When an AI Alert Becomes an Accusation of Cheating

Eleanor Canina opened her grades and found a zero.

The Green Hope High School freshman had submitted an English assignment about Romeo and Juliet. According to reporting by The Herald-Sun, republished by Government Technology, the teacher’s explanation was brief: “evidence of AI, Please redo.” Canina said she had written the assignment herself. [1]

The teacher reported running her work through three detection tools, which returned figures of 62, 75 and 87 percent. Her mother suggested reviewing earlier writing and the document’s revision history. The teacher, who was grading work for a class he did not directly teach, acknowledged the limitations of detection but maintained his position. WRAL later reported that another teacher reviewed the writing history and regraded the assignment. [1, 2]

The problem began before anyone determined whether Canina had cheated. A technological indication had already produced an academic consequence.

That sequence deserves attention because schools are managing several changes at once. They must prepare students to use AI effectively, continue teaching disciplines whose work AI can now perform, and decide how teachers should use the same tools to improve instruction. Academic integrity sits at the intersection of those obligations.

Students still need to do the thinking an assignment is designed to develop. Schools also need a sound basis for deciding that a student failed to do it. The controversies around AI detection show how easily those questions become confused.

What the Accusation Can Change

At Adelphi University, Orion Newby received a zero on an essay for a World Civilization course after Turnitin returned a 100 percent AI-generated score. An academic integrity finding and a required plagiarism workshop followed.

In January 2026, a New York court annulled the finding, ordered its removal from his academic record and rescinded the associated sanctions. The opinion identified problems with the evidentiary basis and the university’s procedures, including the denial of an advisor of his choosing, a failure to consider the two AI-detection results he submitted indicating human authorship, and an appeal handled by the official who issued the determination. It also noted that the professor’s subsequent email undermined the original allegation. [3]

The ruling did not scientifically establish that every AI detector is unreliable. It showed what can happen when an institution’s decision and review process do not support the consequence imposed.

A third dispute illustrates the significance of a grade even without an expulsion. In May 2026, The San Francisco Standard reported that a Palo Alto family filed a federal civil rights suit against the school district after a student was accused of AI use. The complaint alleged that a detector flagged 76 percent of an essay and that the resulting treatment reduced the student’s semester grade to a C. The family disputed the accusation; the district declined to comment because of the litigation. Those are reported allegations, not judicial findings. [4]

An integrity finding can affect a student’s reputation and academic record. A lower grade can affect access to advanced courses, scholarships or admissions. The significance depends on the circumstances, but calling the consequence “just a grade” does not establish that it is minor.

There is also an instructional cost. A student who begins writing to avoid suspicion may spend less effort developing an argument and more effort trying to predict what a detector dislikes.

What a Detector Actually Reports

“How reliable are AI detectors?” sounds like a question with a single answer. It is not.

Performance depends on the detector and its version, the length and kind of writing, the models it is tested against, and whether the text has been edited. A result established on one collection of essays does not automatically establish performance on another classroom’s work.

Even the number on the screen requires interpretation. Turnitin describes its percentage as the share of qualifying text that its model identifies as potentially AI-generated or modified with certain AI tools. An 80 percent result is not an 80 percent probability that the student cheated. Turnitin also warns that its results can misidentify text and should not be the sole basis for adverse action against a student. [5]

Running several detectors does not resolve that distinction. Different systems may respond to similar features of the same prose. Their agreement should not be assumed to provide independent confirmation.

Research also shows why the student population matters. A 2023 study by Weixin Liang and colleagues tested seven detectors on human-written essays and found substantial misclassification of writing by non-native English speakers. The authors connected the errors to constrained linguistic variation and showed that editing could change detection results. Those findings concern the tools and samples tested; they are not a current error rate for every commercial detector. [6]

The practical lesson is that familiar-looking prose can be an unreliable basis for attributing authorship.

That creates a particular problem for assignments that require a tightly prescribed form. A rubric may demand a formal thesis, consistent paragraph structure, conventional transitions and polished language. A detector may treat some of those same features as evidence of machine generation. Schools should investigate that possibility rather than assume that conformity to a writing rubric corroborates misconduct.

