A responsible scanner should be willing to say “inconclusive”. Scores are evidence produced by a method under assumptions—not a mathematical proof of authorship.
Human writing can look model-like. Highly structured technical prose, formulaic academic language, non-native English, edited corporate copy and standardized documentation can all produce patterns similar to material in AI-training datasets.
AI-generated text can fall outside a classifier’s training distribution. Editing, translation, short excerpts, unusual prompts and newer models can change the statistical profile.
The deep classifier used by AI Doc Scan is primarily trained for English and performs best on domains similar to its training data. The app therefore avoids presenting its English model as authoritative for Serbian/BCS or other languages.
There is less evidence in short text. This affects both generic classification and statistical watermark detection. AI Doc Scan avoids forcing very short standalone pages through the deep classifier and reports diagnostic-only results when appropriate.
When a person substantially rewrites generated material, the final text may no longer resemble the original model distribution. That is not a flaw in the concept of authorship—it is precisely why AI participation, provenance and final human responsibility are separate questions.
A score of 70 does not mean “AI wrote 70%”. It means the selected model produced an AI-associated probability or aggregate signal under this scanner’s methodology.