The Privacy Question: Are You Allowed to Run Student Work Through Third-Party AI Detectors?

Usually yes, when the tool is one your school has licensed and signed a data agreement with — and much murkier when you personally paste a student’s essay into a random free web detector, because what separates those two situations isn’t the technology, it’s the paperwork.

Key takeaways

• Is it legal to run student work through an AI detector? Generally yes when it’s your institution’s licensed tool — that’s what FERPA’s school-official exception and a signed data agreement are built for.
• The risky version is going freelance: pasting an identified student essay into a free web checker under your personal account, with no contract behind it.
• Student essays are often part of a student’s education record, and they’re frequently personal writing.
• Free web tools may retain, log, or train on what you paste; the only way to know is their current terms.
• Detectors output probabilities, not proof, so the privacy cost buys you a hint, not a verdict.

Why this is a real question, not paranoia

A student essay is not just text to be graded. Once submitted and kept, it’s typically part of the student’s education record — the same legal family as a transcript. It’s also often deeply personal writing: reflection papers about family and mental health, creative work the student may want to publish someday.

Paste that into a web tool and you’ve handed a copy to a company. If the company is under contract with your school, there are rules about what it can do with the copy. If it isn’t, the rules are whatever its terms of service say — and free tools fund themselves somehow, which is why their terms can quietly permit retention, logging, or training on submissions.

That’s the issue in miniature. Nobody loses sleep when a school runs essays through its licensed plagiarism checker; a data agreement governs every byte. The worry is an individual teacher, alone at 11 p.m. with a suspicious paragraph, pasting a named student’s essay into the first free detector a search turns up.

Is it legal to run student work through an AI detector?

First, plainly: this is general information, not legal advice — check with your institution’s counsel or privacy office before you build a workflow on any of it.

In the US, the frame is FERPA. Schools that accept federal funding generally can’t disclose personally identifiable information from education records to third parties without consent. The exception doing the work here is the school-official exception: a vendor can be treated as a school official when the institution contracts with it, keeps direct control over the data, and limits its use to the contracted purpose. That’s precisely what a school’s license agreement with a detection vendor exists to do, and it’s why the institution-licensed tool is the clean path — the vendor operates, legally, as an arm of the school, bound by the agreement.

Now flip it. Paste an identified essay into a free checker with no agreement behind it and none of that structure exists. Whether a particular paste crosses a legal line depends on the facts — how identifiable the work is, its record status, what your state adds — but the general shape makes privacy officers wince, and “I didn’t know” has never been a policy.

K-12 adds layers: COPPA governs what online services can collect from children under 13, and most US states have their own student-data-privacy laws that reach further. If you teach younger students, there are extra considerations when the writers are minors worth reading before you touch any outside tool.

Outside the US, the rules get stricter, not looser. Under GDPR, processing a student’s personal data needs a lawful basis and a defined purpose, and European institutions typically require a data-processing agreement before any third party sees it. “Paste first, wonder later” does not survive contact with a data-protection office.

A licensed tool and a free web checker are not the same act

These aren’t the same thing at different price points. They’re different legal acts.

With a licensed tool, there’s a contract. Good ones spell out how long submissions are retained, forbid the vendor from training models on student work, name the subprocessors, restrict who can see results, and give the school a deletion path and audit rights. When something goes wrong, the school has someone to hold accountable.

With a free web checker on your personal account, there’s a clickwrap agreement you accepted in half a second. Its terms may allow the service to keep what you paste, tie it to your account and IP, or use it to improve the product. There’s usually no deletion promise, no subprocessor list, and no one at your institution who even knows the disclosure happened. Even paid, reputable tools store things as a matter of course — we’ve looked at whether Originality.ai stores the content you scan, and for nearly every cloud detector the honest answer is “yes, in some form.” The question is never *whether* text leaves your machine. It’s who’s bound by what once it does.

