Dramaturgies of Deception: AI Humanizers and the Performance of Legitimacy in Higher Education Assessment

Synthesis: Presenting the first exploratory, systematic investigation of AI humanizer websites, Roe and colleagues frame these services through Goffman's (1959) dramaturgy and Multimodal AI critical discourse analysis, cataloguing 55 humanizer sites and deeply analyzing three. Findings show humanizers are readily available, offer free and premium tiers, and perform similar functions—deleting misconduct from the discourse, reframing humanization as a rational response to flawed detection, and appealing to mystification through advanced technology and implied institutional endorsement. The authors argue humanizers are a diagnostic signal in a feedback loop of performative assessment whose disruption requires structural assessment reform rather than technological solutionism.

Key Findings

1. Humanizers are readily available. A systematic search surfaced 55 humanizer sites (e.g., targeting queries like "bypass AI detection," "pass Turnitin AI detector"), and the top-ranking sites held high ecological validity as the first results users would encounter.
2. Similar commercial structures. Sites consistently offered both free and premium paid services, pricing tiers, free trials, and even affiliate-marketing programs, performing near-identical functions.
3. Trust and authority artifacts. The three analyzed sites deployed corporate and university logos, detector logos and screenshots, user statistics, satisfaction ratings, testimonials, and social-media presence to construct legitimacy.
4. Mystification strategies. Opaque references to proprietary "deep learning" or "advanced algorithms," scale claims (e.g., "over 1 billion words a month," "trillions of parameters"), one-click framing, and inflated success rates (99–100%) obscured how services actually work.
5. Deletion of misconduct. Drawing on van Leeuwen's recontextualization, sites discursively delete misconduct and cast humanization as a rational, defensible response to surveillance and flawed detection rather than as facilitating cheating.
6. Front/back-stage leakage. Despite polished front-stage performances, sites lacked registered business addresses, displayed unexplained logos, and linked to blank or inactive social-media pages—evidence that the "mask slips" and the backstage is less sophisticated than presented.
7. A self-reinforcing feedback loop. Humanizers are a legible node in a cycle in which assessment demands a performative display of independent authorship, generating a market for AI tools, which generates institutional surveillance, which in turn generates a market for evasion.
8. Commercial actors are the architecture, not accelerants. Because each layer of the cycle (Large Language Models (LLMs) providers, detector vendors, humanizer platforms) is inhabited by commercial interests with no structural incentive to interrupt it, new tools alone will not break the loop.

What this means for practice

• Administrators. Stop treating humanizer sites as a policing problem and read them as a diagnostic signal that assessment demands a performance of independent authorship; better detectors or blocking do not disrupt the cycle, structural assessment reform does.
• Administrators. Account for the epistemic, academic, and financial harm that detector-based enforcement causes students, since the same market that supplies evasion is generated by surveillance.
• Assessment professionals. Redesign toward process-visible, interactive tasks whose value cannot be faked by reformatted text, rather than adding another detection layer.
• Instructors. Read humanizer marketing as a mirror of mainstream AI marketing: deleting misconduct, borrowing institutional logos, and mystifying with proprietary algorithms and success rates of 99–100% appear in softened form in everyday AI promotion.
• Researchers. Track the commercial architecture — LLM providers, detector vendors, and humanizer platforms each profit from the loop — because no layer has a structural incentive to interrupt it.

Limitations

• Only the front stage is observable: the backstage operation of humanizers is concealed, so the authors cannot establish who operates the sites, from where, or how the tools work.
• The study did not test whether humanizer tools actually evade AI detection technologies; the authors state this efficacy question was outside the scope of the work.
• The 55 humanizer sites were found through Google searches, which the authors chose for ecological validity, and no data were collected on how students actually reach these sites, leaving social media, word of mouth, and messaging channels unmeasured.
• Sites were constantly changing, updating, or disappearing — several became inaccessible during data collection — which the authors name as a barrier to replicability, and the deep analysis covered 3 of the 55 sites, so the findings are not easily generalizable.

Citation

Roe, J., Perkins, M., Bannister, P., Furze, L., & Wood, J. (2026). Dramaturgies of deception: AI humanizers and the performance of legitimacy in higher education assessment. Preprint, arXiv.