AI Writing Detection as Classification Infrastructure
This essay argues that AI writing detectors are not accuracy problems waiting to be solved. They are classification infrastructure being installed across higher education, and infrastructure theory predicts that their deepest failures will compound, not self-correct, as the system becomes more embedded. It begins with a case. The case is not the argument. The case is what the argument explains.
In February, a judge found that Adelphi University’s decision to punish a student for AI plagiarism was “without valid basis and devoid of reason.”3 Orion Newby, a freshman with documented learning differences who had worked with university-provided tutors throughout his writing process, submitted an essay on Christianity and Islam. His professor ran it through Turnitin’s AI detector. The tool flagged it as wholly AI-generated. Newby denied using AI.
The school gave him a zero and ordered him into an anti-plagiarism course, a remediation program for a category he’d been wrongly sorted into, then warned that a second offense could mean suspension. When he appealed, the school declined to reexamine the allegations. It never interviewed his tutors. It never examined his drafts. It had a score.
Newby’s family spent six figures on legal fees to prove what a conversation could have established for free.
His case is not unusual. Only the outcome is.
A French-born MBA student is suing after GPTZero flagged his exam answers. The university suspended him for a year. His lawsuit alleges discrimination against non-native English speakers.
A student with OCD and generalized anxiety disorder is suing after professors interpreted her formal, structured writing style as evidence of AI generation.
Students describe running their own work through multiple detectors before submitting, trying to pre-clear assignments they wrote themselves.
One recent graduate told NBC News that she and her classmates had a group chat called “Academic Felons for Life.”
The conversation about these tools has mostly been a technology debate: Do they work? Are they accurate enough? But these questions miss the structural problem entirely.
AI writing detectors are not technology products being evaluated by educators. They are classification infrastructure being installed across education in real time.56
Infrastructure, in the precise sense that scholars of technology use the term, is not just the thing underneath the thing. It is a system that gets woven into institutional processes so thoroughly that it becomes invisible: taken for granted, difficult to question, and nearly impossible to remove.6 Infrastructure persists not because it is optimal but because the cost of replacing it exceeds the cost of living with its flaws.
Everything gets built on top of it, and eventually, removing the foundation means demolishing the building.
Detection scores feed into grading workflows, academic integrity databases, and student disciplinary records. Faculty build syllabi around the assumption that submitted work will be scanned. Policies reference detection thresholds.
When Purdue announced that Turnitin’s AI detection update would activate automatically, they disclosed that the university “does not have the ability to turn it off.”7
Purdue University, April 2023
The institution that licensed the tool could not control when the tool changed underneath them. That is not a product. That is infrastructure.
Tap any node. Then tap “Remove the tool.”
And infrastructure theory predicts that the tools’ three deepest problems will compound, not self-correct, as the system becomes more embedded.5
AI detectors produce continuous probability scores: a percentage likelihood that text was generated by a machine.12 But institutions need categories: cheating or not cheating, flagged or clean. So they impose thresholds. Where you draw the line determines who gets accused.
Eight students. Two actually used AI. Move the threshold. Tap any student to see why they scored where they did.
Turnitin hides scores in this range. They show an asterisk instead.
Turnitin acknowledged the boundary problem seven weeks after launch. Their Chief Product Officer admitted they had discovered “a higher incidence of false positives” when less than 20% of AI writing is detected.2 She did not disclose the actual rate. Their fix was to display an asterisk instead of a number for scores between 1% and 19%.
They did not improve the tool in that range. They stopped showing you the score. The infrastructure kept running; they dimmed the dashboard light.
The chance of flagging a fully human-written paper as AI-generated. This is what Turnitin advertises.
The number the company advertises and the number the user experiences are measuring different things.
Students who know they will be scanned change their writing.8 Not to write better, but to write in ways that avoid triggering detection. They write less clearly, less concisely, less precisely, because clarity and predictability are exactly what the detectors flag.
Turnitin’s own documentation explains that it evaluates how statistically likely each word choice is given its context: language models choose optimal words, so human writers need to choose suboptimal ones to avoid suspicion.13 Better writing now looks more “AI-like” to the machine.
The system produces the adversarial behavior it was designed to prevent, then points to that behavior as justification for its own existence.
Faculty are adapting too. In-class handwritten exams replace take-home essays, not because they are pedagogically superior but because they are scan-proof. The classification system is restructuring pedagogy around itself.
The humanizer industry alone drew 33.9 million website visits in a single month.11 Students are paying for the privilege of making their own writing worse so that a machine will believe they wrote it.
The system is producing exactly the adversarial behavior it was designed to prevent, then pointing to that behavior as justification for its own existence.
Researchers at Stanford tested seven AI detectors and found that they classified over 61% of TOEFL essays written by non-native English speakers as AI-generated, while performing near-perfectly with essays by U.S.-born students.1
The reason is structural: non-native speakers tend toward simpler syntax, more predictable vocabulary, and more formulaic constructions. These are the exact features the detectors associate with machine output.
