If you’ve ever copied a piece of writing into an AI detector, you’ve probably seen some pretty interesting results. You might paste in something you know was written entirely by a human and get a result claiming it is 80%, 90%, or even 100% AI-generated. Then you take the same paragraph and run it through another detector and suddenly you’re looking at a completely different result.
So which one is right?
That’s the problem.
AI writing detectors have become increasingly common as more people use ChatGPT, Claude, Gemini, and other generative AI tools to write everything from school papers to business emails. Schools use them to look for potential academic misconduct, publishers use them to evaluate content, and businesses are increasingly using them to determine whether applicants or contractors are relying too heavily on AI.
The problem is that an AI detector doesn’t actually know who wrote something.
It makes an educated prediction based on characteristics it has learned to associate with AI-generated writing. Sometimes that prediction may be useful. Sometimes it may be completely wrong.
That distinction is important, particularly when a detector produces a number that looks incredibly precise. Seeing “93% AI-generated” can make the result feel scientific, but that number does not mean there is a 93% chance that ChatGPT wrote the document. It is a prediction generated by a particular model using its own methodology, training data, thresholds, and assumptions.
In other words, AI detection is classification, not authorship verification.
And once you understand that, the entire AI detector debate starts to make a lot more sense.
How Do AI Detectors Actually Work?
The basic idea behind AI detection is fairly straightforward. Companies build machine-learning systems and train them using examples of human-written and AI-generated text. The system looks for patterns that tend to occur more frequently in one group than the other and then uses those patterns to classify new writing.
Exactly what those patterns are depends on the detector.
Early AI detection tools received a lot of attention for concepts such as perplexity and burstiness. Perplexity is essentially related to how predictable language is to a language model. If a sentence follows an extremely predictable pattern, it has lower “surprise” to the model. Burstiness, meanwhile, looks at variation in sentence structure and language across a larger piece of writing.
This made intuitive sense. Human writers tend to be inconsistent. We write a short sentence. Then a really long one. We occasionally use an odd phrase, change our sentence structure, make a grammatical mistake, or take a tangent that a machine wouldn’t necessarily predict.
AI-generated writing, particularly older generations of language models, could sometimes appear more uniform.
There is an important caveat here, however. You will still see articles describing perplexity and burstiness as though they are the complete explanation for how modern AI detectors work. That is no longer accurate.
GPTZero, for example, says that it moved away from relying primarily on perplexity and burstiness in 2023 and adopted a deep-learning architecture instead. Those concepts are still useful when explaining the history of AI detection, but modern detection systems are considerably more sophisticated than simply looking at how predictable a paragraph appears.
Modern detectors can analyze many different characteristics of writing and use models trained specifically to distinguish human and AI-generated text. The exact details are generally proprietary, which is another reason it can be difficult for an outside observer to determine exactly why a particular passage received a particular score.
Why Does Human Writing Get Flagged as AI?
This is probably the most interesting problem with AI detection, and it is something I’ve personally run into.
You can write something yourself, without using AI at all, put it into a detector, and receive a result claiming that the writing is AI-generated.
At first that sounds ridiculous.
But once you understand what these systems are looking for, it isn’t particularly surprising.
Think about the characteristics of good professional writing. It tends to be clear, organized, grammatically correct, relatively concise, and easy to understand. Most professional writers aren’t trying to make every sentence completely unpredictable. They’re trying to communicate an idea efficiently.
Unfortunately, some of those characteristics can also appear frequently in AI-generated writing.
Consider two business emails.
One person might write:
“Hey, just wanted to follow up on the project. I’ll have the updates ready Monday. Let me know if anything changes.”
Another might write:
“I wanted to provide a quick update regarding the project timeline. I’ll have the requested updates completed by Monday morning. Please let me know if there are any additional changes or requirements in the meantime.”
The second version is more polished and formal, but there is absolutely nothing about it that proves AI was involved.
Someone who has spent years writing professional emails might naturally communicate this way. A detector doesn’t know that person’s history. It only sees the finished text and compares its characteristics against patterns it has learned.