Good writing encompasses many styles. Neither polish nor predictability establishes who produced a passage. If a teacher expects students to adopt a particular academic register, that expectation belongs in the interpretation of the evidence.

What the Instructions Actually Prohibit

Even accurate identification of AI involvement leaves another question unanswered: was that involvement prohibited?

Research, brainstorming, outlining, drafting and editing are different activities. An assignment may permit some and prohibit others. Outside school, each can be a legitimate use of AI under the applicable rules. Inside school, the appropriate boundary should follow the learning objective.

A student practicing independent argument formation may need to develop a thesis without assistance. A student learning to evaluate sources may benefit from comparing an AI answer with the underlying evidence. A student practicing revision might be allowed to ask for feedback while remaining responsible for every change.

“Do not use AI to write this essay” does not necessarily explain all of those boundaries. It may leave students unsure whether they can ask for an explanation of a concept, receive comments on their draft or use a writing application that offers both grammar correction and generated rewrites.

The boundary becomes harder to infer as AI enters familiar tools. Google describes AI Overviews as a core Search feature and offers a Web filter for users who want results consisting of links. A student can encounter generated material while conducting an ordinary search, without first opening a chatbot. [7]

Encountering an overview, using it to locate a source and submitting its language as one’s own are different actions. A policy should distinguish them.

Spellcheckers, grammar tools and autocomplete provide useful precedents. Their acceptability depends on what is being assessed. A spelling assessment can reasonably prohibit spellcheck. A history assignment may have little reason to do so. As AI becomes more commonplace, schools will need to make similarly specific decisions about functions that students may otherwise assume are ordinary features of their tools.

Adults’ use of AI does not entitle students to outsource an exercise designed to develop independent reasoning. It does make the school’s explanation more important. Students should understand what they are being asked to practice and why a familiar form of assistance is restricted.

That clarity is also an access issue. A student whose main exposure to new technology comes through school may receive little opportunity to learn it if classroom policy consists only of prohibitions. Students with access elsewhere can continue developing those skills. This is a foreseeable equity concern that schools should examine, rather than an established outcome of every AI restriction.

Schools can protect independent practice while also providing supervised opportunities to learn responsible AI use. Both belong in an education intended to prepare students for the world they will enter.

What Hidden Prompts and Watermarks Can Establish

Some educators have tried to obtain more concrete evidence than a stylistic prediction.

Alcorn State University history professor Jason Gibson hid the word “Madagascar” in white font within an assignment prompt, placed where it made no contextual sense. Reporting by TODAY, carried by NBC Los Angeles, described students submitting answers containing the unexpected reference. [8]

This technique can provide a more specific reason to investigate. An irrelevant term planted in the prompt offers a possible account of how an answer was produced.

It remains a partial solution. Hidden text can become visible, be removed or fail to influence a model. Pasting as plain text strips formatting; it does not necessarily delete the underlying words. It may instead reveal words previously concealed by their color or size.

Educators also need to consider whether an accessibility tool or display setting could expose it, and whether permitted assistance could produce the same tell. By Gibson’s own account, 32 of his 35 students across two sections failed a portion of the midterm; two appealed, and one grade was changed after the student explained that dark mode had made the hidden text invisible on her screen. The appearance of a planted phrase warrants examination of the student’s process and the assignment’s rules. It does not supply that examination.

Provider watermarking takes a different approach. Instead of predicting authorship from familiar language, it builds a detectable statistical pattern into generated text. Article 50(2) of the EU AI Act establishes machine-readable marking and detectability obligations for covered synthetic outputs, with an exception for specified standard editing or changes that do not substantially alter the input or its meaning. The law’s obligations and voluntary codes used to demonstrate compliance are distinct. [9]

Anthropic says its Claude text watermark uses a version of Google DeepMind’s SynthID-Text approach. It also explains the limit: a watermark can indicate likely Claude involvement, but cannot distinguish generation from heavy editing. Short passages and lightly proofread text can provide too little signal; a complete rewrite can remove it. A negative result does not establish exclusively human authorship. [10]

Watermarking may therefore improve evidence about a tool’s involvement. It still cannot determine whether that involvement violated an assignment.

