When a university uses a commercial detector to decide whether a student cheated, the implicit assumption is that the software can tell human writing from LLM output. A controlled study on published English abstracts dismantles that assumption. Researchers compared four disciplines and two time windows, 2013–2015 and 2023–2025, separating original texts from those edited with a light “refine abstract only” intervention. That operation, a proxy for guideline-compliant AI assistance, is flagged as suspicious in 64 to 80 percent of cases depending on the detector, Pangram or GPTZero. Unmodified 2023–2025 originals are flagged in only 9 to 15 percent of cases, with non-STEM rates far above STEM.

The issue is not that detectors miss some cases: it is that the risk order is inverted. Someone who uses AI transparently to refine an academic abstract is more likely to come under suspicion than someone who starts from a generated draft and modifies it with tools designed to evade detection. The study states this explicitly: honest AI editing carries a higher sanction risk than humanizer-assisted evasion. That changes incentives in a profound way. The rational student or researcher is not pushed to follow the rules but to invest in obfuscation. The detector, born as a deterrent, becomes an accelerator of adversarial techniques.

For institutions the cost is not only reputational. If a commercial software score is treated as standalone evidence of misconduct, a governance problem opens up: who can verify the threshold used, the false positive rate, the model’s training criteria, the data provenance? Most commercial detectors are opaque services. They do not allow inspection of the scoring logic. In regulated contexts or disciplinary proceedings, this is a structural flaw, not a side issue. Universities need auditable procedures, not scores arriving from an external API without context. That is exactly the kind of problem pushing some organizations to evaluate self-hosted alternatives for sensitive controls: not out of isolationism, but to document and reproduce decisions.

It is no coincidence that the study uses proxy human/AI labels at a tau=0.50 threshold. This signals the fragility of any binary classification in a field where assisted writing is now a continuous practice, not an exceptional event. The boundary between “written by a human” and “generated by an LLM” dissolves when tools intervene on individual sentences, structure, and vocabulary. A detector making a hard decision on a final text is monitoring a process it cannot see. It is like judging the legitimacy of a photograph by looking only at the final image, without knowing whether the negative was developed in a darkroom or retouched digitally. The difference lies in the process, not the output.

There is also a second-order effect on academic policy. If a light “refine abstract only” edit is indistinguishable from a violation, the rules that allow it become unenforceable. Universities face a fork: ban every generative intervention and impose controlled workflows, or shift attention from hunting synthetic text to evaluating the process, revisions, sources, and demonstrated competence. The third path, trusting vendor scores, is the weakest because it moves disciplinary responsibility onto a black box.

For those evaluating on-premises deployment, there are well-known trade-offs between data control and operational complexity; AI-RADAR covers some of these scenarios at /llm-onpremise. Here the point is not necessarily to bring the plagiarism detector in-house, but to recognize that the opacity of an external service is a hidden cost: without access to the inference logic and calibration data, an institution cannot defend a disciplinary decision in a documented way. Data sovereignty is not only about servers; it is about the possibility of audit.

The study’s operational conclusion is blunt: detector scores should not be used as standalone evidence of misconduct. This is not a position against AI, but against blind delegation to software that cannot distinguish transparent assistance from a fully generated text. If the goal is academic integrity, the first step is to stop rewarding evasion and punishing cooperation.