This is where the fundamental limitation comes into play: writing style is not the same thing as authorship.
The False-Positive Problem Is Bigger Than Many People Realize
False positives are one of the biggest reasons I would be extremely cautious about using an AI detector as definitive evidence.
A false positive occurs when a detector identifies human-written content as AI-generated. This isn’t simply something that happens occasionally with a poorly designed free website. Researchers have documented significant false-positive problems across AI detection systems.
A Stanford-led study published in Patterns examined seven AI text detectors and found that they incorrectly classified 61.3% of essays written by non-native English speakers as AI-generated on average. One detector incorrectly classified 97.8% of the essays in the sample.
Why would that happen?
One factor appears to be predictability. Writers who use English as a second language may naturally use simpler vocabulary and more predictable sentence structures. Those characteristics can overlap with patterns that AI detectors associate with generated text.
That creates a particularly troubling situation. A person can be completely honest about writing something themselves and still have a computer system effectively accuse them of using AI because their writing happens to look statistically similar to text produced by an AI model.
That’s a pretty serious limitation when the result is being used to make decisions about someone’s education, employment, or reputation.
AI Detection Gets Even Harder When Humans and AI Work Together
There’s another problem that becomes obvious when you look at how people actually use AI.
Most real-world AI-assisted writing isn’t simply 100% AI or 100% human.
A writer might create an outline themselves, use AI to brainstorm some ideas, write several paragraphs independently, ask an AI tool to reorganize another section, rewrite the result, add personal examples, correct the grammar, and then edit the entire article before publishing it.
At that point, what exactly is the detector supposed to identify?
The document contains both human and AI-assisted work.
This is becoming increasingly important as AI becomes integrated into normal writing workflows. People aren’t necessarily asking ChatGPT to “write my entire article” anymore. They’re using it as an editor, brainstorming partner, research assistant, copy editor, or second set of eyes.
A 2026 peer-reviewed study comparing Turnitin and Originality found that both systems had difficulty with hybrid human/AI writing. The study also found that detection performance changed depending on the genre of the writing and the length of the material being analyzed.
That is a much more realistic test of AI detection than simply asking whether a completely untouched ChatGPT response looks like AI.
The AI Models Are Constantly Changing
AI detectors also have a moving target.
A detection system might perform well against one generation of language models and then behave differently when a new model arrives.
ChatGPT has changed dramatically since the original GPT-3.5 era. The same is true of Claude, Gemini, and other systems. Modern models are much better at following a requested writing style, varying sentence structure, incorporating personal details, and producing writing that sounds less formulaic.
That creates something of an arms race.
AI companies improve their models. Detection companies improve their detectors. AI companies improve their models again.
There isn’t a permanent “AI writing fingerprint” that can simply be identified forever.
This is one reason claims that a particular punctuation mark, word, or writing habit is proof of AI should be taken with a very large grain of salt.
No, an Em Dash Doesn’t Prove Someone Used AI
We’ve all seen the jokes by now.
Someone uses an em dash and suddenly someone else says, “That’s definitely ChatGPT.”
The same thing happens with words such as “delve,” “comprehensive,” “robust,” “seamless,” “landscape,” “leverage,” and “transformative.”
There is some legitimate reasoning behind the observation. Language models do have vocabulary preferences, and some words may appear more frequently in AI-generated material than they do in certain types of human writing. GPTZero, for example, has developed tools that analyze vocabulary and stylistic characteristics associated with AI-generated writing.
But there is a huge difference between saying that a word is more common in AI-generated text and saying that using the word proves AI was involved.
Humans have been using the English language for a very long time.
The same goes for punctuation. An em dash isn’t an AI fingerprint. Neither is a semicolon, colon, perfectly grammatical sentence, or neatly organized paragraph.
If anything, the obsession with individual “AI words” demonstrates how easy it is to confuse a statistical tendency with actual evidence of authorship.
