CDU's international students, or by CDU students broken down by international, domestic or language background, at all. The behavioural data simply does not exist at that resolution.

What the Australian evidence does address is where suspicion tends to land, and it points away from nationality. The academic-integrity synthesis hosted in TEQSA's toolkit finds that misconduct is predicted more by English-as-an-additional-language status and by unfamiliarity with local academic expectations than by international-student status, and that the strongest predictors of all are a lack of understanding of the rules and finding the work too difficult, rather than where a student is from. The same source carries a nuance that has to be reported alongside it: international students do continue to be over-represented in recorded misconduct cases, and some larger contract-cheating studies have found higher rates among them. The source's own reconciliation is that language features and unfamiliarity, not nationality or culture, drive both the elevated suspicion and part of the measured difference, and that the appearance of misconduct can itself reflect implicit bias. Both halves belong together, and neither resolves into a licence to scrutinise a student because of where they come from.

There is a further trap that this section connects directly to the detection material. AI-writing detectors misclassify writing by non-native English speakers as AI-generated far more often than they misread native writing. In the most-cited study, seven detectors wrongly flagged essays by non-native writers at an average rate of 61 per cent, while classifying native-speaker essays almost perfectly. That is a property of the tools and of second-language writing style, not a measurement of anyone's behaviour, and it means the group about whom staff most readily form suspicions is the same group the software most readily flags in error. A commercial survey has separately reported that international students expect and use AI study tools more than domestic students, but that figure comes from a vendor's own panel, aggregates all international students with no country detail, and has not been independently confirmed; it is not a basis for anything here.

What this leaves a marker with

The measured picture is genuinely useful, and it is narrower than the anecdotes. A meaningful share of students use generative AI, the real share is probably higher now than the surveys caught and is best treated as unknown at the individual level, most students are ambivalent rather than cavalier, staff are adopting the same tools in parallel, and nothing in the evidence lets anyone read AI use off a cohort, a language background or a style of writing.

That last point is where this section hands over. If the population data cannot diagnose the paper in front of you, the weight falls on what a detector score and an integrity process actually can and cannot establish about a single submission. That is the subject of the two companion sections, How Detection Technology Works and Detection as Evidence, Not Verdict.

Sources

Figures on this page are drawn from Australian survey and regulator-hosted sources wherever possible; international evidence is labelled as such, and commercial or secondary reporting is identified where it is used or set aside. Access date 21 August 2026.

• The national student survey of generative AI use among Australian university students: preliminary findings, Newell and Dahlenburg, ASCILITE 2024. The one-third use figure, ChatGPT at 53 per cent, and the roughly 80 per cent devaluing-my-degree concern; preliminary, n=399.
• Charles Darwin University submission to the Review of Regional Migration Settings, July 2024. International students at 17 per cent of total and 27 per cent of degree enrolments; India, Nepal and Bangladesh as primary source countries.
• An overview of culture and academic integrity: myth-busting the notion that international students are more likely to engage in academic misconduct, Guy Curtis, TEQSA Academic Integrity Toolkit, date unavailable (cites sources to 2025). EAL rather than nationality as the operative variable, and the over-representation counter-nuance.
• GPT detectors are biased against non-native English writers, Liang et al., Patterns, 10 July 2023 (open text at arXiv 2304.02819). The 61 per cent average false-positive rate on non-native essays; United States data.
• 71 per cent of Australian uni staff are using AI, The Conversation, October 2024. Staff adoption across 3,421 staff at 17 universities; low confidence among non-adopters. Author summary of the underlying survey.
• Brown professor suspects most of his class used AI to cheat, Emma Whitford, Inside Higher Ed, 8 July 2026. Used only as a labelled United States anecdote.
• Newer 2026 Australian figures referred to and held at arm's length: an "almost 80 per cent of students use AI" claim (The Conversation, 17 March 2026), whose primary survey is not publicly documented, and a "more than half of assignments showed some AI" figure (The Conversation, 16 July 2026) drawn from Turnitin detection data on submissions rather than a student survey. Both are secondary or vendor-based and are not relied on here.