How Do Professors Check for AI?
Learn how professors assess whether student work was generated with AI, including the role and limitations of AI detectors, writing evidence, and follow-up conversations.
Overview
Professors may check for generative-AI use by combining an AI-writing detector with ordinary academic review: checking whether claims and citations are real, comparing a submission with drafts or earlier work, and asking a student to explain their argument or writing process. The exact tools and rules vary by institution, course, and assignment. An AI detector estimates whether text resembles patterns associated with machine-generated writing; it does not provide a universally reliable record of who wrote the text or whether a student broke a course rule.
The practical answer to “can professors detect AI?” is therefore: sometimes they can identify suspicious or AI-like work, but no single score or stylistic impression reliably establishes authorship. Turnitin itself says its AI report can misidentify human, AI-generated, and AI-paraphrased text and should not be the sole basis for adverse action. Using the AI Writing Report – Turnitin Guides calls for further scrutiny and human judgment under the institution’s policy.
What professors use to check for AI
AI-writing detection tools
Some instructors or academic-integrity offices have access to AI-writing reports in systems such as Turnitin; other commercially available detectors have included GPTZero, Winston, ZeroGPT, and Originality.ai. These products are not universally available or used, and universities may set different rules about whether faculty can use them. A 2024 experiment, for example, tested five detectors on student-authored, AI-generated, and student-edited microbiology answers; that list describes the study’s test set, not a survey of what all professors use. Frontiers | Students are using large language models and AI detectors can often detect their use found substantial differences in performance among tools.
A Turnitin AI Writing Report marks qualifying prose it classifies as likely AI-generated or likely AI-generated and modified by an AI paraphrasing or bypassing tool. The displayed percentage represents the portion of qualifying text identified by the model—not the probability that a student cheated. Turnitin’s AI score is separate from its Similarity Report score. Its current guide says the detector requires at least 300 words of long-form prose, does not reliably assess code, poetry, tables, or bullet points, and displays an asterisk rather than a numerical result for detections below 20%, because false positives are more common in that range. Using the AI Writing Report – Turnitin Guides
Similarity checking is different
A similarity checker compares submitted text with material in its indexed sources and reports matching passages. It can help locate copied or insufficiently attributed text, but a match score does not by itself prove plagiarism—and a low match score does not establish that the writing is human-authored. Turnitin explicitly treats its AI-writing indicator and similarity score as independent measures. Using the AI Writing Report – Turnitin Guides
Human review and process evidence
Professors may notice a sudden change in style, an answer that does not address the assignment, references that cannot be verified, or methods and terminology that seem inconsistent with the course. Such observations can prompt questions, but style alone is a weak basis for a misconduct finding. A student’s writing can change for many reasons, and polished or formulaic prose is not proof of AI use.
To investigate, an instructor may compare a submission with drafts, outlines, earlier coursework, or other preparatory work; check whether sources and factual claims exist; and ask the student to explain their reasoning or how they developed the work. Cornell’s guidance identifies unverifiable citations, unexpectedly advanced methods, inability to explain submitted work, and inconsistency with drafts or prior work as forms of evidence to consider. It recommends using an initial impression as the start of an inquiry, not as its conclusion. AI & Academic Integrity | Center for Teaching Innovation
Writing-platform records can also document work in progress when a course uses a platform that actually records those steps. Such records are not universal: a professor cannot infer a student’s private browser history or external AI-account activity from an essay alone. Whether process records exist depends on the tools used for the assignment and the institution’s rules.
| Method | What it can show | What it cannot establish by itself |
|---|---|---|
| AI-writing detector | Text patterns its model classifies as likely AI-related | Who wrote the text or whether a course rule was violated |
| Similarity report | Passages matching indexed sources | AI authorship or plagiarism without review of the matches |
| Drafts and version history | Some evidence of how a document developed | A complete record of every tool or activity used |
| Conversation with the student | Whether the student can explain their ideas, sources, and process | Guilt based solely on hesitation, style, or imperfect recall |
The table summarizes limits described in Using the AI Writing Report – Turnitin Guides and AI & Academic Integrity | Center for Teaching Innovation.
How AI-writing detectors work—and what a score means
Most AI-writing detection systems analyze statistical and linguistic patterns in text and classify the text against examples or patterns associated with human and machine writing. Some approaches have examined how predictable a sequence of words is to a language model; researchers use the term perplexity for one measure of that predictability. A detector’s output is an inference from the submitted text, not a digital signature that necessarily identifies a particular AI service. Commercial systems may use different methods, and vendors do not always publish enough technical detail for independent users to reproduce their results.
OpenAI’s own short-lived classifier illustrates the distinction between a detector score and proof. In January 2023, OpenAI described a classifier trained on paired human and AI text. On its English challenge set, it correctly identified 26% of AI-written text as “likely AI-written” and mislabeled 9% of human-written text. OpenAI said the tool was unreliable on short text, code, and non-English writing, and that editing could defeat it. The company discontinued the classifier on July 20, 2023, citing its low accuracy. These figures describe that specific discontinued classifier and evaluation, not current commercial systems. New AI classifier for indicating AI-written text | OpenAI
Detector percentages can also be misread. If a report says that some proportion of qualifying prose was detected, that is not the same as saying the same proportion of the whole assignment was certainly written by AI, nor is it a probability that the student committed misconduct. The report does not determine whether AI use was permitted, whether it was disclosed, or whether the use violated the assignment instructions. Those are policy questions that require context.
How accurate are professors and detectors?