OpenAI Tried to Build an AI Detector Too
One of the most interesting examples comes directly from OpenAI.
In January 2023, OpenAI released its own AI text classifier designed to distinguish AI-generated writing from human writing. The experiment didn’t last.
OpenAI eventually discontinued the classifier because of its low accuracy. In its own published evaluation, the system correctly identified only 26% of AI-generated text in its challenge set and incorrectly labeled human-written text as AI-generated 9% of the time. OpenAI also warned that the classifier was not fully reliable.
That doesn’t mean AI detection technology is useless.
It demonstrates something more interesting: even the company developing some of the world’s most advanced language models found that reliably determining whether text had been generated by AI was extremely difficult.
The lesson isn’t “AI detectors can never work.”
The lesson is that the problem is much harder than a simple scanner looking for a hidden signature.
What About Turnitin?
Turnitin is probably one of the most recognizable names in this conversation because of its widespread use in education.
Its AI Writing Detection system is designed to identify text that may have been generated by AI and related systems. The important part is that Turnitin itself acknowledges the limitations of its technology.
According to Turnitin’s current documentation, its AI writing detection model may misidentify human-written, AI-generated, and AI-paraphrased writing. Turnitin specifically says that its AI score should not be used as the sole basis for taking adverse action against a student.
That is worth paying attention to.
If one of the largest companies in academic integrity software is telling its customers not to treat its AI detection score as definitive proof, it would be a mistake for everyone else to treat a random free detector’s “99% AI” result as an absolute fact.
What About GPTZero?
GPTZero is another major player in the AI detection market and has evolved considerably since it first became popular.
The company now describes its system as using a deep-learning model rather than relying solely on the older perplexity and burstiness approach. It also provides sentence-level analysis and additional information intended to explain why sections of text were flagged.
That additional information is useful because it gives the user more context than a single percentage.
But there is still an important distinction between explaining a prediction and proving authorship.
A detector can become considerably better at identifying characteristics associated with AI-generated text without ever becoming capable of looking at a document and knowing with certainty that ChatGPT produced it.
What About Originality.ai?
Originality.ai is another popular option, particularly among publishers, SEO professionals, agencies, and businesses.
It is worth paying attention to because professional content publishers have a legitimate reason to care about AI-generated material, plagiarism, and content originality.
The more useful way to evaluate Originality, however, is to look beyond the company’s own marketing claims and examine independent research.
A 2026 peer-reviewed study comparing Originality and Turnitin found that Originality performed better overall in that particular test, with an accuracy score of 0.69 compared with 0.61 for Turnitin. Neither system performed particularly well when the writing contained a mixture of human and AI-generated material.
That doesn’t mean Originality is “the best AI detector” under every circumstance.
It means that different tools can perform differently depending on what they’re being asked to detect.
That is a much more useful way to look at these products.
Why “98% Accuracy” Doesn’t Tell You Everything
This is where AI detection gets particularly technical.
You may see a company advertise that its detector is “98% accurate.” That sounds fantastic until you ask what that number actually represents.
Imagine a test containing 1,000 documents. If 900 of them are AI-generated and only 100 are human-written, a system can potentially achieve a very high overall accuracy while still performing poorly on human writing.
That’s why researchers use other measurements, including precision, recall, sensitivity, specificity, and F1 scores.
You don’t need to be a statistician to understand the underlying point.
Accuracy is only meaningful when you understand the test conditions behind it.
A credible AI detector should tell you what data was used, which AI models were tested, how much human writing was included, what types of writing were evaluated, and how often human writing was incorrectly flagged.
Without that information, a giant “98% accurate” badge doesn’t tell you nearly as much as it appears to.
Free AI Detectors Are Especially Difficult to Compare
There are now countless free AI detection websites.
Some are legitimate products offering a limited version of a larger service. Others provide very little information about how their systems work.
The easiest way to demonstrate the problem is to take one paragraph and run it through several services.
