Direct answer
Teachers detect AI writing by combining three layers of evidence rather than trusting any single signal: close reading for clusters of stylistic markers (uniform sentence rhythm, generic transitions, confident vagueness, thin local detail), a Turnitin AI writing report read as an indicator rather than a verdict, and process evidence such as drafts, version history, or an oral explanation of the work [1]. No stylistic marker is reliable on its own, and Turnitin's own guidance stresses that detection results must sit alongside professional judgement and institutional policy [1]. The defensible sequence is therefore: read critically, check the report, then verify with the student before any concern is formalised.
What Are The Reliable Signs That A Piece Of Student Writing Was Generated By AI?
There is no single tell. Published teaching guidance converges on the point that AI-generated prose is identified by clusters of weak signals, not by one giveaway phrase, and that any marker loses value the moment a student edits the draft [2]. A paragraph that reads smoothly but says little is suggestive; six paragraphs that all read that way, with no local specifics, is a pattern worth investigating. Treat stylistic evidence as a reason to look closer, never as proof.
The markers most often cited in teaching-centre guidance are consistent: unusually uniform sentence length and rhythm, transitions that signpost structure without adding argument ("Furthermore, it is important to note that…"), confident generalisation where a specific citation or example would be expected, and a striking absence of the idiosyncratic errors, digressions, and register shifts that human drafting produces [2]. Human writing is uneven — that unevenness is often the most useful signal available, because it is genuinely hard to fake consistently across a whole assignment.
Two cautions matter here. First, formulaic writing is not the same as AI writing: heavily scaffolded essays, EAL writers working from templates, and students who have been drilled in a rigid structure can produce prose that trips every stylistic heuristic [2]. Second, markers disappear after revision — a student who rewrites an AI draft in their own voice leaves prose that reads human, because it partly is. That is precisely why guidance recommends weighting process evidence above prose alone [2].
The practical test is triangulation. If the prose looks generic, the source list is vague or unverifiable, and the student cannot explain a key paragraph in conversation, you have a pattern. If only the prose looks generic, you have a hypothesis. Record what you noticed and where, because a claim you cannot point to on the page will not survive an appeal.
How Does Turnitin's AI Writing Detection Actually Work, And How Should Teachers Read The Report?
Turnitin's AI writing detection works by segmenting long-form prose and evaluating how predictable the text is at sentence level, then reporting the proportion of qualifying text that the model classifies as AI-generated [3]. The report is built to function as an indicator that supports an educator's judgement — Turnitin's own framing is that results are meant to inform a conversation and a decision, not to substitute for either [3]. Read it as one data point in a case file, not as an outcome.
The detail teachers most often misread is the asterisk. When the AI score displays as *% rather than a number, the result sits below Turnitin's 20% confidence threshold — a low-confidence signal, not a clean bill of health and not a finding of misconduct [3]. Equally, a high percentage is not a confession: it tells you which passages the model considered predictable, and those passages are exactly where your close reading should start. Coverage matters too, because the score only reflects the text the detector could actually evaluate [3].
Three reading habits make the report far more useful. First, open the flagged passages and read them yourself before forming a view — the report highlights text, and your judgement about that text is the actual evidence. Second, check what the score excludes: short submissions, reference lists, and heavily quoted material may sit outside the analysed text, so a percentage can describe less of the essay than you assume [3]. Third, compare the flagged passages against your own stylistic notes; when the report and your reading point at the same paragraphs, you have corroboration rather than a single automated opinion.
It also helps to know what the report cannot do. It cannot tell you who wrote the text, whether a tool was used with permission, or whether the assignment brief invited AI assistance in the first place — those are policy questions, not detection questions [3]. Keeping that boundary clear protects you in moderation meetings and makes your written feedback much harder to challenge.
What Should A Teacher Do When They Suspect AI Use — And How Can They Verify Before Raising A Concern?
Start with your institution's policy, not with the report. Academic integrity guidance is consistent that the first question is what the institution permits for that assessment, because the same text can be entirely acceptable under one brief and a breach under another [4]. Confirm whether AI assistance was allowed, what disclosure was required, and which escalation route applies before you put anything in writing.
Then gather process evidence, which is the strongest material available to you. Drafts and version history, planning notes, annotated bibliographies, in-class or handwritten writing samples, and a short oral conversation about the student's sources and choices all test understanding in ways prose cannot [4]. Ask the student to explain a specific paragraph — why that example, where that source came from, what they would change. A student who wrote the work can usually do this, even imperfectly; a student who did not often cannot [4].
Frame the conversation as a request for explanation rather than an accusation. Guidance on handling suspected AI use recommends neutral, specific language — naming the passages that concerned you and inviting the student's account — because an accusatory opening forecloses the explanation you actually need and creates appeal risk if you were wrong [4]. Take notes, keep the exchange factual, and follow the documented escalation route rather than resolving it informally on your own.
Finally, calibrate your confidence before you act. A single stylistic impression is not enough to raise a formal concern; a Turnitin report plus your own annotated reading plus an inability to explain the work is a defensible case [1][4]. Where the evidence is thin, the proportionate response is a developmental conversation and a note in your marking record — not a misconduct referral.
Reading a Turnitin report is far easier once you have seen a real one, with its cover page, score band, flagged passages, and similarity summary laid out exactly as your institution's system presents it. That is what turnitin0 was built for: teachers and students can run a pre-submission check on a draft and receive the same two PDFs your LMS produces — the AI detection report and the similarity report — in under fifteen minutes in 98% of cases, without the file entering Turnitin's student paper database. Seeing the report before you sit down with a student turns an abstract percentage into a concrete, quotable piece of evidence you can point to on the page.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Is a high Turnitin AI score enough to accuse a student?
No. Turnitin presents the AI writing report as an indicator that supports educator judgement, not as a finding of misconduct [1]. A defensible concern combines the report's flagged passages, your own annotated reading, and process evidence such as drafts or an oral explanation.
What does *% mean on a Turnitin AI report?
It means the score falls below Turnitin's 20% confidence threshold, so the system shows an asterisk instead of an exact percentage [3]. Treat it as a low-confidence signal: it is neither proof of AI use nor a guarantee that the text is original.
Can human writing be falsely flagged as AI-generated?
Yes. Formulaic, heavily scaffolded, or template-driven prose — including work by EAL writers — can resemble AI output stylistically, and detection markers weaken further once a draft is edited [2]. That is why guidance recommends weighing process evidence above prose alone.
Should I tell students which detection tools I use?
Transparency generally helps. Stating your institution's AI policy, the tools used, and how reports are interpreted reduces disputes and makes a conversation about a flagged passage far easier to hold [4].
What evidence should I keep if I raise a concern?
Keep the flagged passages with your own annotations, the Turnitin report, drafts and version history, notes from your conversation with the student, and a record of the policy clause you are applying [4]. Specific, dated, page-level evidence is what survives moderation and appeal.