How to Check If a Student Used AI — A Step-by-Step Guide

A student submits an essay that reads differently from their usual work. The grammar is flawless, the structure is textbook-perfect, and the argument flows with a smoothness you have not seen from them before. Your instinct says something is off, but instinct is not evidence.

This is the core tension teachers face with AI-generated writing: suspicion is easy, but proof is hard. Accusations carry serious consequences, including academic penalties, disciplinary records, and damaged trust. Getting it right matters. Here is a practical, step-by-step approach for evaluating whether a student used AI, built on evidence rather than gut feelings.

Step 1: Look for Manual Signals

Before running any software, read the submission carefully and compare it to what you know about the student. Several patterns can indicate AI involvement without needing a detection tool.

Sudden Quality Jumps

A student who has been writing at a consistent level all semester and suddenly produces dramatically better work warrants closer attention. Pay particular attention to vocabulary sophistication, sentence complexity, and argument structure. Genuine improvement is gradual and usually shows in some areas before others. A leap across every dimension at once is unusual.

Voice Inconsistency

Every writer has a voice. Some students write in short, direct sentences. Others prefer longer, more complex constructions. If a submission reads like it was written by a completely different person than the one who wrote previous assignments, that disconnect is meaningful.

Look specifically at word choice. A student who normally writes "shows" and "talks about" suddenly using "demonstrates" and "elucidates" has either made a conscious stylistic shift or did not write those sections themselves.

Generic Examples and Missing Perspective

AI models produce examples that are illustrative but impersonal. A human student writing about social media and mental health might reference their own experience deleting Instagram for a month. An AI-generated response will cite "studies" and "researchers" without naming them, discussing the topic from a detached, encyclopedic distance.

If a submission on a personal or opinion-based prompt contains zero personal perspective, that absence is worth noting.

Suspiciously Clean Transitions

Human essays have friction points. Transitions can be slightly awkward, and paragraphs do not always connect perfectly. AI-generated essays flow with mechanical smoothness, every topic sentence perfectly placed. That kind of structural perfection is uncommon in authentic student drafts.

Step 2: Use a Detection Tool

Once you have noted any manual signals, run the text through an AI detection tool to add quantitative data to your observations.

Paste the full submission, not just a paragraph or two. Detection accuracy improves with longer text because the statistical measures need sufficient data. A 200-word excerpt can generate misleading scores that a full essay would not.

Review the results carefully. Good detectors provide more than a binary verdict. Look for:

• An overall probability score that indicates how likely the text is to be AI-generated, on a scale rather than a simple yes or no
• Sentence-level highlights that show exactly which sentences triggered AI signals and which read as human
• Multiple metrics such as perplexity, burstiness, entropy, and stylometric features, so you can see whether the text is flagging on one measure or across the board

A text that flags on every metric with a high probability score is a stronger signal than one that flags on a single metric at a borderline threshold. Multi-signal analysis reduces the chance of false positives and gives you a more nuanced understanding of the results.

Step 3: Compare to Known Student Work

Pull up previous assignments from the same student. Side-by-side comparison is one of the most effective detection methods available because it leverages context that no algorithm has.

Look at vocabulary level, sentence length patterns, citation habits, use of contractions, and how they structure introductions. If a student's previous essays used straightforward language and the flagged submission is full of advanced academic vocabulary, that shift is data.

This comparison does not prove anything on its own, but combined with detection results and manual observations, it builds a picture that is difficult to explain away if the work is not genuine.

Step 4: Have a Conversation

This step is the most important and the most often skipped. Before drawing any conclusions, sit down with the student and talk about their submission.

Ask open-ended questions: "Walk me through your argument in this essay." "Why did you choose this particular angle?" "Can you explain what you meant in this paragraph?" "What was the hardest part of writing this?"

A student who wrote the essay themselves can discuss their ideas fluently and elaborate on their points. A student who submitted AI-generated text often struggles to go beyond what is on the page. They may not remember their sources, may not be able to restate their thesis in different words, or may give answers that feel rehearsed.

Frame it as a discussion, not an interrogation. Many teachers routinely do brief oral follow-ups on written assignments because it serves as both a learning opportunity and a verification step.

What Not to Do

Getting this wrong can harm a student's academic career and your relationship with your class.

Do not rely on a single detector. No tool is 100% accurate. False positives happen, especially with formal writing, ESL students, and heavily edited text. A detector score is a data point, not a verdict.

Do not accuse without evidence. Present your observations as a conversation, not a confrontation. A student deserves to hear the specific concerns, not just "the detector flagged you."

Do not punish based on a score alone. A detection probability of 70% means a 30% chance the text is human-written. No student should face consequences based on a probability estimate without corroborating evidence.

Building a Proactive Approach

The most effective strategy is prevention rather than detection after the fact. For more on what tools are available, see our guide to the best AI detectors for teachers. Structural changes to assignment design can make AI use harder to hide and less tempting to attempt.

Require rough drafts. Collecting outlines, first drafts, and revision notes creates a paper trail. A student who submits a polished final essay but cannot produce any earlier version has a harder time defending authenticity.

Collect in-class writing samples. A baseline sample of each student's writing, produced in a monitored environment, gives you a reliable comparison point. Even a short in-class reflection provides useful data.

Assign oral presentations or defenses. Asking students to present their argument and discuss their writing process is the single most effective method for verifying authenticity. It also builds communication skills, serving a pedagogical purpose beyond detection.

Design assignments that resist AI. Prompts that require personal reflection, reference to specific class discussions, or analysis of materials that are not publicly available online are much harder for AI tools to handle convincingly.

Privacy Considerations

When using detection tools with student work, FERPA (the Family Educational Rights and Privacy Act) applies. Student submissions are educational records, and uploading them to third-party services raises legitimate privacy concerns.

Cloud-based detectors that transmit text to external servers may store or use that text in ways that conflict with student privacy protections. Some services explicitly state that submitted text may be used for model training.

Privacy-first tools that run analysis entirely on the user's device avoid this problem. ShaamAI Detector runs all analysis in the browser with no text ever sent to a server, making it a safer choice for student work. Before adopting any detection tool institutionally, review its data handling practices and confirm compliance with your school's privacy policies.

A Balanced Approach

AI detection is not about catching students in a lie. It is about maintaining the value of the work your students do and ensuring that grades reflect genuine learning. The best approach combines multiple sources of evidence: your knowledge of the student, manual observations, quantitative detection results, and direct conversation.

No single tool or method is sufficient on its own. But together, they give you a responsible, evidence-based process for addressing AI use without rushing to judgment or ignoring the issue entirely.