The Adjunct’s Dilemma: Handling Suspected AI Use With No Institutional Backing

You teach two sections on a semester contract, you’ve got no office, and a paper just landed in your inbox that doesn’t sound like the student who wrote it. You run it through a detector out of curiosity, it says “98% AI,” and your stomach drops. Now what? Because the honest answer, for a lot of adjuncts, is that there’s no policy, no chair returning your email, and no one whose job it is to have your back.

This is the part of the AI conversation that rarely gets written about. Tenured faculty debate detection in the faculty senate. Contingent faculty deal with it alone, at 11 p.m., wondering whether raising it could cost them next semester’s classes. Here’s how to handle a suspected case fairly when the institution gives you nothing to stand on.

Key takeaways

• A detector score is a probability, not proof. Acting on a number alone puts your contract at more risk than the student’s grade.
• Document the assignment and the student’s known work, not just the suspicious paper. A gap in writing is a stronger signal than any percentage.
• Talk to the student before anyone says the word “cheating.”
• Know what authority your contract actually gives you. Grading discretion and the power to file integrity charges are not the same thing.
• Redesign so AI shortcuts don’t pay off. That’s the one lever fully under your control.

Why your situation is different

Academic integrity systems assume a faculty member with standing: a permanent appointment, a department that backs its people, and time to follow a formal process. Adjuncts often have none of that. You may not be able to file a misconduct charge without a tenured sponsor. You may have no access to the integrity office. And the unspoken risk hangs over everything: an accusation goes sideways, a student complains, a parent emails the dean, and a contract simply doesn’t get renewed, no reason given.

So the question isn’t only “is this fair to the student.” It’s also “what can I actually do here that won’t blow up on me.” Those two questions point the same way more often than you’d think, because the careful path protects both of you.

First, understand what a detector score is not

A detector did not “catch” the student. It produced a statistical guess. These tools estimate how predictable the text is, roughly how closely the word choices track what a language model would tend to produce. If you want the mechanics, here’s how detector scores are actually calculated. The short version: it’s pattern-matching on probability, and it gets things wrong.

Three facts worth keeping in your back pocket, all verifiable:

• OpenAI retired its own AI-text classifier in July 2023, saying it wasn’t accurate enough.
• A 2023 study from Stanford (Liang et al.) found GPT detectors disproportionately flag writing by non-native English speakers as AI-generated.
• Detectors output probabilities, and two of them will routinely disagree on the same paragraph.

If you build a case on a score, you’re building it on sand. Worse, you’re exposing yourself. When a student pushes back and asks you to explain what “98%” means, “the software said so” is not a defense that survives a department chair who’d rather avoid the headache.

Build your read on the work, not the tool

Set the detector aside and ask older, sturdier questions. These are the things that actually hold up.

Compare to known work. Do you have anything the student wrote in class, by hand, or in a low-stakes assignment? A sudden jump from rough-but-real to flawless-but-hollow is the single most telling signal. It’s also something you can describe in plain language.

Look for specificity gaps. AI-generated work tends to be confident and oddly empty. It gestures at the course themes without engaging the actual readings, names no specifics, and answers a slightly different question than the one you asked. That’s a content judgment you’re qualified to make.

Check the prompt fit. Did the paper respond to your particular assignment, or to a generic version of it? Models often miss the constraints you added out loud in class that never made it onto the handout.

None of this requires software. All of it is defensible, because it’s your professional reading of a student’s work, which is exactly what you were hired to do.

A mini-scenario

Maria adjuncts a first-year composition course. A student named Devon turns in an essay that’s noticeably smoother than his earlier drafts. A detector flags it at 91%. Maria’s first instinct is to mark it as misconduct. Instead, she pauses.

She pulls Devon’s in-class writing sample from week two. The voice is different, sure, but Devon is also an ESL student, and she remembers the Stanford finding about false flags. So the detector score is worth even less to her now, not more. She schedules a quick chat and asks Devon to walk her through how he got from his outline to the final draft, and to expand on one claim in person.

Two things could happen. Devon explains his process, shows his notes, clarifies the claim, and that’s that, no accusation made, no harm done. Or he can’t say much about his own paper and goes quiet. Even then, Maria doesn’t shout “cheater.” She gives feedback grounded in the work’s actual weaknesses, notes the gap in writing, and, if her contract allows, asks her chair how a suspected case should be handled. Either way, she never staked anything on a percentage she couldn’t defend.

Have the conversation before anyone says “cheating”

A direct, low-key talk does more work than any tool. Ask the student to explain a specific argument, describe their research process, or talk through one source. Frame it as wanting to understand their thinking, not as a trial.

This serves three purposes. It’s fair, since plenty of “suspicious” papers come from students who genuinely improved or got legitimate help. It’s informative, because a student who wrote their paper can almost always talk about it. And it’s protective: a documented, respectful conversation is the thing that makes you look reasonable if the matter ever escalates.

Keep a short record either way. A few lines in your own notes about what you observed and what was said. Not to build a prosecution, but so your account stays consistent if you’re ever asked.

Know your actual authority

Before you do anything formal, find the line you’re not allowed to cross. Read your contract and the faculty handbook, and look specifically for two different powers that often get confused:

1. Grading discretion. You almost certainly have this. You can evaluate the quality, depth, and originality of work and grade accordingly.
2. Filing integrity charges. You may not have this, or it may require a chair’s sign-off. File without standing and you can end up being the one who violated process.

When the policy is silent, the safest real move is to grade the work on its merits. A thin, evasive, source-free paper earns a thin paper’s grade on the rubric you already published, whether or not a machine wrote it. You’re on solid ground assessing the writing in front of you. You’re on shaky ground policing how it was produced with no policy behind you.

The lever you fully control: assignment design

You can’t fix your institution’s missing policy. You can make AI shortcuts mostly pointless.

• Grade the process. Require an outline, a draft, and a reflection. Faking the whole pipeline is more work than just doing the assignment.
• Anchor to class. Reference a specific in-class discussion, a guest comment, a local example. Models can’t fake what only happened in your room.
• Add an oral or in-class component. A two-minute “tell me about your argument” check, or a short handwritten in-class piece, gives you a real baseline.
• Ask for personal stakes. Prompts that require the student’s own experience or position are harder to outsource and more interesting to read.

This takes more setup, but it’s far less stressful to run. Detection software you can’t trust gets replaced by structure you designed on purpose.

Knowing the tools without depending on them

You don’t have to pretend detectors don’t exist. It helps to understand how they behave, because then you can speak to their limits instead of treating a score as gospel, and you can recognize a likely false flag when you see one. For broader background, there’s more reading on AI detection and writing. The point isn’t to wave a percentage at a student. It’s to understand the tech well enough that a percentage never does your thinking for you.

The takeaway

The adjunct’s dilemma isn’t really about catching cheaters. It’s about acting fairly and protecting yourself at the same time, with none of the institutional cover other faculty take for granted. Skip the urge to escalate on a score. Read the work, talk to the student, know your contract, and redesign so shortcuts don’t pay. That’s a position you can defend to a student, a chair, or yourself.

If you want to understand the detection side well enough to talk about it honestly, check how a passage reads to a detector and see what the scores actually do, and don’t do, before you ever rely on one.