Artificial learning: how sessional academics view contract cheating and digital academic misconduct in assessment

Synthesis: Andriamora and Andriamora (2026) asked twenty experienced sessional academics across four Sydney institutions how they read student misconduct in the papers they mark, and got a picture in which traditional contract cheating is in decline while Generative AI use is treated as a fact of the environment rather than a deviation from it: seventeen of twenty participants described widespread AI use, and one estimated 25 of 70 recently marked papers were fully AI-written. Their detection practice is craft knowledge read off the page — fabricated references (nineteen participants), disconnection from course materials (eighteen), rubric misalignment (fifteen) — with no instrument behind it, since at one institution the AI detector had been switched off for flagging grammar checkers. Participants traced the drivers to institutions as much as to students: outsourced essay formats, time pressure, fear of failure, and non-native language underpin the behavior, and offshore services still sell a 2,000-word paper for $25. What they asked for was not better policing but assessment redesign, spoken and invigilated formats, and consistent policy — with the frank admission that sessional staff carry the integrity load without the standing to change any of it.

Key Findings

1. AI use has displaced contract cheating rather than joined it. Seventeen of twenty participants described widespread generative AI use, one estimating 25 of 70 recently marked papers were fully AI-written, while most saw outright ghost-writing in decline.
2. Fabricated references are the most reliable cue markers still have. Nineteen of twenty participants reported invented or incorrect citations, one finding all 21 references in a paper were fake and another encountering DOIs that do not exist.
3. Detection is untrained craft, and the instruments are distrusted. At one institution the AI detector had been disabled for flagging grammar-checking software, leaving participants to estimate prevalence without clear guidance from the university.
4. The drivers sit in assessment design and student circumstance. Fifteen participants blamed easily outsourced essay formats, fourteen named time pressure, thirteen fear of failure, twelve non-native language, and eighteen pointed to offshore services selling a 2,000-word paper for $25.
5. Rules that differ between units leave students unable to comply. Fifteen participants described confusion about what was permitted, one citing an institution that banned AI in one unit while encouraging it for research in another.
6. The recommendations end at curriculum and learning outcomes, not policing. Participants proposed in-class presentations, monitored check-ins, case studies on local data and rubrics that demand course content, and several argued for redefining what learning outcomes should measure.

What twenty markers actually see

The design is deliberately narrow. Twenty casual academics working across the University of Sydney, Central Queensland University, La Trobe University and Torrens University were recruited purposively through the principal researcher's professional networks, all with more than 10 years of cumulative teaching, tutoring and marking experience; nine taught in business management or MBA programs, six in accounting or finance, five in information and communication technology, and twelve were working across three or more institutions at once. Semi-structured interviews of 45–60 min ran from December 2024 to May 2025 and were analyzed thematically. That breadth of simultaneous employers is what makes the accounts worth reading: the same person can compare four integrity regimes in one week.

On prevalence, the participants mostly reported a substitution rather than an addition. Contract cheating was still present — one participant suspected 30 of 100 recently marked papers, another described international students offshoring work to ghost-writers in India or Bangladesh at low prices — but the dominant account was that AI had absorbed the demand. Several framed the technology as an inevitability that prohibition cannot reverse, comparing it to the computer or the internet rather than to a cheat sheet.

The cues that survive when the detectors fail

Asked how they actually identify work that is not the student's, participants described reading for the absence of course context. Eighteen named disconnection from lecture material and unit content; fourteen named generic or outdated sources and examples from the wrong national context; fifteen named submissions that ignored the rubric entirely. Fourteen also reported comparing the writing against the student's own classroom performance, a form of continuous Assessment that a marker with a sustained relationship has and a rotating anonymous marker does not.

