When COPPA Reaches a K-12 AI Detector, It Binds the Vendor

COPPA binds an operator whose service is directed to children, or who has actual knowledge that it is collecting personal information online from a child under 13 ⁸. A detector a teacher pastes an anonymized excerpt into may never collect from a child at all. One that students sign into does. The rule's own list of personal information never names an essay ¹.

Paragraph (11) of the definition in 16 CFR 312.2 is where an essay lands. Information about a child that the operator collects online from the child, and combines with an identifier, is personal information under the rule ¹. A name on the cover page is such an identifier. So is the email address behind the account a submission travels through, or a user name that works as contact information. So is a persistent device or cookie identifier picked up on the way. Text becomes covered by association rather than by content.

Two questions settle most of it. Does anyone under 13 use the tool directly? Does a submission carry a name, a student number or an account identifier? A building where teachers paste anonymized excerpts sits in a different position from one that wires the detector into the gradebook. A contract that names students under 13 as users is hard to read as anything but actual knowledge.

A school can consent in a parent's place. The terms come from FTC guidance rather than from the rule itself. The agency says a school may act as the parent's agent where it contracts with a provider for the benefit of its students ². That power stops at the educational context: the information may serve the school and no other commercial purpose ².

The April 2025 amendments were expected to write that path into the rule text. They didn't. The Commission had proposed definitions of School and School-authorized education purpose, plus a codified school authorization exception. It declined to finalize them, to avoid conflicting with planned Department of Education changes to the FERPA regulations, and said it would keep enforcing COPPA in the ed tech context under its existing guidance ³. The proposal is on the record, the exception is not in the rule, and in a dispute that gap is the ground district counsel has to stand on.

The same guidance puts compliance on the vendor. The FTC says the operator is responsible for complying with the Rule, and that a provider should not state in its terms of service that the school is responsible ². Terms that push that duty onto a district belong in the detector RFP, flagged before anyone signs.

Section 312.10 of the amended rule sets four duties on the operator. Keep a child's personal information only as long as reasonably necessary for the purpose it was collected for. Delete it with reasonable measures once that purpose ends. Never retain it indefinitely. Maintain a written retention policy naming the purposes, the business need and a timeframe for deletion ⁴.

That written policy has to appear in the notice on the service, which makes it readable before a contract is signed ⁴. It is also the cheapest check a K-12 district has. A vendor selling into schools with no published children's retention policy is either outside COPPA's scope or out of compliance with it, and asking which produces a useful answer either way. The amended rule took effect June 23, 2025, and regulated entities have until April 22, 2026 to comply with nearly all of it ³. Nothing in that schedule waits for a district's procurement calendar.

The FTC put the same limit in plainer words in 2022. Its education technology policy statement says it is unreasonable for an ed tech provider to retain children's data for speculative future potential uses ⁵. A pile of student text kept with no stated purpose or deletion date is the retention that line describes, and it is where a training corpus usually starts.

Not on school consent alone. The FTC's 2022 policy statement says an ed tech operator collecting under school authorization may use that information only to provide the requested online education service, and prohibits use for marketing, advertising or any other commercial purpose unrelated to that service ⁵.

The amended rule adds a second lock, on the disclosure side. In adopting the separate-consent requirement, the Commission named three kinds of disclosure that are not integral to a service: transfers for money or other consideration, transfers for advertising, and transfers to train or otherwise develop artificial intelligence technologies ³. Each of those needs consent of its own. A detector whose model provider may keep or train on the text it receives is making that transfer. One whose provider returns a score and retains nothing is the service-provider disclosure the preamble treats as integral ³. The contract has to say which, and a privacy page that does not say is the question to ask.

Vendors tend to answer the easier half of this. The rule text above governs disclosure to a third party. An operator's reuse of text it already holds is governed by the use and retention limits instead ³⁵. Whether a vendor training its own model on submissions counts as a purpose unrelated to the requested service is the question the guidance leaves open, which is why a district that wants no training writes the ban into the contract rather than reading it out of the policy statement ⁵. A contract that names only transfers leaves the more likely practice unaddressed.

The FTC's own guidance lists the questions a school should put to an operator before authorizing collection ². What personal information does it collect from students? How does it use that information? Does it use or share the information for commercial purposes unrelated to the requested service? Will it let the school review that information and have it deleted? What measures protect security and confidentiality? What are its data retention and deletion policies for children's personal information? On the agency's reading, a wrong answer to either of the middle two removes the school's power to consent in the parent's place ². That list is what the rule and the guidance say rather than legal advice, and district counsel reads it against the terms on the table.

COPPA stops at 13, and the district's exposure does not. The rule defines a child as an individual under the age of 13, so a ninth grader's essay sits outside it entirely ¹. The protections a detector contract needs for a 15-year-old come from FERPA and from state student-privacy statutes instead.

A policy scoped to students under 13 covers the lower grades and none of a high school. California is the sharpest version of that gap, in section 22584 of its Business and Professions Code. That law covers personally identifiable information or materials a pupil creates or provides to an operator for school purposes, and binds a service with actual knowledge that it is used primarily for K-12 school purposes and was designed and marketed for them ⁶. A detector designed and marketed for K-12 and used mostly by districts meets that test; one sold across schools and colleges may not, and the vendor's own marketing is the evidence either way. The same law requires the operator to delete covered information under the district's control when the district asks ⁶. No age appears in either provision. One set of terms for every grade means drafting to the stricter standard.

Whether a detector report on a named student becomes an education record the district has to produce is a FERPA question, taken up in detection and FERPA. Age does not change that question either. The parent asking for the file will ask the district.

Privacy compliance says nothing about whether a detector reads K-12 writing correctly. The published evidence divides by who did the writing. Seven detectors tested by Stanford researchers in 2023 were near-perfect on essays by US eighth graders, and misclassified more than half of the TOEFL essays written by non-native English speakers, an average false positive rate of 61.22 percent ⁷.

Rewriting those same human essays with richer vocabulary cut the average false positive rate to 11.77 percent, which means the flag was tracking range of expression rather than authorship ⁷. The risk the study measured was for writers with a narrow English vocabulary, and whether a building's writers resemble that set more than the eighth graders is an assumption a district checks against its own population, not something the study settles. The exposure in a multilingual district is worked through in detectors and non-native English writers. None of that is fixed by a privacy clause.

A K-12 buyer weighing Human meets the same limit from the other side. Human's detector is tuned for college-level academic writing. Any reading it returns on a seventh grader's paragraph points at the drafts and no further. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating.