Schools use a combination of automated detection software, manual review techniques, and institutional investigation procedures to identify AI-generated student work. No single method is foolproof, so most schools layer multiple approaches together before reaching a conclusion.
AI Detection Software Built Into Learning Platforms
The most visible layer of detection happens through software integrated directly into the learning management systems (LMS) where students submit their work. Turnitin, long known for plagiarism checking, now includes an AI detection feature that flags portions of text it believes were machine-generated. Copyleaks is another major player, offering AI detection plugins that work inside Canvas, Blackboard, Moodle, Google Classroom, Brightspace, Schoology, and Sakai. When a student submits an assignment through one of these platforms, the detection tool can automatically scan the text and generate a report for the instructor.
These tools use several analysis methods working in parallel. One approach, sometimes called “AI phrases,” looks for phrases that appear far more frequently in AI-written text than in human writing. Another, called “AI source match,” checks whether portions of the submission match AI-generated content that has already been published elsewhere online. Some tools also use a layered analysis that blends multiple detection methods to assess how likely it is that AI influenced a piece of text. Instructors can also upload assignment templates so the scanner ignores the original prompt or provided text, focusing only on the student’s response.
How the Software Measures “Human” vs. “AI” Writing
Under the hood, most AI detectors rely on two core linguistic metrics: perplexity and burstiness.
Perplexity measures how predictable a piece of text is. AI language models work by predicting the most likely next word in a sequence, which means their output tends to be highly predictable. A low perplexity score signals that the text follows patterns an AI model would produce. Human writers, by contrast, tend to make more unexpected word choices, use idiosyncratic phrasing, and structure sentences in less predictable ways, resulting in higher perplexity.
Burstiness refers to the natural variation in how humans write. People tend to mix short, punchy sentences with longer, more complex ones. They might dive deep into one idea, then shift abruptly to another. AI-generated text, on the other hand, often maintains a more uniform rhythm and sentence length throughout. It can also repeat themes, words, or structural patterns in ways that feel unnaturally consistent. When detectors see text that lacks the natural clustering and variation of human writing, they treat it as a red flag.
Manual Checks Teachers Use
Many instructors don’t rely on software alone. One of the most effective manual techniques involves checking a document’s edit history. When students write in Google Docs, for example, every keystroke and revision is logged in the version history. An instructor can open the document, click File, then Version History, and see exactly how the text was built over time. A genuine student draft will show dozens or hundreds of small edits: sentences added, words deleted, paragraphs rearranged across multiple sessions. An AI-generated submission, by contrast, typically shows very few entries, with large blocks of text appearing all at once, as if pasted in from an outside source.
Teachers also compare a suspicious assignment against a student’s previous work. If a student who typically writes short, informal paragraphs with occasional grammar issues suddenly submits a polished, highly structured essay with advanced vocabulary, that shift raises questions. Some instructors build this comparison into their course design from the start by requiring first drafts, outlines, and research notes submitted to a shared folder throughout the semester. This creates a paper trail of the student’s thinking process that is difficult to fabricate.
Oral follow-ups are another common technique. An instructor might ask a student to explain their thesis, walk through their argument, or define a term they used in the paper. A student who genuinely wrote the work can usually discuss it in depth, while someone who submitted AI-generated text may struggle to explain their own reasoning or recall specific sources they cited.
Why Detection Tools Aren’t Reliable on Their Own
AI detectors have serious accuracy limitations, and schools increasingly recognize this. OpenAI, the company behind ChatGPT, shut down its own AI detection tool after it correctly identified only 26% of AI-written text while falsely flagging 9% of human-written text as AI-generated. One study found that detectors identified ChatGPT text with 74% accuracy, but that rate plummeted to 42% when students made even minor tweaks to the generated content.
The tools are also surprisingly easy to fool. Researchers found that through simple prompt engineering, like asking ChatGPT to “write like a teenager,” Turnitin’s detection rate dropped from 100% to 0%. When researchers had ChatGPT polish genuinely human-written academic work to sound more scholarly, Turnitin failed to detect any AI involvement at all.
Perhaps most concerning, these tools disproportionately flag non-native English speakers. Stanford researchers discovered that while detectors performed near-perfectly on essays by U.S.-born eighth graders, they misclassified over 61% of essays written by non-native English speakers as AI-generated. Among TOEFL exam essays (written by international students proving their English proficiency), 97% were flagged by at least one detector. This happens because non-native speakers sometimes use simpler, more predictable sentence structures and vocabulary, which mimics the low-perplexity patterns detectors associate with AI.
What Happens When a School Suspects AI Use
When an instructor suspects AI-generated work, most institutions follow a structured process rather than issuing an immediate penalty. The first step is typically verification through multiple methods. Schools generally advise instructors not to rely on a single detection tool’s score. Instead, they’re encouraged to run the text through additional detectors, compare it against the student’s previous writing, and review any available drafts or edit histories.
If the evidence warrants it, the instructor meets with the student. During this conversation, the teacher may share the detection tool’s flagged sections and ask the student to explain their writing process. The discussion often focuses on specific qualities of the text: formulaic structure, repetitive word choices, or a noticeable departure from the student’s usual writing style. This meeting serves both as an investigation and an educational moment.
When the evidence is inconclusive, many schools offer the student a chance to redo the assignment. Given the known limitations of detection tools, this approach protects students from being penalized based on unreliable software output alone. Some instructors may assign additional work or require that the student complete the exercise under supervised conditions.
When the evidence clearly shows a student submitted AI-generated text in violation of course policies, the consequences mirror those for other forms of academic dishonesty: grade reductions, a failing grade on the assignment, or withdrawal from the course. Instructors are expected to document the investigation in writing, and the case may be referred to a dean or academic integrity office for further review. At many schools, repeated violations can result in suspension or expulsion under the same academic honesty codes that cover plagiarism and cheating.
How Course Design Is Shifting
Beyond detection, many schools are redesigning assignments to make AI-generated submissions harder to produce or easier to spot. Instructors increasingly require process-based evidence: annotated bibliographies built over weeks, reflection journals explaining research decisions, or staged submissions where each draft builds visibly on the last. Some courses have shifted toward in-class writing assessments, oral exams, or presentations where the student must demonstrate their knowledge in real time.
Other instructors take a transparency-based approach, allowing students to use AI tools but requiring them to document exactly how they used them, what prompts they entered, and what they changed in the output. This shifts the focus from catching AI use to evaluating whether the student engaged meaningfully with the material. The detection landscape is still evolving rapidly, and most schools treat their policies as works in progress, updating them semester by semester as the technology on both sides advances.