The widespread adoption of AI text detectors in academic settings since late 2022 has created an unrecognized problem: rigorous academic writing is increasingly penalized as “AI-generated.” Universities and journals have rapidly deployed commercial detection tools such as Turnitin, GPTZero, and Originality.ai despite limited evidence of their reliability regarding scholarly texts. This study examines the empirical reliability of AI text detectors on academic writing and investigates the epistemological and ethical implications of delegating authorship verification to algorithms. We conducted a systematic review of peer-reviewed studies published between 2022 and 2025 that met strict inclusion criteria, namely quantitative accuracy metrics on academic corpora, peer-reviewed publication, and reproducible methodology. Data were synthesized from 21 verified sources including controlled experiments, meta-analyses, and detection limit studies, while philosophical analysis drew on creativity theory, philosophy of science, and the Chinese Room argument. On academic texts, false positive rates vary substantially across detectors (range: 2–41%, mean: 14–22%), representing a three- to sixfold increase compared to news corpora, whereas false negative rates reach 44–47% with minimal human editing and exceed 60% under combined obfuscation. Inter-detector agreement is weak (Cohen's κ = 0.35–0.44), which makes single-detector decisions unreliable, and no systematic cross-linguistic studies meeting inclusion criteria exist, representing a critical research gap. Theoretically, large language models can produce combinatorial and limited exploratory novelty but cannot generate transformational novelty, that is, the capacity for paradigmatic rupture that defines human authorship. Current AI detection systems are therefore epistemologically flawed and technically unreliable for high-stakes academic decisions. We propose a three-level human-centered framework (screening – expert review – appeal) that restores human agency, since academic integrity requires trust, transparent procedures, and human judgment rather than algorithmic suspicion.