Contract cheating, ghost-writing, and AI: why process records address all three

Process records address contract cheating and AI writing for the same reason: both forms of academic dishonesty appear in the evidence not as suspicious text, but as an absent or anomalous composition event. A ghost-written essay and a ChatGPT essay may both read perfectly; only the process record — what happened at the keyboard before the text existed — distinguishes them from each other and from a genuine student who simply writes well.

The problem that detectors cannot solve

Academic integrity discussions since 2022 have concentrated heavily on AI-text detectors. That focus is understandable, but it has obscured a problem that predates large language models by decades: contract cheating, in which students pay or commission another human to complete assessed work on their behalf.

Bretag et al. (2019), in a systematic review that is now the standard definitional reference in the field, define contract cheating as "when students use a third party to complete all or part of their academic assessment" — a definition broad enough to include essay mills, professional ghost-writers, and informal peer assistance. Lancaster and Clarke (2017) placed the behavior in institutional context: it is "the outsourcing of assessed student work," a supply chain that had already generated a professional industry well before AI became capable of producing academic prose.

The QAA's guidance document for UK higher education institutions, which surveyed the evidence, reported that "15.7% of students — roughly 1 in 6 — say they have paid for academic work at some point during their studies." That is not a marginal problem or a future risk. It was a large-scale integrity failure before ChatGPT was released.

Why detection tools fail on both problems

A ghost-writer who is good at their job produces well-formed academic prose that passes every stylometric test. An AI model asked to write in a student's established register may do the same. In either case, the artifact — the finished document — contains no detectable anomaly. This is the core failure surface of detection tools, and it affects AI detection and contract-cheating detection in the same way.

AI-text detectors are trained to identify statistical regularities in AI-generated prose: lower perplexity, more predictable token sequences, particular rhythmic flatness. A human ghost-writer produces none of those signals; detection tools trained on AI text will not flag it. Stylometric approaches — which ask whether this text resembles the student's prior writing — have had some success in contract-cheating cases, but they are only applicable when prior samples exist, require expert effort to apply, and become unreliable when a ghost-writer has been briefed on the student's style.

The OIA's casework note on AI and academic misconduct makes the evidentiary problem explicit from the institutional side: decision-makers relying on detector output alone face significant procedural exposure, because the output is a probability over the artifact, not a record of what happened. The note emphasizes that institutions must be able to demonstrate "reliable evidence" and that "the burden of proof rests with the institution." A detector score does not meet that standard unaided.

What both approaches share — and what limits them both — is that they examine the wrong object. The finished text is the output of a process. If the process was not the student's own, the text may still look exactly like the student's own. The evidence is not in the artifact.

What the process record shows instead

The distinction appears clearly when you shift from examining the product to examining the composition event. All three scenarios that look identical in the text produce different process signatures:

	Contract cheating	AI writing	Genuine authorship
Process record present?	No, or minimal post-receipt editing only	Partial — paste event(s) with thin or absent prior session	Yes — multi-session composition record
Composition duration vs word count	Near-zero typing time for the submission word count	Very short active writing time; time-to-submission dominated by prompt/paste	Extended duration consistent with the word count and task difficulty
Paste events	Single large paste at submission, or document uploaded wholesale	One or more bulk pastes of substantial content	Small pastes of quoted material only, if any
Revision rate	Absent (nothing to revise — the document arrived complete) or surface-only cosmetic edits	Low; pasted content rarely revised in depth	Characteristic revision pattern: deletions, rewordings, structural moves
Pause/burst pattern	No composition pauses; no burst structure	No composition pauses in pasted sections; may show brief typing around paste point	Pause/burst rhythm consistent with thinking and production alternating

This is not inference from the text; it is a direct record of behavior at the keyboard. Crossley et al. (2024) demonstrated that these behavioral signals are recoverable with high accuracy in controlled conditions: their study found 99% classification accuracy distinguishing composed writing from transcription using keystroke features alone, with a random forest model trained on features including pause rates before words and sentences, insertion counts, deletion counts, and revision structure. That result is for written-versus-transcribed, not for the specific contract-cheating scenario — and we should be precise about that — but the behavioral principle is the same: composition leaves a signature, and the absence of that signature is itself informative.