You might get something like this:
| Detector | Result |
|---|---|
| Detector A | 96% Human |
| Detector B | 82% AI |
| Detector C | 51% AI |
| Detector D | 100% Human |
| Detector E | 74% AI |
Now imagine that paragraph was written by you.
Which result should you believe?
The answer isn’t automatically the highest number, the lowest number, or the detector with the nicest-looking website.
You have to look at the methodology.
That’s why I would be cautious about articles that simply rank “the 10 best AI detectors” based on a few quick tests. The technology changes too quickly for a ranking like that to remain universally meaningful, and different detectors may perform very differently depending on the type of content being tested.
What Should You Look For in an AI Detector?
If you’re considering using an AI detector professionally, the company’s methodology should matter more than its marketing.
Look for independent testing rather than relying entirely on the company’s own accuracy claims. Check whether the researchers report false-positive rates, because incorrectly accusing a human writer can be just as important as correctly identifying AI-generated content.
It’s also worth checking whether the system has been tested against several different AI models and whether researchers have evaluated hybrid writing rather than only completely untouched AI output.
The type of writing matters too. A detector that performs well on college essays isn’t automatically going to perform equally well on technical documentation, journalism, marketing copy, scientific papers, or short business emails.
Finally, look for transparency. The more information a company provides about its testing methodology, limitations, datasets, and independent evaluations, the easier it is to make an informed judgment.
Short Text Is Another Problem
There’s a major difference between analyzing a 2,500-word article and analyzing a 15-word email.
A long article gives a detection system a lot more information to work with. It can examine vocabulary, sentence patterns, structure, repetition, transitions, and other characteristics across a much larger sample.
A short email might contain almost none of that information.
Consider:
“Thanks for the update. I’ll review everything and get back to you Monday.”
There simply isn’t much text there for a classifier to analyze.
If a detector tells you that sentence is 94% AI-generated, the number may look extremely precise, but there is very little evidence contained in those 11 words.
This is another reason I wouldn’t take an AI detector’s result on a short email, social media post, or isolated paragraph particularly seriously.
Can Editing Change an AI Detection Result?
Absolutely, and this is another reason detection scores should be treated carefully.
Research has shown that editing or paraphrasing AI-generated material can substantially change the output of AI detectors. Stanford researchers found that relatively small changes to writing style could significantly affect detection results.
That makes sense when you think about what the detector is actually doing.
It’s analyzing the characteristics of the text in front of it.
Change the text and you change the characteristics.
This doesn’t mean that every detector can be easily fooled, nor does it mean someone should try to manipulate a detector. It simply demonstrates that the result is tied to the statistical properties of the submitted text rather than some permanent identifier attached to the original writing process.
The Difference Between Detection and Provenance
This may ultimately be the most important concept in the entire discussion.
There is a huge difference between asking:
“Does this writing look like AI?”
and asking:
“Can we prove where this writing came from?”
The first is a classification problem.
The second is a provenance problem.
If you really need to establish authorship, the history of the document can often be much more useful than the final text itself.
Drafts can show how an article developed. Google Docs and Microsoft Word can preserve revision history. Source files can demonstrate how a project evolved. Notes, research materials, commits, and previous versions can provide context that a detector simply cannot see.
The finished paragraph doesn’t contain the complete history of how it was created.
Trying to reconstruct that history from the final text alone is inherently limited.
So Are AI Detectors Worth Using?
Yes, with the right expectations.
A detector can be useful as a screening tool. If you’re an editor and a document suddenly looks dramatically different from a writer’s previous work, an AI detector might provide another reason to investigate.
What it shouldn’t do is become the entire investigation.
A result that says “94% AI” shouldn’t automatically turn into “this person definitely used ChatGPT.”
That’s an enormous leap.
The better approach is to consider the detector alongside writing history, previous samples, revision history, source material, and the author’s ability to explain their work.
The detector becomes one piece of evidence rather than the judge, jury, and executioner.
What If Your Completely Human Writing Gets Flagged?
This is probably the most practical question for everyday users.
If you know you wrote something yourself and an AI detector says otherwise, don’t immediately assume you’ve somehow been caught by a hidden AI fingerprint.
