Can you spot an essay written by AI?

Generative AI has undoubtedly become one of the most disruptive pieces of technology in classrooms over recent years. In higher education, instructors are increasingly confronted by the challenges it poses to evaluate learning. In particular, the question of whether instructors can truly tell if an essay was written by AI has quickly become one of the defining pedagogical anxieties of our era. Currently, much of the public discourse has centered on detection: are there “fingerprints” that reveal AI writing? Should schools invest in better AI detection software? And, perhaps the most daunting question for educators, can we truly distinguish between human-written and machine-generated prose? 

In my experience, the answer to these questions is both yes and no. 

Yes — because large language models (LLMs) often write in recognizable patterns, mimicking polish and institutional academic tone without necessarily demonstrating originality or depth. These patterns become more obvious after grading hundreds of undergraduate essays. Instructors may pick up some recurring tendencies of AI writing. LLM AIs tend to write in highly balanced sentence structures and polished transitions. Someone untrained and unaware of these tendencies would find no issue in a generated passage. But, an overuse of “not only X, but also Y” structure? Frequent deployment of lists-of-threes? Even punctuation habits — the (now) infamous em dash — can raise an eyebrow from experienced instructors who just feel that something is off with the passage they had just read. 

No — because I should emphasize that none of these features should be taken as definitive proof of AI writing. Many strong student writers also use impeccable grammar and polished transitions, lists-in-threes, or em dashes. To say there is a definitive checklist approach to AI detection is unreliable, and at times dangerous. No instructor would want to falsely accuse an honest student, especially in an age where AI writing is becoming increasingly sophisticated. 

I recently clicked into a subreddit frequented by Canadian high school students. Students were sharing advice of how to write “not like AI”. Many expressed discontent that their papers were falsely flagged as AI-generated, and they had to adapt by avoiding certain grammatically complex phrases while artificially increasing sentence variety. This sort of behavior points to not just anxiety from being surveilled by black-box detection logic, but also a reconsideration of how students prioritize aspects of their learning journey. Students are now learning how to “perform humanness” for algorithmic scrutiny, which distorts writing pedagogy itself. 

A strange standardization 

Nevertheless, instructors are not machines and cannot mathematically compute the likelihood of AI writing. What experienced instructors often recognize is not a particular “tell,” but a broader mismatch between the student and the level of writing. AI-generated essays often come with a generalized fluency, characterized by clean grammar and balanced rhetoric. However, this is rarely the case for authentic student thinking, which tends to possess characteristics that are more uneven and unrefined. While the prose, on the surface, may sound competent, it does so in a strangely standardized fashion. This tends to be that “fishiness” instructors pick up. 

In practice, suspicion of AI writing rarely pops up as a “gotcha” moment. Rather, it is often a cumulative impression formed through familiarity with the student’s prior work and classroom participation, where candid learning can be more easily observed. In other words, the strongest form of AI detection available to instructors is not software, but pedagogy.  

This realization came to me in my sociology course this year. Early in the school year, students were required to complete a handwritten in-class writing assignment responding to a specific prompt. Because the exercise was done on the spot and without digital assistance, I was able to gain a direct sense of how each student writes in their own, authentic voice — how arguments were organized, evidence was cited and personal styles were manifested. Later, that baseline became invaluable. When a student who had struggled with basic university-level writing suddenly turned in a polished graduate seminar paper, it raised legitimate questions. 

Good pedagogy is the best defence 

I remain skeptical toward the claim that better AI detection will solve the problem. Detection software can assist instructors but should never be allowed as definitive evidence. More importantly, overreliance on surveillance risks turning education into an adversarial process that students learn to circumvent instead of developing authentic critical thinking skills. Rather, instructors should focus on designing assignments to make student thinking visible – scaffolded assignments, low-stakes in-class writing, oral discussions and opportunities for students to take over and demonstrate, in their own terms, how their ideas develop over time. In an age of generative AI, it may be that the most reliable academic safeguard is not technology at all. It may simply be good pedagogy and meaningful familiarity with students’ thinking and writing.