Can an AI Detector Prove Who Wrote It? No.

01

What the score actually says

The strongest reason to reject detector-as-proof comes from the toolmaker’s instructions. Turnitin says its overall percentage represents qualifying prose its model identifies as likely AI-generated or likely AI-generated and then modified. That is a classification of text. It is not an identification of a writer.

There is another easy category error here. The AI percentage is independent of Turnitin’s similarity score. Nor is it a probability that an individual used AI. Reading “23%” as “23% chance this person cheated,” for example, assigns a meaning the report does not claim to provide.

Scope matters too. Turnitin defines qualifying text as prose sentences in long-form writing. It says the detector does not reliably detect AI-generated poetry, scripts, code or short-form and unconventional material such as bullet points, tables and annotated bibliographies. A document-level number therefore comes with boundaries before any investigation even begins.

None of this makes a detector pointless. A passage flagged by a model can provide a reason to inspect the work more closely. The problem begins when a signal designed to initiate scrutiny is treated as though the scrutiny has already happened.

02

A flag and a finding are different things

Turnitin’s warning is unusually clear: its model “may not always be accurate” and can misidentify human-written, AI-generated and AI-paraphrased text. The company says it should not be used as the sole basis for adverse action. Determining misconduct requires further scrutiny, human judgment and the institution’s applicable policies.

The product also contains a concrete concession to uncertainty. Turnitin says its testing found a higher incidence of false positives from 1% through 19%. It therefore no longer surfaces a numerical score or highlights in that range, displaying an asterisk instead. Reports created before July 8, 2024 may still contain those lower numerical scores.

That safeguard does not mean every score at 20% or higher is correct. Turnitin separately states that false positives are possible. Equally, its guidance does not establish that every flag is wrong. The useful boundary is simpler: a model result can support a review, while authorship or misconduct requires evidence beyond that result.

A percentage has the visual authority of measurement, which makes the distinction easy to miss. But measurement and conclusion are separate jobs. Detector output answers what a model classified within eligible text. An authorship finding asks who created the work and, often, whether their conduct violated a particular rule.

03

The paper trail matters more than the aura of precision

Practitioner accounts show why the issue becomes messy in practice, although they cannot establish detector accuracy or false-positive rates. In one Reddit post, a teacher said colleagues stress-tested an unnamed detector with writing they had produced themselves and saw it flagged. The teacher characterized detector results as circumstantial evidence needing additional proof.

That teacher described a more useful source of context: a visible writing process. Assignments were structured so the teacher could examine how drafts developed, and students were advised to preserve editing history. Version history is not infallible proof of authorship either, but it can add evidence about how a document came together.

A separate graduate student reported being told that Turnitin gave a final essay a 23% AI score despite the student denying AI use. The student later showed annotated source material and Google Docs change history, reporting no official penalty but a lowered grade. That is one person’s account, not a controlled test.

These stories should therefore be used for what they demonstrate: the practical stakes and the kinds of evidence people reach for when a score is disputed. They do not tell us how frequently detectors err. They do show why resolving authorship from one percentage can leave important evidence outside the frame.

04

What a detector can legitimately do

There is a useful middle ground between “the detector proved it” and “detectors are worthless.” Turnitin explicitly designed its report to help educators identify text that might have been prepared with generative-AI or text-altering tools. “Might” is doing important work: the report can identify material for closer examination.

That examination should respect the detector’s coverage. A mixed document may contain prose alongside tables, bullets, code or other material the model does not reliably detect. The percentage applies to qualifying text, so reviewers should first understand what the tool actually evaluated before interpreting the result.

For businesses, schools and other organizations, the transferable lesson is procedural. Separate automated screening from the decision that follows. A flag can prioritize attention; the decision should rest on the relevant policy and the fuller evidentiary record. That preserves the detector’s useful role without quietly promoting it into an authorship oracle.

The same caution applies in the other direction. These sources do not establish that every detector finding is false or that screening has no value. They establish the narrower point needed to bust this myth: classification alone does not identify a document’s author or independently establish misconduct.

05

How to do it

A defensible review starts by deciding what kind of question you are answering. “Did a model flag this prose?” is different from “Who wrote this?” and different again from “Did this person violate our policy?” Keeping those questions separate prevents an automated signal from becoming a verdict by default.

Process evidence can include contemporaneous drafts, research notes and document version history. Such records are context, not magic receipts: none necessarily proves authorship alone. The writer’s explanation and the substance of the work can also be considered alongside whatever procedures the organization requires.

Most importantly, document the basis for the eventual decision independently of the detector percentage. That makes clear whether the finding rests on a fuller record or merely repeats the model’s classification in official language. A detector can be part of that record without being allowed to become the whole record.