Try another detector and see what happens.
If you’re dealing with something important, keep the original drafts and revision history. Having evidence showing how something was written is considerably more meaningful than arguing over whether a website thinks your prose looks “AI-like.”
And don’t get too hung up on individual words or punctuation.
If you naturally use an em dash, keep using it.
If you naturally write formally, that’s fine too.
You shouldn’t have to deliberately make your writing worse simply to convince a statistical classifier that you’re human.
That may be the strangest consequence of the current AI detector craze: people are starting to worry that writing clearly and professionally might make them look less human.
That’s backwards.
The Future of AI Detection Is Probably More Complicated
The idea that every document will eventually be classified as either “human” or “AI” doesn’t really match the direction technology is taking.
AI is becoming a normal part of creative and professional workflows.
A developer may use AI to generate boilerplate code and then rewrite most of it. A marketer may use AI to brainstorm headlines but write the article themselves. A designer may use AI to generate concepts and then build the final design manually. A writer may use AI to proofread an article they wrote from scratch.
Those are all different levels of AI assistance.
Trying to force all of those scenarios into a binary human/AI classification becomes increasingly difficult.
And that’s before we consider the fact that the underlying AI models continue to improve.
The technology is moving toward human and AI working together, not humans disappearing from the process.
What Should We Actually Care About?
Maybe the biggest question isn’t whether a particular paragraph was touched by AI.
Maybe we should be asking whether the final product is actually good.
Is the information accurate?
Was the research done properly?
Are the sources credible?
Is the content original?
Does it represent the company’s actual expertise?
Is the website accessible?
Does the code work?
Is it secure?
Is it performant?
Can another developer understand and maintain it six months from now?
Those questions matter a lot more than whether somebody used an AI tool to help rewrite a paragraph.
That philosophy is also why we’ve been developing our Controlled AI Integration (CAI) approach at Mediamato. AI can be extremely useful in web development and content production, but generated output still needs human review, testing, accessibility checks, SEO review, and technical accountability before it becomes part of a production website. Learn more about Controlled AI Integration at Mediamato.
AI is a tool.
It isn’t a replacement for knowing what you’re doing.
The Bottom Line: AI Detectors Are Signals, Not Proof
AI detectors have come a long way since the earliest attempts to identify ChatGPT-generated writing, and some modern systems are genuinely sophisticated.
That doesn’t mean they can determine authorship with certainty.
They are statistical classification systems trying to identify patterns associated with AI-generated writing. Sometimes they get it right. Sometimes they don’t. Their performance can vary depending on the AI model, the writing style, the length of the document, the language background of the writer, the genre, and whether the material contains a mixture of human and AI-generated content.
That’s why the most reasonable way to interpret an AI detector result is simply:
“This text has characteristics that resemble AI-generated writing.”
Not:
“We have proven that AI wrote this.”
Those are two very different statements.
And if a detector tells you that your completely human-written email is 94% AI-generated, don’t immediately start questioning your punctuation choices.
You might just be a predictable writer.
And honestly, sometimes predictable is exactly what good business communication is supposed to be.
How to Evaluate an AI Detector Before Trusting It
Before paying for an AI detection service or using one to make an important decision, look for independent research rather than relying solely on the company’s marketing claims. Check the false-positive rate, find out which AI models were tested, see whether human and AI writing were mixed together during testing, and look for results across different types and lengths of writing.
Most importantly, find out what the company’s percentage actually represents.
If a service tells you that something is 99.8% AI-generated, ask:
99.8% according to what model, what dataset, what testing methodology, and what false-positive rate?
That’s the question that matters.
Sources and Further Reading
The research and documentation behind this article includes work from Stanford’s Human-Centered AI group, peer-reviewed research published in the International Journal for Educational Integrity, OpenAI’s own documentation regarding its discontinued AI classifier, GPTZero’s documentation on its current detection methodology, and Turnitin’s documentation describing the limitations of its AI Writing Detection system.
