Why AI Detection Does Not Prove Authorship

AI-text detectors estimate whether writing resembles examples associated with a model. They do not observe how a document was produced and cannot establish who wrote it, which tools were used, or whether the use complied with an assignment.

What detector results mean

Most detectors classify patterns in text. Some use statistical predictability or stylometric features; others use classifiers trained on samples of human and model-generated writing. A result such as “70% AI” is a product-specific score, not a measured percentage of sentences written by a machine.

Performance changes with the model, subject, genre, length, language, editing, and decision threshold. A tool tested on long, unedited English prose may behave very differently on a short answer, translated work, code, technical writing, or text edited with an embedded writing assistant.¹ A 2023 journal study tested 14 detectors on human-written, translated, generated, and edited texts and found that the tested tools were not consistently reliable across those conditions.²

Even apparently low false-positive rates can cause harm when applied across a large cohort. A detector can also miss model-generated text after ordinary editing or paraphrasing. Independent testing found substantial variation among tools and large drops in detection accuracy after simple modifications; the authors did not recommend the tested tools for determining integrity violations because of accuracy limits and the risk of false accusations. They distinguished this from possible non-punitive uses that support learning.³

Newer evaluations can find stronger results for particular tools. A 2026 study found that one detector performed well on its constructed dataset, while still advising against using detector results as sole evidence in high-stakes decisions. Its separate analysis of real theses lacked known authorship ground truth, so the flagging rate cannot establish the prevalence of misconduct. Results belong to the tested tools, models, and text conditions; they do not establish authorship or whether an assignment rule was broken.⁴

Why the authorship question is harder than two labels

Student writing can include brainstorming, translation, grammar correction, autocomplete, generated passages, human revision, peer feedback, or no AI at all. “Human” and “AI” are therefore often not mutually exclusive categories. The educational question is whether the student met the stated learning outcome and followed the published rules.

Bias and unequal error rates are also serious concerns. A 2023 study found that the tested detectors frequently misclassified non-native English writing samples as AI-generated.⁵ This finding concerns those samples and systems, not every language, genre, accessibility aid, or current detector. Do not infer misuse from multilingual, concise, formulaic, or accessibility-supported writing. No student should have to change a legitimate writing style merely to avoid a proprietary score.

Respond to a concern fairly

Design assessment around evidence of learning

The better long-term response is not surveillance but clearer assessment design:

• state whether the task is AI-free, AI-limited, AI-permitted, or AI-integrated;
• use staged drafts, annotated decisions, source checks, or reflection when those activities genuinely support the learning outcome;
• include oral or performance follow-up when students must demonstrate that capability;
• assess application to specific course material rather than generic fluency;
• provide equivalent, accessible ways to demonstrate learning; and
• avoid collecting complete AI transcripts or writing histories by default.

See Setting course and assignment AI rules for copyable policy blocks and Assessing learning in the age of AI for assessment redesign.

References & Footnotes