Human readers can sometimes recognize AI-generated work, but intuition is not dependable enough to settle a case. A study presented college instructors and students with six pairs of psychology essays. Each pair contained one essay written by a college student during a proctored exam and one produced by ChatGPT in response to the same prompt. Among 140 college instructors, the average correct identification rate was 70%; 145 students averaged 60%. ChatGPT, tested in separate runs, averaged 63%. Random guessing in a two-choice task would yield 50%. The instructors therefore did better than chance in this experiment but still misidentified the source in roughly three of every ten choices. The authors used a convenience sample and a small set of essay pairs, so the result is not a universal estimate of faculty accuracy. Can you spot the bot? Identifying AI-generated writing in college essays | International Journal for Educational Integrity | Springer Nature Link
Automated systems can separate some AI-generated text from human writing under controlled test conditions, but results depend on the model, language, text length, assignment, and how much the text has been edited. In a 2024 experiment with 153 students in an introductory microbiology course, students wrote approximately 500-word responses, then obtained AI-generated versions and attempted to edit them. Several detectors differentiated the original human answers from raw AI output; however, their false-positive rates on human writing varied, and edited AI text was harder to classify. In that study, reported false-positive rates for four tools ranged from 9.8% to 45.8%. Those rates apply to the tested systems and study materials, not every detector or current version. Frontiers | Students are using large language models and AI detectors can often detect their use
Other experiments have found different results. A 2023 study of 14 detection products reported false positives and false negatives, with results affected by machine translation and text alterations. The authors argued that the tools tested were unsuitable as stand-alone evidence of academic misconduct. Their findings reflect products and test conditions at that time; they should not be treated as a measure of every newer system. Testing of detection tools for AI-generated text | International Journal for Educational Integrity | Springer Nature Link
NIST’s later benchmark work offers a counterpoint to claims that detection never works: its 2025 report says detector performance varied significantly by generator and discriminator, that some detectors identified nearly all tested generators while some generators fooled most discriminators, and that discriminator systems improved over multiple rounds. The study used a curated benchmark of human- and machine-generated article summaries, not a universal sample of student essays or real-world misconduct cases. It supports the conclusion that detectors can improve and perform well in particular test settings, but does not establish that a score proves who wrote an individual assignment. 2024 NIST GenAI (Pilot Study): Text-to-Text Evaluation Overview and Results | NIST
Why AI checks can be wrong
False positives and false negatives
A false positive is human-written text incorrectly flagged as AI-generated. A false negative is AI-generated text that the detector fails to identify. Both matter in education: false positives can lead to an unjust accusation, while false negatives can let unauthorized use go unnoticed. Since detectors generally return classifications rather than verifiable records of a student’s activity, their output needs corroboration.
A 2023 study found a particular fairness concern in seven then-available detectors: the systems often classified human-written TOEFL essays by non-native English writers as AI-generated. In that test, 89 of 91 essays were flagged by at least one detector and 18 were flagged by all seven. The authors linked the pattern in part to text predictability and cautioned against using the tested detectors in evaluative settings. This was a bounded study of older systems and specific samples, not proof that every detector treats every multilingual writer the same way; it does show why a score should not be treated as neutral evidence without scrutiny. GPT detectors are biased against non-native English writers
Changing tools and incomplete coverage
Text generation systems and detection systems change. NIST found meaningful differences among generators and detectors, and Turnitin’s product documentation describes updating its model as new language models emerge. Results produced at different times—or by different detector versions—may therefore not be directly comparable. 2024 NIST GenAI (Pilot Study): Text-to-Text Evaluation Overview and Results | NIST Turnitin product updates – Turnitin Guides
Detectors may also miss short submissions, non-prose formats, or work that differs from the material on which the tool was evaluated. Turnitin, for example, specifies a minimum of 300 words of qualifying prose and says it does not reliably detect AI writing in code, poetry, or other short-form and unconventional formats. Using the AI Writing Report – Turnitin Guides
What happens after a professor sees a suspicious result?
A responsible review starts by checking the assignment’s rules: whether AI was prohibited, allowed for limited tasks, or allowed with disclosure. A detector cannot answer that question. The instructor may then check the flagged passages, verify citations and factual claims, compare relevant drafts, and ask the student to explain the ideas and process. Whether a formal report or sanction follows depends on the institution’s procedures and evidence standards.
University policies illustrate the variation. Cornell recommends gathering objective evidence and considering a conversation with the student. Columbia says an AI-detector report cannot be the sole proof in its disciplinary process. San José State University likewise states that AI-detection results alone cannot support an academic sanction and requires the instructor to present evidence and give the student an opportunity to respond. These are institutional examples, not a single rule binding every university. AI & Academic Integrity | Center for Teaching Innovation Generative AI Policy | Office of the Provost Academic Integrity | Student Conduct and Ethical Development
This is a general description of a careful review, not a mandated procedure. Cornell, Columbia, and San José State publish different details about how instructors should handle suspected violations. AI & Academic Integrity | Center for Teaching Innovation Generative AI Policy | Office of the Provost Academic Integrity | Student Conduct and Ethical Development
What students should take from this
Students should read the syllabus and assignment instructions rather than assume that all AI assistance is either permitted or prohibited. If the rules are unclear, asking the instructor before using a tool is safer than guessing. Where use is allowed, students should follow disclosure or citation requirements and verify any sources or claims the tool supplies. Keeping notes, outlines, drafts, and source records can help explain a genuine writing process if questions arise; it is not a guarantee against an incorrect flag.
In short, professors may use detectors, source-matching tools, their knowledge of a student’s prior work, and follow-up questions. These methods can guide an inquiry, but none of them—especially an AI-detector score or a change in writing style—should be treated alone as proof. Course policy and corroborated evidence determine whether the use was unauthorized.