For more on AI and web development, check out Mediamato’s articles on Controlled AI Integration, the risks of AI-generated code, and why custom WordPress themes with ACF can outperform page builders.
If you’re interested in having your website evaluated for performance, accessibility, SEO, maintainability, or AI-generated code, you can also explore Mediamato’s web development and digital marketing services.
Research
- Stanford Human-Centered AI – research into AI detector bias and false positives among non-native English writers.
- International Journal for Educational Integrity – 2026 comparative evaluation of Turnitin and Originality across human, AI-generated, and hybrid writing.
- OpenAI – documentation and evaluation of its discontinued AI text classifier.
- Turnitin – current documentation regarding AI Writing Detection and its limitations.
- GPTZero – documentation explaining its current detection technology and evolution from earlier statistical approaches.
Frequently Asked Questions About AI Detectors
Are AI detectors accurate?
AI detectors can be useful for identifying text that has characteristics commonly associated with AI-generated writing, but they are not 100% reliable. False positives and false negatives can occur, and results can vary considerably between detection tools. An AI detector’s score should therefore be treated as an indication rather than definitive proof of who wrote a piece of content.
Can AI detectors detect ChatGPT?
AI detectors can analyze ChatGPT-generated text and look for patterns associated with language models, but they cannot guarantee that ChatGPT was responsible for a particular piece of writing. Detection accuracy can also change depending on the ChatGPT model used, how much the output has been edited, the length of the content, and whether human and AI-generated writing have been combined.
Can human-written content be falsely flagged as AI?
Yes. Human-written content can absolutely receive an AI-generated classification. Research has documented false positives across AI detection systems, including cases involving non-native English writers. This is one of the biggest reasons an AI detector result should not be treated as conclusive evidence of AI authorship.
What is the most accurate AI detector?
There isn’t one AI detector that can be considered the most accurate in every situation. Tools such as GPTZero, Turnitin, Originality.ai, and others use different models and methodologies, and independent testing has found that their performance can vary depending on the type and length of writing being analyzed. When evaluating an AI detector, independent research and false-positive rates are more useful than a company’s advertised accuracy percentage.
Can an AI detector prove that someone used AI?
No. An AI detector can indicate that text resembles writing generated by an AI model, but that is different from proving how the content was created. Drafts, revision history, source material, previous writing samples, and other forms of document history can provide much stronger evidence of authorship than a detector score alone.
Why do AI detectors give different results?
Different AI detectors use different models, training data, detection methods, and scoring thresholds. They may also be designed around different types of writing or AI models. As a result, the same paragraph can receive very different scores from different services. This disagreement is one reason AI detection results should be interpreted cautiously.
Can AI detectors detect AI-edited or human-edited content?
Sometimes, but the reliability varies. Once AI-generated content has been substantially edited by a human, its characteristics can change enough to affect the detector’s classification. The same issue occurs with AI-assisted writing where only certain portions of a document were generated or modified by AI. Hybrid writing is particularly challenging for current detection systems.
Do certain words or punctuation prove that text was written by AI?
No. Words such as “comprehensive,” “delve,” or “leverage,” and punctuation such as em dashes, do not prove that AI was involved. Certain vocabulary and stylistic patterns may occur more frequently in AI-generated writing, but individual words or punctuation marks are not reliable evidence of authorship.
Should you trust an AI detector’s percentage score?
Treat it as a signal rather than a verdict. A score such as “94% AI-generated” represents the detector’s prediction based on its own model and methodology. It does not necessarily mean there is a literal 94% probability that ChatGPT or another AI system wrote the content. If the result has serious consequences, it should be considered alongside other evidence rather than used by itself.
What should you do if your human-written content is flagged as AI?
Don’t assume the detector is automatically correct. Try the content with another reputable detector, keep your original drafts and revision history, and compare the result with your previous writing. If you know you wrote the material yourself, the detector’s result doesn’t change that fact. It simply means the characteristics of your writing happened to resemble patterns the detection system associates with AI.