The evidentiary consequence: why contract cheating cases are harder

In practice, the AI-writing case and the contract-cheating case are not evidentiary equals. AI writing typically leaves partial process evidence: a student who uses ChatGPT must paste the output somewhere, and that paste event is usually recoverable. The composition session exists; it is just structurally anomalous — bulk content appearing without prior keystrokes, implausible typing speed for the final word count, minimal revision depth.

Contract cheating is harder because there is often no process record at all. A student who submits a file received via email, or who copy-pastes from a document shared on Google Drive, may leave nothing more than a single paste event in a brief session. The process record does not show anomalous composition; it shows no composition. That absence is evidence, but it is weaker than positive evidence of anomalous behavior — and institutions must be careful not to treat the absence of a record as proof of misconduct when the absence may equally result from a student who composed elsewhere and pasted in.

This is exactly why prospective process capture changes the landscape. If a student is notified before beginning an assignment that a session record will be created, and that record is sealed at submission, the contract-cheating strategy — write elsewhere, paste in — becomes the thing the record documents directly. The absence of composition in the record is no longer ambiguous; it is the record of what happened. The student had the opportunity to compose in the monitored environment and did not do so. That is meaningfully different from discovering after the fact that a submission arrived without a record.

Lancaster and Clarke (2017) anticipated this precisely: they argued that the most durable response to contract cheating would be "assignment designs and submission processes that make outsourcing more difficult or more visible." Prospective process capture is one such design.

What the literature establishes and what it does not

The Crossley et al. 99% accuracy figure is the strongest result in this area, and it is worth being precise about its scope. The study tested written versus transcribed text in a controlled experiment with recruited participants and specific tasks. It is not a certified false-positive rate for real-world misconduct adjudication, and it does not separately evaluate the contract-cheating scenario from the AI-writing scenario. It is strong evidence that composition and non-composition are distinguishable from keystroke data — which is the foundational behavioral claim this analysis depends on — but it should not be quoted as a product accuracy claim.

The stylometry literature has reported some success in contract-cheating identification when baseline samples are available, but that approach requires prior verified samples, expert application, and degrades when ghost-writers are briefed on style. It is a retrospective tool with high operational overhead, not a systematic institutional approach.

What the Bretag et al. (2019) definitional work and the QAA guidance establish on the institutional side is that contract cheating has scale, that detection is difficult with current tools, and that deterrence — making outsourcing visible before it happens — is part of any serious response. The process record approach is consistent with that institutional framing. It does not replace academic judgment; it creates conditions under which the evidence for that judgment is stronger.

Implications for institutional policy

The OIA guidance is instructive here. Its core message for AI cases — that detector scores alone do not meet evidentiary standards for misconduct findings — applies with equal force to contract cheating. Institutions that build misconduct cases on detector output face procedural challenge; institutions that build misconduct cases on verifiable process records occupy a substantially stronger position, because the record describes what happened rather than inferring what the text suggests.

This points toward several policy conclusions that follow from the evidence:

First, detection tools and process records address different problems. Detection tools examine the artifact after the fact; process records document behavior before and during composition. Treating them as competing solutions to the same problem misunderstands both.

Second, the contract-cheating and AI-writing problems are structurally similar at the process level — both appear as absent or anomalous composition — which means a single process-capture approach can generate evidence relevant to both, without requiring two separate detection systems.

Third, prospective disclosure matters. A student informed in advance that a session record will be created is deterred from the outsourcing strategy in a way that a student simply warned about detectors is not, because the deterrent is architectural rather than probabilistic.

Fourth, process records do not resolve every case. A student who composes entirely elsewhere and retypes the output character by character will produce a composition-shaped record. This adversarial behavior is possible; it is also substantially more effortful than outsourcing, which is itself a partial deterrent. No integrity tool eliminates all adversarial strategies, and process records should be presented to decision-makers alongside that caveat.

The argument for process records is not that they are infallible. It is that they address the evidentiary gap that both contract cheating and AI writing currently exploit — the gap between what a finished document looks like and how it came to exist. That gap is where the evidence lives.

Authmarks is built on this approach: a consented, session-level process record, sealed at submission and verifiable by institutions, designed to make the composition event available as evidence rather than leaving the finished text as the only thing there is to examine.

---

Related reading

• Temporal analysis and writing evidence
• Investigating AI writing concerns: a guide for educators
• Writing process evidence vs AI detection
• Using process data in academic integrity cases