One of the challenges of teaching right now is the use of AI by students. As an instructor, I want my students to know how to use AI, but I also want them to understand core principles so they can use AI, and understand it’s results better. Unfortunately it seems that some students have embraced AI in a method of using it as a short cut from learning, while others avoid it all together for fear of “cheating”, and won’t be prepared for the world they walk into upon graduation.
Detecting improper use of AI is a real problem for instructors, especially for some types of classes. How do you know the right answer was generated via AI and not by hand? Coding classes, like what I teach, and math classes have this very real challenge.
As a side note: students have been cheating since the beginning of time. I’ve been catching cheaters for over 25 years – it’s just the methods have changed as the time and technology has changed. I can’t remember the last time I saw a crib note, but notes on computers/phones etc – well, that’s a different story now.
In other types of classes, some instructors think it might be easier to detect the use of AI by using an AI detector. After all, this is what AI Detection companies claim. However, the question is can an AI detector, by using AI itself, detect the use of AI in an assignment?
There are several critical reasons to distrust AI detectors. The technical and structural flaws of these tools are alarming. Those flaws are why those who are neurodivergent or non-native speakers get flagged at a higher rate than others.
However, the structural flaws of AI detectors also means that well written, academic style papers, have a very high chance of being detected as AI.
These same flaws are why notable false positives in popular works created before any form of AI was in existence. Wait till the end to see some examples of this.
So what are some reasons why AI detectors are unreliable?
The “Perplexity” & “Burstiness” Flaw
Most detectors rely on two main metrics to judge a text. Unfortunately, these metrics often penalize high-quality human writing, which is what formal and academic writing demand.
Perplexity (The “Surprise” Factor): This measures how “surprised” an AI is by your word choice. AI writes predictably (what’s the most likely word “B” that will follow word “A”); humans write chaotically. The more predictably your write – the more it looks like AI. Those with rule based thinking systems, like those on the autism spectrum have very defined structures. They tend to write more mechanically, and have a very organized, logical flow to their writing.
The Problem: If you write clearly, logically, and use standard academic phrasing, your “perplexity” score drops. Effectively, the better and more coherent your writing, especially in an academic setting, the more likely you are to be flagged as AI.
Burstiness (The “Rhythm” Factor): This measures sentence variation. Humans tend to vary sentence length. You will write a short sentence. Followed by a long, complex winding one that goes on and on, maybe even covering more than one point. AI tends to be monotonous, writing sentences of similar length over and over.
The Problem: Technical writing, legal documents, and strict academic papers require consistency. They are naturally “low burstiness,” causing detectors to flag them. Additionally, if you have anything that follows a writing pattern, like a poem, it will follow a pattern which can cause issues.
The “False Positive” Trap (The 99% Fallacy)
A detector might claim “99% confidence,” but that does not mean there is a 99% chance the text is AI.
Mathematical Impossibility: Because AI models are trained on human text, there is no unique “AI fingerprint” that only AI possesses. The detector is merely guessing based on probability, not finding evidence. The less it looks like standard human text, the more is assumes it is AI. It uses the flaws above to make a logical assumption on the likelihood of something being AI generated.
The “Innocent Victim” Bias: False positives disproportionately affect those who don’t try to cheat. Students who use “humanizers” or paraphrasing tools can often easily bypass detectors. This leaves the detectors primarily catching legitimate writers who happen to have a formal style and didn’t think they needed to “trick” the system.
This becomes a natural problem which leads into the next section as to what if your mind works very logically.
Bias Against “Standardized” or “Rule-Following” Minds
Autistic/ADHD brains tend to follow strict rules, preferring structure, and consistency. This means their works are very likely to be flagged.
But they are not the only ones. Students who are non-native speakers may have recently learned the formal rules, and are following them, instead of finding a voice and style of their own in what is a literal foreign language to them. This reduces the “perplexity” factor.
- Many neurodivergent and ESL writers prefer structure, clear rules, and lack of ambiguity.
- AI detectors view “adherence to structure” as a robotic trait.
- Therefore, a human who strictly follows grammatical rules and essay structures is often penalized for being “too perfect.”
The “Black Box” Problem
A black box in a computer system is one in which we give it data, and get a result, but do not know, or even care in some cases how the result was generated. In many areas of computer science, this is a good thing. However, when we have to justify that a student did or didn’t cheat, we need that box to be transparent because their academic livelihood and future job prospects just might depend upon it.
Consider a situation where a teacher accuses a student of plagiarism. The teacher can point to the source document (i.e. “You copied this paragraph from Wikipedia“) and validate their claim.
With AI detection programs, there is a profound lack of transparency in how these scores are generated. The accusation is “the computer feels like you didn’t write this.” This is also how the same document can be shown to multiple AI detectors, and get (in some cases wildly) different scores. How is anyone able to defend themself against a “feeling” from an algorithm that the developers themselves, let alone the people using it, cannot fully explain.
Unfortunately, I’ve seen professors assume that the AI detection tool is an all-knowing god, and be completely dismissive of the student, or outside reasons as to why the result might have been falsely generated.
Historical Proof
The ultimate counter-argument to these detectors is to provide it with examples of things which we know could not have been written by AI. Historical documents written 50 to 250 years before modern AI tools will often show false positives. Why, well these documents that we all learn and study are written in a formal, complex method, was used to train AI. By training AI models with these documents, they are the source of the AI generated content, not a result of it.
The Declaration of Independence
- Age: Written (1776) nearly 250 years before AI existed.
- Detectors’ Verdict: JustDone (73%), TextGuard (89%).
- Why? Because the Founding Fathers wrote in a highly formal, complex, and structured style. This is the very style detectors are trained to “detect”.
Other Examples with Scores
To test out the capabilities of some AI detectors, I uploaded several popular and well known texts to various tools to see if they detected the work as being generated by AI. All of these were clearly written by human hand (unless time-travelers went back in time and wrote them with AI), and I’ve included the year they were written.
| TextGuard | JustDone | Originality.ai | Scribbr | GPTZero | |
|---|---|---|---|---|---|
| Declaration of Independence (1776) | 89% | 73% | 57% | 62% | 63% |
| Gettysburg Address (1863) | 89% | 90% | 70% | 64% | 61% |
| Twas the Night Before Christmas (1822) | 86% | 73% | 63% | 70% | 53% |
Summary: Why Detectors Fail
There are lots of reasons why AI detectors are not the best tool to use.
| Feature of Writing | Human Intent | AI Detector Interpretation |
| Clear, Logical Flow | “I want my reader to understand this easily.” | “Low Perplexity. Likely AI.” |
| Consistent Structure | “I am following the essay guidelines perfectly.” | “Low Burstiness. Robotic.” |
| Formal Vocabulary | “I am using professional academic language.” | “Predictable tokens/words. Likely AI.” |
| Grammar Perfection | “I proofread this carefully.” | “Too clean. No human error detected.” |
There are other ways to ensure students are doing the work themselves, but it will take time, effort, and in some cases – compromises with other aspects of education. Which of those you’re willing to trade for, will have to be up to you.
The Challenges of AI Detectors was originally found on Access 2 Learn
