White Paper July 2026
Heads We Win, Tails You Lose: Why AI Detection Cannot Be the Answer to Academic Integrity
Why a detector score cannot meet the standard of proof a misconduct case requires.
This is relevant to you because:
- You're running a detection tool, or being asked to buy one, and want its accuracy claims tested.
- You run misconduct processes that must meet a real standard of proof, and need to know what qualifies.
- You're aware a detector flag falls hardest on international and second-language students, regardless of wrongdoing.
- You're accountable for GDPR and EU AI Act compliance, and suspect staff are already using free detection sites.
Abstract
AI detectors report a probability that text resembles machine-generated writing. They do not establish authorship, and this paper argues that gap is structural. Drawing on peer-reviewed analysis and independent testing of the leading commercial tools, it identifies three failures: no ground truth against which a flag can be checked, a base rate fallacy that makes the same detector reliable in one cohort and worse than chance in another, and false positive rates that climb sharply for second-language writers.
It then traces the institutional consequences, including an inverted burden of proof and explainability duties under the EU AI Act, and sets out why UNIwise builds WISEflow Originality on verifiable similarity detection rather than probabilistic inference.