There is a further tradeoff to examine, but it should be stated carefully. Some watermarking configurations improve detectability at the expense of text quality. If a deployment creates poorer prose that users must repair, the additional human editing—or processing through another model—could impose costs and weaken the mark. That is a conditional concern, not an established effect of every deployed watermark. The SynthID-Text researchers reported quality preservation for their production configuration, including a large user-feedback experiment. [11]

Stylistic convergence also matters, principally for detection based on style. Models learn from human language; people increasingly work with model-generated language; future models may encounter that mixed material. A 2026 study of scientific publications reported convergence toward U.S. scientific English among publications classified as AI-assisted, while acknowledging that its classification used indirect lexical markers. [14]

It would be premature to infer that all writing will converge or that all detection will fail. It is reasonable to question whether fixed assumptions about what “AI writing” looks like will remain dependable. Keyed watermarks address a different signal and are not automatically defeated by ordinary stylistic resemblance.

Why a Safeguard Needs a Working Process

California’s Department of Education model AI policy says detection software should not be the sole basis for disciplinary action or a grade penalty. It is a model for local use, not proof that every district has adopted the same rule. [12]

In September 2026, a Santa Barbara Unified School District task-force group discussed treating detection as an initial step in a broader human review. The reporting described policy development, with a proposal intended for later board consideration. [13]

These are useful safeguards. Their effectiveness depends on what happens after the alert.

Adding “the writing seemed too sophisticated” to a detector score does not necessarily create independent evidence. Nor does requiring a student to disprove the accusation before a teacher will consider the drafting record.

A meaningful review should identify the specific restriction allegedly violated, explain what the tool’s output measures, examine available evidence of the writing process, and give the student a fair opportunity to respond. An appeal should be capable of changing the outcome.

Drafts, source notes, classroom work and a discussion of the argument can help. None is infallible. Version history can support a student’s explanation, but it should not become a new technological verdict—or a compulsory evidence archive that every student must maintain to avoid suspicion.

The school’s responsibility is to evaluate the evidence in context and explain its decision. That responsibility remains even when the software supplies a confident number.

Sources

1. T. Keung Hui, “North Carolina Student Fights Accusation of AI Use,” The Herald-Sun / The News & Observer reporting, republished by Government Technology, April 24, 2026. Read.

2. Destinee Patterson, “Wake County student says clear AI policies needed after being accused of cheating,” WRAL, May 5, 2026, updated May 6. Read.

3. Matter of Newby v Adelphi University, 2026 NY Slip Op 26021, Supreme Court, Nassau County, January 28, 2026. Official court opinion. Read.

4. “A Palo Alto high schooler was accused of AI cheating. His family filed a civil rights suit,” The San Francisco Standard, May 11, 2026. Read.

5. Turnitin, “Using the AI Writing Report.” Official product guidance. Read.

6. Weixin Liang et al., “GPT detectors are biased against non-native English writers,” 2023. Author paper with link to the published Patterns version. Read.

7. Google Search Help, “Find information in faster & easier ways with AI Overviews in Google Search.” Official product guidance. Read.

8. Rachel Paula Abrahamson, “A professor created a trap in his midterm to catch students using AI,” TODAY / NBC Los Angeles, July 28, 2026. Read.

9. Regulation (EU) 2024/1689, Article 50(2), EUR-Lex. Read.

10. Anthropic, “How Claude’s text watermark works,” August 14, 2026, updated September 1. Official technical explanation. Read.

11. Sumanth Dathathri et al., “Scalable watermarking for identifying large language model outputs,” Nature 634, 818–823, October 23, 2024. Read.

12. California Department of Education, “Model Policy: Artificial Intelligence In Education.” Read.

13. Sofia Wallace, “Students, parents, teachers want their voices heard on artificial intelligence in Santa Barbara schools,” Santa Barbara News-Press, September 25, 2026. Read.

14. “Does scientific writing converge to U.S. english? evidence from generative AI-assisted publications,” Scientometrics, 2026. Read.