Using AI Detectors to Grade Student Work
This post summarizes an article originally published by ETS on November 22, 2024. Read the complete original article on ETS.
As students increasingly use generative AI in their coursework, educators face a difficult question: how can they distinguish a student’s own writing from text produced by an AI system? AI-generated text detectors can provide useful information, but their results require careful interpretation.
No detector is perfectly accurate. A tool may incorrectly flag authentic human writing, or it may fail to identify AI-generated content. Because either error can have serious consequences for a student, detector output should not be treated as definitive proof of misconduct.
Understand what the tool can—and cannot—do
Before using a detector, educators should examine the accuracy information provided by its developer. Reported performance should include both how often the tool identifies AI-generated text correctly and how often it mistakenly classifies human writing as AI-generated.
Performance also depends on the material being evaluated. Longer passages generally provide detectors with more usable evidence than short responses. Research discussed in the original article suggests that approximately 50 words should be treated as a minimum for reliable detection; attempting to classify a sentence or a few words is especially uncertain.
Human–AI collaboration complicates detection
Many assignments are not produced entirely by a person or entirely by an AI. A student may use AI to brainstorm, revise selected passages, improve grammar, or reorganize a draft. In these mixed-authorship situations, the label “AI-generated” becomes ambiguous, and detectors may not reliably identify the extent or nature of the assistance.
Educators therefore need clear policies that explain:
- which uses of AI are permitted for each assignment;
- which uses are prohibited;
- how students should disclose AI assistance; and
- why those distinctions matter for the skills being assessed.
Clear expectations are more useful than asking a detector to resolve a policy question after the work has been submitted.
Consider fairness and possible bias
Detector performance may vary across demographic and language groups. The direction and size of those differences are not always predictable, making it risky to rely on a detector for consequential decisions without additional review.
This is particularly important when evaluating multilingual writers or students whose writing differs from the data used to train a detector. A flag should prompt careful examination, not an automatic penalty.
Use multiple sources of evidence
Different detectors may produce different results for the same passage. Comparing more than one tool can reveal whether a result is consistent, but agreement among tools still does not prove authorship.
Whenever possible, educators should also consider evidence beyond the submitted text, such as:
- drafts and revision histories;
- notes or outlines;
- keystroke or writing-process data;
- a student’s previous work;
- conversations with the student; and
- other records of how the assignment was developed.
These sources can provide context that text-only detection cannot capture.
Treat detection as a signal, not a verdict
Responsible use of AI detectors requires restraint. Their best role is to identify cases that may deserve closer review, alongside other evidence and professional judgment. They should not be the sole basis for grading, accusations, or other high-stakes decisions.
As AI tools continue to evolve, educators will need to combine sound policies, transparent expectations, and multiple forms of evidence. The goal should be to support learning and academic integrity while protecting students from decisions based on uncertain technology.