For AI-written work the cues shifted to style and to citation. Seventeen participants cited structural repetition — the same transitional phrases, a uniformly polished register, a lack of logical build — and twelve described a shallowness in reflective tasks where AI cannot produce the personal particularity the genre requires. The most concrete indicator was invented sourcing: nineteen of twenty participants had found fabricated or incorrect references, one reporting all 21 references in a paper as fake, another describing non-existent DOIs. Participants were explicit that none of this is training; it is experience, described by one as "just a gut feeling," and they expected it to decay as generation improves. Their own summary was that "we're always a step behind."

Why the drivers are structural

The contributing factors split along the same line. On the student side, twelve participants named studying in a non-native language, thirteen named fear of failure or of being wrongly accused of plagiarism, and fourteen named disengagement from learning as such — students seeking the credential rather than the competence. On the institutional side, fifteen blamed curriculum and assessment that invite outsourcing, and eighteen pointed at a mature offshore market that makes the transaction cheap and, in students' eyes, undetectable.

Between those poles sits the policy environment, and it is where the paper is sharpest. Fifteen participants reported that students do not know what is permitted, and the illustration is not a student failing to read the rules but a faculty with no coherent rule to read: one unit bans AI outright, another encourages it for research, and students are left guessing. Time pressure (fourteen) and bare convenience (seventeen, particularly for students with weak academic skills) do the rest. Participants doubted that prevalence could be estimated at all without better guidance, and described reporting suspicions into a process whose decisions they never see.

From detection to design

The recommendations move quickly from surveillance to pedagogy. On design, participants proposed in-class presentations on the reasoning that students may outsource the slides but must deliver the talk themselves, monitored weekly check-ins, case studies built on local or live data, tasks that explicitly incorporate AI while bounding it, and rubrics that require direct use of course content and theory. Several endorsed invigilated examinations while noting that phones still reach them, which connects their accounts to the wider problem of Remote Proctoring.

On support, they wanted consultations, scaffolded tasks and continuous feedback, plus the structural change of letting markers assess their own students instead of rotating anonymous graders. On policy, they asked for consistent institutional messaging and for awareness campaigns that borrow the reach of the cheating providers themselves. The last strand is the largest: several participants argued for redefining learning outcomes around critical thinking, ethical reasoning and real-world capability, on the grounds that assessment designed before generative AI cannot measure judgment in a world that has it. That connects to the corpus's work on Reducing AI Misuse and on what Critical Thinking assessment should look like. Participants were clear-eyed about the obstacles: casual staff have no curriculum authority, oral formats are resource-intensive in a massified system, and templates persist.

What this means for practice

• Instructors. Read for sourcing and context rather than style alone: check whether the reference list exists and whether the paper engages the unit's own materials, and tell students explicitly that you do this — nineteen of twenty participants found fabricated references, and that cue survives model improvements better than phrasing.
• Administrators. Audit whether your AI rules actually reach students. If two units in the same program permit and prohibit the same behavior, the compliance gap you are measuring is a communication failure, and the participants' evidence is that students respond to ambiguity by guessing rather than asking.
• Assessment designers. Pair any high-stakes written submission with a spoken or in-class component that requires the student to defend the work, and build the rubric to demand unit-specific content — the cheapest structural change the participants describe, and the one their detection cues already reward.

Limitations

• Twenty participants, all in Sydney, purposively recruited through the principal researcher's networks with the researcher's insider status acknowledged; a different network would sample different institutions and disciplines.
• Self-reported perceptions only: no student accounts, no institutional misconduct records, and the prevalence estimates are recall-based figures (one participant's "30 of 100") rather than audited counts.
• Interviews ran December 2024 to May 2025, a single cross-sectional window in a period when institutional AI policy was changing month to month.
• The paper reports no participant-level demographic beyond counts, so differences between the business, accounting and ICT groups cannot be examined.
• The authors disclose AI assistance with language editing only; the coding and analysis is theirs.

Citation

Andriamora, L., and L. Andriamora (2026). Artificial learning: how sessional academics view contract cheating and digital academic misconduct in assessment. Assessment & Evaluation in Higher Education, 1–17.