AI Detectors Are Wrong Often Enough to Hurt Kids

Ohio's Department of Education and Workforce, in materials supporting its 2026 model AI policy, flagged a 61.3% false-positive rate for AI-detection software specifically on writing by non-native English speakers. A separate documented case, William A., shows a student earning a 3.4 GPA denied a free appropriate public education because AI-detection tooling masked, rather than helped identify, a real learning disability. Detection software is not a neutral solution to the AI-in-schools problem - it introduces its own serious equity risk.

AI Detectors Flag Non-Native English Speakers' Writing as "AI" 61.3% of the Time. That's Not a Rounding Error.

The most common institutional response to "students might use AI on assignments" has been to buy detection software. Ohio's own state education department, in the research behind its mandatory July 2026 district AI policy, put a number on why that response is flawed on its own terms.

61.3% - the false-positive rate for AI-detection tools flagging writing by non-native English speakers as AI-generated, when it wasn't.

That's not a tool that's slightly imprecise. That's a tool that's wrong on this population more often than it's right in the other direction, which functionally means multilingual learners face disproportionate discipline, grade penalties, or academic-integrity accusations for writing in their own voice.

The second case Ohio's own materials flag

A separate documented case, referred to as William A., involves a student who earned a 3.4 GPA while a real learning disability went unaddressed - AI-detection and related tooling contributed to masking the disability rather than surfacing it, and the student was denied a free appropriate public education as a result. Ohio's model policy analysis notes this case explicitly as a caution against over-reliance on detection tooling as a stand-in for actual diagnostic understanding of a student.

Also flagged in the same analysis: the state's model AI policy is "largely silent on special education," a gap that leaves districts exposed exactly where detection-tool false positives do the most damage.

Why detection was always the wrong layer to solve this at

Detection tools try to answer a binary question - did a human or an AI produce this text - using pattern-matching on writing style. Multilingual learners, students with certain learning disabilities, and even strong writers with unusual voice can all produce text that pattern-matches to "AI-like" for reasons that have nothing to do with whether they used AI. The tool is guessing at authorship from the wrong signal.

The question a teacher actually needs answered isn't "did a machine write this." It's "does this student understand the underlying skill." Those are different questions, and only one of them is answerable by a detector running after the fact on a finished piece of writing.

What actually answers the right question

Diagnosing understanding requires evidence gathered during instruction, not forensic analysis of a submitted essay - a worked problem, an in-class response, a pattern of answers across an assessment that reveals a specific misconception rather than a stylistic fingerprint. That's a fundamentally different kind of system: not a gatekeeper deciding whether a piece of work is legitimate, but a diagnostic layer working with the teacher to understand what a student actually knows, using assessment evidence the school already collects, before the detection question even needs to be asked.

For a multilingual learner specifically, that shift matters enormously. A system built around identifying misconceptions treats their errors as information about what to reteach. A detection tool built around identifying "AI-like" text treats their voice as a red flag. One of those is educationally useful. The other has a documented 61.3% error rate against exactly this population.

See Teacher-Facing AI in Action.

Related reading

• NPR teacher trust poll
• Ohio AI policy deadline
• teacher-facing AI

FAQ

What is the false-positive rate for AI detectors on non-native English speakers?

61.3%, according to research cited in Ohio's Department of Education and Workforce materials supporting its 2026 model AI policy.

What is the William A. case?

A documented case where a student with a 3.4 GPA was denied a free appropriate public education because AI-detection and related tooling contributed to masking a real learning disability rather than identifying it.

Does Ohio's model AI policy address special education specifically?

No - legal analysis of the policy notes it is "largely silent on special education," a gap that leaves districts exposed given documented detection-tool equity problems.

What should schools use instead of detection software to address AI concerns?

A system that diagnoses understanding from assessment evidence gathered during instruction - identifying specific misconceptions - rather than forensically analyzing finished writing for stylistic AI markers.

Sources

• KJK, "Ohio's July 1, 2026 School AI Policy Deadline," June 12, 2026
• Ohio Department of Education and Workforce, AI Model Policy, January 2026