What to check in a detector’s terms before pasting student work

If you’re evaluating any tool that isn’t already under an institutional agreement, four questions cover most of the ground:

• Retention. Is pasted text stored, and for how long? “As long as necessary” with no specifics is a red flag.
• Training use. Does the vendor use submissions to train or improve its models? Student work should never become someone’s training data by accident.
• Deletion path. Can you actually delete a submission, and does deletion purge it from backend systems or just hide it from your view?
• Subprocessors. Who else touches the data — hosting providers, analytics vendors, model providers? A tool that won’t say hasn’t been asked hard questions before.

If those answers aren’t findable in a few minutes, that’s an answer too.

Safer practices that don’t require a law degree

• Use the school’s licensed tool if one exists. The license is what converts a third-party disclosure into a sanctioned one.
• If there’s no licensed tool, ask before adopting your own. Your privacy officer would much rather answer a question now than clean up a disclosure later.
• Anonymize before you paste anywhere unvetted. Strip names, student IDs, course headers, and identifying details inside the prose itself. It reduces the exposure; it doesn’t make the tool vetted.
• Never paste an entire identified essay into an unvetted free tool. If you must check something, a short anonymized excerpt is a smaller disclosure than a whole named document.

What students can ask — and challenge

Students aren’t bystanders here. They can reasonably ask which tool the school uses, what happens to their text after submission, how long it’s kept, and whether they can see the result — we’ve covered whether students can see their Turnitin AI score separately. Under FERPA, students (or parents, for minors) also have the right to inspect their education records and request amendment — real procedural footing for a flagged student.

They can also challenge the substance. Detectors report a probability that text resembles AI writing, not proof of what a student did. The 2023 Stanford research by Liang and colleagues found commercial detectors misclassifying a large share of non-native English writers’ essays as AI-generated, and OpenAI retired its own classifier in July 2023 over low accuracy. A tool that uncertain shouldn’t get a student’s personal writing without a data agreement — and its score should never be the whole case against anyone. There’s more on detection and student writing worth reading before you build a policy.

Frequently Asked Questions

Does FERPA prohibit teachers from using third-party AI detectors?

No. FERPA doesn’t ban any category of tool. What it regulates is the disclosure of education records to third parties without consent. When your institution licenses a detector under a contract that meets the school-official exception, running student work through it is generally the sanctioned path. The trouble starts when an individual instructor pastes identifiable student work into an outside service the school never vetted. That’s general information, not legal advice — your institution’s counsel or privacy office is the authority for your situation.

Is a school-licensed AI detector different from a free web checker?

Meaningfully, yes — and the difference is the contract, not the technology. A licensed tool operates under a data agreement that typically limits retention, restricts training on submissions, names subprocessors, and gives the school deletion and audit rights. A free web checker runs on whatever its public terms of service say, which may allow logging, retention, or training on pasted text — and you agreed to those terms alone, on the school’s behalf, without the authority to do so.

Do free AI detectors keep the text you paste?

Some do, and with many you simply can’t tell. Free tools have to fund themselves somehow, and their terms may permit retaining submissions, logging them alongside your account details, or using them to improve models. Before pasting anything sensitive, check four things in the current terms: how long text is retained, whether it’s used for training, whether there’s a real deletion path, and who the subprocessors are. If you can’t find those answers in a few minutes, treat that as your answer.

Should I anonymize student work before running it through a detector?

If you’re using anything other than your institution’s licensed tool, yes — strip names, student IDs, course headers, and any identifying details inside the prose itself before pasting. Anonymizing reduces the privacy exposure, but it doesn’t make an unvetted tool vetted, and an essay can identify its author through content alone. Treat anonymization as a harm reducer, not a permission slip, and still check the tool’s terms first.

Can students refuse to have their work run through an AI detector?

Usually not unilaterally — if the institution’s policy says submitted work may be checked with licensed tools, that’s typically a condition of the course. But students can do a lot short of refusal: ask which tool is used, ask what happens to their text and how long it’s kept, request access to the score under the school’s records policies, and challenge a flag through the appeals process. A school that can’t answer those questions has a policy problem worth raising.

Closing

The privacy question isn’t a reason to abandon detection; it’s a reason to do it through the front door — the licensed tool, the data agreement, the process students can see. If you want a feel for what these tools do before deciding what’s safe to feed them, run a piece of your own writing through a checker and see how much weight a probability carries.