In early 2023, as generative AI tools first entered academic discourse, conversations naturally turned toward containment and verification: Can software detect machine-generated writing? Nearly four years later, the empirical and pedagogical verdict is definitive.
AI text detectors do not work as advertised, their outputs cannot serve as evidence of academic misconduct, and their punitive use runs directly counter to a community founded on mutual trust and honor.
The Flawed Mechanics: Probability Is Not Authorship
AI detection platforms (including Pangram, Turnitin, GPTZero, and Copyleaks) operate primarily on statistical likelihoods—specifically metrics known as perplexity (how predictable a word is given the preceding context) and burstiness (the variation in sentence structure and length).
These tools do not detect digital footprints, watermarks, or origin metadata. They measure structural predictability. If human writing is clear, direct, structured, or formulaic, detectors frequently misread that clarity as machine generation.
Recent peer-reviewed research and cross-platform evaluations reveal significant systemic flaws:
Catastrophic False Positive Rates on Polished Prose: Research from the University of Maryland (Saha & Feizi) highlights that existing detectors exhibit high false-positive rates, often misclassifying human work that has undergone minor polishing or editing as machine-generated text. As human writers revise for clarity and brevity, their perplexity scores naturally drop—triggering detection algorithms simply for being well-edited.
Documented Linguistic Bias: Landmark findings led by Liang et al. (Stanford University) demonstrated that over 61% of human-written TOEFL essays by non-native English speakers were falsely flagged as AI-generated by leading detection tools. Subsequent multi-institution benchmarks consistently reveal that writers whose native language is not English face false-positive probabilities two to six times higher than native writers. Detectors actively penalize concise, non-idiomatic vocabulary.
Discipline Discrepancies and Inconsistency: Research evaluating detector consistency (e.g., Sultan Qaboos University; University of Chicago Booth School) found that detection accuracy varies wildly depending on subject matter, tool selection, and text length. In scientific and technical domains where precise, shared terminology is mandatory, detector reliability collapses, often producing false negative rates as high as 40% alongside elevated false positives.
A probabilistic score—whether it outputs 30% or 95%—is a measurement of statistical conventionality, not an audit of human thought.
Why AI Detectors Violate Our Honor System
The unreliability of detection software is a technical reality, but its impact on institutional culture is a community concern.
An Honor System functions on an explicit presumption of integrity: students are trusted as honorable scholars until substantive, credible evidence demonstrates otherwise. Subjecting student writing to black-box detection software subverts that relationship in several distinct ways.
Inverting the Presumption of Trust: Relying on an algorithm to screen assignments reframes teaching from mentorship into surveillance. It presumes dishonesty by default, outsourcing faculty discernment to an opaque proprietary engine.
The Burden of Impossible Defense: When an algorithm assigns a high “AI probability” to an authentic essay, the student faces a Kafkaesque challenge: proving they wrote something that originated inside their own mind. Requiring students to preserve screen recordings, keystroke trackers, or extensive time-stamped revision histories to clear their names introduces acute anxiety and erodes psychological safety.
Disproportionate Harm: Because false positives disproportionately strike multilingual learners, international students, and students adhering strictly to formal rubric structures, using detector outputs to justify disciplinary action introduces unvetted, algorithmic bias into judicial proceedings.
Submitting student work to an external detector to generate accusations of cheating violates the compact of mutual respect that defines an honor-based academic environment.
What to Do Instead: Process-Centered Pedagogy
Abandoning AI detectors does not mean ignoring the pedagogical questions raised by large language models. It means addressing them through intentional assignment design, relational teaching, and authentic assessment:
Focus on the Writing Process: Incorporate staged deliverables—annotated bibliographies, proposal memos, rough drafts, peer critique transcripts, and reflective self-assessments. When faculty engage with ideas as they evolve across several weeks, authorship is self-evident.
Anchor Assignments in the Classroom Context: Tie prompts directly to specific seminar discussions, local field observations, archival resources, or contemporary class debates that generic language models cannot replicate or synthesize without context.
Have Direct, Open Conversations: If an essay appears uncharacteristic of a student’s prior work or voice, use that observation as a catalyst for a conversation, not an accusation. Asking a student to explain their thesis development, walk through their research trail, or elaborate on a specific interpretive choice clarifies understanding far more effectively than an algorithmic confidence score ever could.
Our Stance
The Harte Center does not endorse or support the use of AI detection software. We strongly urge faculty across all disciplines to refrain from using automated detection scores as evidence or grounds for an honor violation.
Our community is built on mutual trust, rigorous inquiry, and personal honor. Upholding those values requires relying on human judgment, pedagogical relationship, and direct dialogue rather than flawed probabilistic software.
References
- Saha, S., & Feizi, S. (2025). Almost AI, almost human: The challenge of detecting AI-polished writing. arXiv. https://doi.org/10.48550/arXiv.2502.15666
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://doi.org/10.1016/j.patter.2023.100779
- Jabarian, B., & Imas, A. (2025). Artificial writing and automated detection (Working Paper). SSRN. https://doi.org/10.2139/ssrn.5407424
