What happened at Thoreau Elementary
Christine Apollos taught kindergarten at Thoreau Elementary School in Concord, Massachusetts for more than two decades before Superintendent Laurie Hunter fired her in August 2025. The district cited a run of unsatisfactory evaluations written by Principal Justin Sparks, who took over the school in 2022 and conducted Apollos's reviews from 2023 onward. Apollos says her ratings were acceptable until Sparks arrived, and that the decline reflects age bias rather than performance.
During arbitration, Apollos introduced a new claim: that Sparks did not write the evaluation himself. She ran his reports through GPTZero, an AI detection tool, and says it flagged several as likely AI generated, including the summative evaluation that recommended her dismissal. She also argues the document cites rubric sections that do not exist, which she frames as the kind of error an AI system would produce and a human evaluator would not.
The evidence problem with AI detection tools
The district and Sparks do not deny using AI entirely. They say the only tool involved was an AI grammar checker, a category most employers would consider unremarkable and would never think to disclose. That leaves the dispute resting entirely on what GPTZero's flag actually proves, and detection tools built to catch AI-written text have a well documented history of false positives on ordinary human writing, particularly from non-native speakers, technical writers, or formulaic documents like performance rubrics that follow a fixed template every evaluator is trained to reuse year after year.
Arbitrator Thomas Kochan ultimately gave the AI allegation no weight, while still calling the evaluation's factual errors troubling. That split matters for any employer facing a similar claim. An arbitrator can treat an AI-detection flag as too unreliable to decide a case, even while agreeing the underlying document has real problems. The detection tool did not settle anything. It just added a layer of dispute on top of the original performance question.
How the arbitrator complicated his own ruling
Kochan did not stop at rejecting the AI claim. According to Apollos's filing, he went further and asked two frontier AI models, GPT-5.4 and Claude Opus 5, to describe GPTZero's reliability, then used their responses to help discount the detection tool entirely. He did this after the hearing closed, without notifying either party, which meant neither side had a chance to question his prompts, challenge his sources, or weigh in on a method that effectively turned the arbitrator himself into an undisclosed expert witness on the very technology at the center of the dispute.
Apollos's lawsuit treats this as the stronger argument than the underlying AI claim itself. She alleges the studies the models cited were outdated relative to the GPTZero version she used, and came from competitors with an incentive to undermine it. Whatever the merits, the sequence is a useful cautionary tale: using AI to evaluate the reliability of AI evidence, outside the record and without disclosure, is its own procedural risk, not a shortcut around one.
Why this is bigger than one school district
Performance evaluations are exactly the kind of document enterprise managers increasingly draft with AI assistance, because rubric-based writing is tedious and formulaic, the same qualities that make it easy for a model to produce convincingly. A manager who runs a draft through a chatbot to tighten language is doing something close to what Sparks is accused of, and most companies have no policy that distinguishes acceptable AI assistance from a problem.
The Massachusetts case also exposes a harder fact: once a document reaches final form, almost nobody can prove how it was written. Detection tools analyze style patterns that correlate loosely with AI output, not authorship in any provable sense. Employers who assume a detection tool will settle a dispute about AI use in personnel decisions are relying on evidence courts and arbitrators are already treating as unreliable on its own.
The governance gap enterprise HR teams still have
Most companies that have adopted generative AI tools for managers have done so informally, through the same consumer-grade chatbots employees already use for everything else, with no logging of what was drafted, what was edited, or who approved the final language before it went into a personnel file. That absence of a record works fine right up until a termination gets challenged, at which point the lack of an audit trail becomes the employer's problem just as much as the employee's, and arguably more so, since the employer is the party a court expects to have kept the better records.
A defensible process does not require banning AI from performance writing. It requires knowing, and being able to show, which tool touched a document, what the manager changed, and that a human reviewed the final version against actual evidence. Few HR platforms currently capture that chain by default, which means most enterprises are one dispute away from discovering they cannot answer the question Apollos is asking Concord-Carlisle to answer now.
What CIOs and CHROs should do before the next dispute
The practical fix is unglamorous: a written AI-use policy for personnel documents, a record of which tools managers are authorized to use for evaluations, and a requirement that any AI-assisted draft be reviewed against specific, cited evidence before it becomes final and goes into a file an employee can later challenge. That review step, a human checking the draft against real observations rather than trusting fluent prose, is what actually protects an employer in a dispute, far more than any detection tool brought in after the fact to argue over what a model probably did.
CIOs buying HR and performance-management software should also start asking vendors directly whether their platforms log AI assistance at the document level, not just at the account level. The Concord case will likely resolve on procedural grounds specific to Massachusetts arbitration law, but the underlying exposure, a personnel decision nobody can fully document, travels to every enterprise that has let AI quietly into its management workflow without a policy to match.
.jpg)