The detectors are not failing randomly. They are failing along the same lines that educational systems have always failed along.
Students with autism, ADHD, and other conditions that produce systematic or highly structured writing patterns face elevated false positive rates.4 The detectors are not failing randomly. They are failing along the same lines that educational systems have always failed along. And calling the result a technical measurement.
But beneath all three problems lies a deeper one: the category itself is incoherent. The detectors assume a binary: human-written or AI-generated. That binary is already obsolete.
Asked an LLM to brainstorm topic ideas, then researched and wrote the essay herself.
Drafted from scratch, asked an LLM to identify weak arguments, revised by hand.
Dictated rough ideas into a voice memo, fed the transcript to an LLM for an outline, then wrote from the outline.
Wrote a full draft, asked the LLM to suggest transitions, rejected most, adopted two.
Asked the LLM to generate a first draft, then substantially rewrote every paragraph in her own voice.
Pasted in a prompt and submitted the output unchanged.
The detector cannot distinguish between the student who used AI to think and the student who used AI to avoid thinking. The classification system was designed for a world where “human” and “machine” writing are discrete categories, and that world no longer exists.
What we are building is a system that punishes students for how their writing sounds rather than how they learned.
The psychiatrist Arthur Kleinman called this the “category fallacy”: importing a classification scheme into a context where it does not fit.9 The human/AI binary made a kind of sense before AI became a writing environment rather than a writing tool. It does not make sense now. But the infrastructure has already been built on top of it.
And the company that built the infrastructure knows it.10
This is the signature of mature infrastructure. The system’s operators can describe its limitations in detail, publish documentation acknowledging those limitations, and build an entire support apparatus for managing the consequences of those limitations, all while continuing to sell and expand the system. The disclaimer does not constrain the infrastructure. The disclaimer is part of it.
Infrastructure theory predicts what happens next. The tools will become invisible in the technical sense: not hidden, but taken for granted. Unquestioned. That is what infrastructure scholars mean by “transparent”: not that you can see through it, but that you stop seeing it at all.
Switching costs compound
Every policy, every workflow, every database that references a detection score makes the system harder to remove.
Disparate impacts accumulate
False flags travel in student records, shaping institutional trajectories long after the detection score itself is forgotten.
The binary collapses
Writing is already becoming collaborative in ways no detector can parse. But the infrastructure persists because everything has been built on top of it.
We have seen this before. The diagnostic categories in the DSM have persisted for decades despite widespread acknowledgment of their scientific limitations, not because they are accurate but because insurance billing, treatment guidelines, and clinical training are all built on top of them. AI detection is on the same trajectory, compressed from decades into years because digital infrastructure installs faster than institutional infrastructure.
What would it mean to treat AI detection as an infrastructure design problem rather than a technology procurement decision? At minimum, three things.
Tap any proposal to see what happens when it is not adopted.
If a detection tool cannot report its false positive rate by race, by language background, and by neurodiversity status, it should not be deployed in any context where those false positives carry consequences.
Every AI detection contract should require re-evaluation on a fixed schedule (two years, not perpetuity) so institutions are forced to actively decide whether the infrastructure still serves them rather than allowing it to persist by default.
An absolute prohibition on using detection scores as the sole basis for academic integrity proceedings. Turnitin’s own documentation already says this, in fine print, on every report, beneath the score that institutions are already treating as a verdict.
Document your writing process.
Drafts, outlines, revision histories, timestamped notes. If you are ever flagged, the process is the evidence. The score is not.
Know that a detection score is not a finding.
Turnitin{'’'}s own documentation says it should not be used as the sole basis for action against you. Quote it.
Do not pre-scan your own work.
Running your essay through multiple detectors before submitting teaches you to write for machines, not for readers. That is the looping effect, and you are inside it.
Read the disclaimer at the bottom of the report you are about to act on.
If the tool{'’'}s own maker says the score should not be the sole basis for action, ask yourself what else you have.
Ask what you are assigning for.
If the answer is learning, and the tool cannot measure learning, the tool is measuring the wrong thing.
Track your own false positive patterns.
Which students get flagged? By what background? If you notice a pattern, the pattern is the data. Report it upward.
Ask your vendor for demographic false positive data.
If they cannot provide it, you are deploying a tool whose failure patterns you cannot see. That is a liability question, not a technology question.
Put a sunset clause in every detection contract.
Two years, not perpetuity. Force your institution to actively decide whether the infrastructure still serves you, rather than allowing it to persist because removing it is harder than keeping it.
Audit the secondary infrastructure.
How many hours of professional development, how many integrity hearings, how many student appeals are being generated by the tool? Count the cost of managing the tool{'’'}s failures against the cost of the tool itself.
Orion Newby got his grade back and his record expunged. It cost his family six figures and a federal court order.3
The infrastructure will sort the next student tomorrow morning, before anyone reads the disclaimer at the bottom of the screen.
Share this essay. Every section has a direct link.