The detection-first frame emerged in the months after ChatGPT's November 2022 release. Turnitin shipped AI detection in April 2023; GPTZero and Copyleaks released competing tools the same spring. Within a semester, many institutions had a detection tool in their LMS and a sentence in their syllabus pointing students at it.
The editorial position
Detection-first policies treat probability scores as evidence of misconduct. They produce false-positive cases that institutions cannot defend, fall hardest on multilingual and neurodivergent students, become unstable every time a paraphraser updates, and reframe the classroom as adversarial. An integrity-first frame keeps detection as one signal among many and puts assignment design, conversation, and process documentation at the center.
The case against detection-first is not a case against detection tools. It is a case against the policy architecture that makes those tools the first move, the primary evidence, and the default frame for the teacher-student relationship.
The detection-first frame and why it spread
The frame spread because it was tractable. Faculty senates and provosts' offices could not redesign every assignment, but they could enable a checkbox in Turnitin. The detector produced a number, and a number felt like evidence. New York City's initial ban of ChatGPT in January 2023, later reversed by Chancellor David Banks, captured the same impulse at policy level: prohibit the tool, detect the violators, restore the old order.
Three years on, the reasons that frame felt urgent are still real. Teachers face workloads that do not allow them to redesign every essay assignment. We are arguing that the specific solution detection-first offered does not solve the problem it claimed to solve.
A policy whose evidentiary instrument falls hardest on multilingual students is not a neutral policy that happens to have a bias. It is a policy that produces unequal outcomes by design.
Working Educators editorial
The procedural false-positive problem
The first crack in detection-first appeared as soon as the cases reached student conduct offices. A Washington Post investigation in August 2023 documented students who had written their own work and could not prove it once a detector had flagged them.
Vanderbilt's response was instructive. In August 2023, the university's Brightspace team disabled Turnitin's AI detector across the institution. Vanderbilt did not argue that no one cheats with AI. It argued that the procedural cost of the tool exceeded its evidentiary value.
The deeper procedural problem is that detection-first inverts the normal evidentiary structure of academic conduct. In a well-run process, the institution carries the burden of showing misconduct occurred. With a detector score driving the proceeding, the student is implicitly asked to prove the negative: to demonstrate that work they wrote is their own.
The disparate-impact problem
Weixin Liang and colleagues at Stanford published a study in the journal Patterns showing that seven leading AI detectors misclassified TOEFL essays written by non-native English speakers as AI-generated at rates above fifty percent, while the same detectors were nearly accurate on essays written by native English speakers.
The mechanism Liang's team identified is straightforward. Detectors learn that AI-generated text tends to be linguistically uniform and predictable. Non-native English writers also produce text that scores as uniform and predictable on those measures, for reasons that have nothing to do with how the text was produced.
A policy whose evidentiary instrument falls hardest on multilingual students is not a neutral policy that happens to have a bias. It is a policy that produces unequal outcomes by design, because the underlying signal it relies on is not the signal it claims to measure.
The fragile-pedagogy problem
Krishna and colleagues showed in their paper on paraphrase attacks that running AI-generated text through an off-the-shelf paraphraser dropped detector accuracy from above ninety percent to near random. Every paraphraser update, every new model release, every adversarial tool that appears changes the detection landscape.
A policy built on a tool with that update cadence is a policy that needs to be rewritten every few months. We have watched institutions issue revised guidance three times in a single academic year. Faculty stop reading the guidance. Students stop trusting the policy.
Pedagogy that depends on a vendor's model staying ahead of an open-source paraphraser is fragile pedagogy. It is asking the wrong system to do the load-bearing work.
The adversarial-classroom cost
The fourth cost is the hardest to quantify but the most consequential. Detection-first reframes the classroom. The teacher becomes an investigator, the student becomes a suspect, and the assignment becomes evidence in a proceeding that may or may not occur.
Students who know their work will be scored by a detector write differently. They write defensively, they avoid the AI-typical patterns the detector punishes, they shorten sentences and avoid the connective phrases that make essays readable. Faculty who treat detector scores as their primary signal grade differently, watching for the score rather than the argument.
The detector does not just produce occasional false positives. It produces a classroom culture in which the relationship between teacher and student is mediated by suspicion. That culture is incompatible with the kind of teaching most teachers want to do.
What integrity-first looks like
An integrity-first frame puts the detector at the back of the toolkit rather than the front. It treats a detector flag as an invitation to a conversation, not as evidence of a violation. It puts assignment design, process documentation, and direct conversation at the center.
Assignment design is the structural piece. Our colleagues writing on oral defense as an AI-resilient assessment have made this case at length: a five-minute conversation about a draft is worth more than a detection score.
Process documentation is the procedural piece. When a detector flag does occur, the response is the protocol described in our piece on handling a student who denies AI use after a flag, not an automatic referral. And the policy itself, built using the principles in our guide to AI policy faculty will follow, names what is permitted, what is not, and what verification looks like.
Frequently Asked Questions
Are you arguing that schools should stop using AI detectors entirely?▼
No. We are arguing that detection should not be the foundation of policy. A detector can be one signal among several, used to start a conversation rather than to conclude one.
What about faculty who do not have time to redesign assignments?▼
This is the real constraint. We do not think it is solved by detection-first; we think detection-first hides the problem rather than addressing it. Even small changes shift the load away from detection without requiring a full redesign.
Does the Liang study still hold up given the newer detectors?▼
The specific accuracy numbers vary across detector versions. The structural finding, that detection signals correlate with linguistic uniformity in ways that disadvantage non-native English writers, has not been refuted.
How does an institution defend itself against AI-assisted misconduct without leading with detection?▼
By making the misconduct harder to commit and easier to spot through assignment design, by documenting student process, and by training faculty to use detector scores as one input rather than as a verdict.
Is this position likely to change as AI tools improve?▼
Our editorial position rests on the structural problems with detection-first, not on the current accuracy of any specific tool. Better detectors would reduce the false-positive rate; they would not change the procedural inversion or the classroom-culture cost.
The Bottom Line
Detection-first integrity policy is a frame that solved an administrative problem in early 2023 and has been producing the cases it was supposed to prevent ever since. The procedural inversion, the disparate impact on multilingual students, the technical fragility, and the adversarial classroom dynamic are not bugs that better detectors will fix. They are features of the architecture.
The alternative is not a single policy. It is an orientation: integrity built through teaching, through assignment design, through process documentation, and through conversation, with the detector as one tool among several rather than the foundation. The work is harder. The result is more durable.
