When ChatGPT became publicly available in November 2022, marketers and executives largely split into two camps. Some became obsessed with AI content generation (“Automate everything!”). And others swung the other way, fixating on AI detectors to make sure no machine-written text ever reaches their websites.

As a content writing agency, we occasionally get requests from clients to run our texts through AI checkers. While we have no issue doing that because we’re confident in the quality we deliver, the importance some companies place on the AI score feels completely misplaced.

So, let’s look at what’s behind AI detectors, how much you can trust their results, and why chasing “100% human” scores might be distracting you from what actually matters.

How AI detection tools “judge” the writing

AI detectors all make the same promise: to tell you whether a piece of writing came from a human or a machine. But can they really know that? Not quite. They just make a probabilistic prediction, better known as a guess.

That guess comes from a machine learning model running behind the tool. It’s trained to spot patterns in massive datasets of both human and AI-written text. When you upload your piece, the tool scans it, compares it to what it has learned, and turns those signals into probabilities. The result is a percentage score, usually labeled as “likelihood of AI-generated content.”

What kind of patterns are these tools looking for? They usually focus on a mix of perplexity, burstiness, and repetition. 

Perplexity 

Perplexity measures how predictable the text is for a language model or, put simply, how expected each next word feels.

For example:

• The dog ran in the park. — low perplexity

Every word is highly predictable with no surprises in phrasing or meaning.

• The dog ran, philosophically speaking, from the concept of Mondays. — high perplexity

The sentence breaks logic and expectation because dogs don’t usually run from abstract ideas (at least, not that we know of).

Remember those viral lists of “words to avoid so AI detectors don’t catch you”? That’s not because detectors scan for specific phrases. It’s because overused expressions make your text more predictable, lowering perplexity, which, in turn, makes it look more like AI writing.

Burstiness

If perplexity looks at what the text is about, burstiness looks at how it is written, measuring the rhythm, pacing, and variation between sentences.

Humans tend to mix things up: a short line here, a longer thought there, sometimes even breaking a rule or two. AI, on the other hand, often sticks to evenly structured sentences that feel smooth but strangely flat.

For example:

• The dog ran in the park. It saw a ball and barked loudly. — low burstiness

The sentences are similar in length and structure.

• The dog sprinted across the park, slicing through the sunlight. For a heartbeat, everything went still. Then, out of nowhere, a ball rolled across the grass. Thrilled for the reasons only it understood, the dog exploded into barking. — high burstiness

The text combines short and long sentences, uses varied syntax, and keeps the rhythm changing, much like how people actually write or speak.

In short, the more diverse the sentence flow, the lower the chances the text gets flagged by AI detectors.

Repetition

Besides word choice and rhythm, AI detectors also look at how often a writer repeats the same patterns across the piece. When too many lines start in a similar way or follow a similar structure, the text sounds mechanical, and that’s when detectors get suspicious.

For example:

• The dog is running across the park. [Later in the text] The dog is barking loudly. [Even later] The dog is chasing a ball. — high repetition

Every sentence uses the same verb form (“is + -ing”) and almost identical phrasing.

• The dog runs across the park, barking as it goes. [Later in the text] A sudden noise sends it skidding to a stop, paws digging into the grass. [Even later] Excited, the little pup returns with the prize, tail wagging and eyes bright with triumph. — low repetition

The sentences start differently, use varied structures, and shift tone.

Repetition is related to burstiness, but they’re not the same. Burstiness focuses on the rhythm and flow of sentences, while repetition tracks patterns and phrasing across the entire text, regardless of pacing.

Many AI detectors confirm that their algorithms rely on these three principles. QuillBot and Winston AI, for example, openly mention perplexity and burstiness as part of their models, while Copyleaks also lists repetitive phrasing and lack of personal style as its key detection signals. Does that mean these AI detectors produce similar results? The short answer is no. But I’ll dig into that (and show you examples) in a few sections.

The blurred line of AI detector reliability

So, let’s say the detector did its job: it scanned your text, ran the calculations, and gave you a score or percentage. Maybe it even highlighted sections in red for “AI” or green for “human.” Looks clear and reliable, right? Not really.

Probability says nothing about quality

AI detectors don’t actually understand what they’re reading. They simply measure probability using math. Logic, meaning, and factual accuracy aren’t part of the equation.

Because of that, completely illogical or chaotic text, which is less predictable and has high perplexity, can sometimes slip past AI detectors more easily than clear, well-structured writing. At the same time, trying to make a well-written text “less predictable” just to avoid being flagged can leave it messy and unreadable.

So, a low AI score says nothing about the actual quality of the text: neither the ideas nor the style. It simply means the text happened to get a low AI score. That’s it.

Different AI checkers provide different scores

Another issue is that the probability scores produced by these tools are wildly inconsistent. Even though most AI detectors rely on the same general principles, each uses its own model, dataset, and thresholds, which leads to very different outcomes.

TechCrunch confirmed this inconsistency by comparing several popular AI detectors, including GPTZero, Copyleaks, GPT Radar, and Originality.ai. For example, a Claude-generated news article was labeled as fully human by most tools, except GPTZero, which got it right. But even GPTZero missed two out of eight samples in the same test (that’s a 25% failure rate), and that was still the best performance among all tested tools.

But the real question isn’t which one to trust. It’s whether you can trust any of them at all. Even OpenAI, the creator of ChatGPT, admits that its own models can’t reliably tell whether something was written by AI. They just make up random answers with no factual basis.

The margin of error is huge

The key thing you should remember when looking at any AI detection score is that these tools don’t really check whether a text is AI-generated or not — because they can’t know that. What they do is measure if the text looks like something an AI might write. And that leaves a lot of room for error.

In general, the mistakes AI checkers make fall into two categories:

• False negatives, when AI-generated text slips through undetected, especially if it’s been lightly edited
• False positives, when human-written text gets incorrectly flagged as AI-generated

We’ll break down both of these further on, but here’s a quick glimpse into how big that margin of error can be. The US Constitution has been repeatedly flagged as AI-generated — a case discussed in multiple articles, including those from Forbes and Ars Technica. 

When running it through ZeroGPT myself, I got a 98% “AI-generated” score. Ouch, bad news for the Founding Fathers.

How AI detectors get fooled by “humanized” text

Earlier, I mentioned that a low AI score doesn’t mean a text is good. But it also doesn’t mean it’s human. Even if you run it through multiple detectors and they all confidently say “human-written,” that result doesn’t prove anything. Why? Because AI detectors are surprisingly easy to fool. 

One study found that while detectors identified human writing with about 96% accuracy, their success in spotting ChatGPT text dropped from 74% to 42% once the content was slightly edited. 

Overall, AI-generated text can be “humanized” enough to bypass detection with the right tools, smart prompts, or minor tweaks. 

AI humanizers 

AI humanizers are the quickest (and laziest) way to avoid getting flagged by AI checkers. You just paste in your text, hit a button, and the tool does its “magic” by:

• Rephrasing sentences to raise perplexity
• Changing sentence structure to add burstiness and reduce repetition
• Dropping in minor grammar quirks or filler words to sound more natural
• Using invisible tricks like extra spaces, hidden characters, or special symbols
• Swapping letters, for example, using a different version of “I” from another language

A Washington Post investigation found that Turnitin’s AI detector failed to recognize ChatGPT-written papers after they were run through Quillbot, a simple paraphrasing tool that lightly rewords sentences. 

Essentially, running the text through AI humanizers doesn’t really make the writing better. It doesn’t even make it sound more human to real readers. These tools simply play around with the math behind AI detection models.

Prompts and minor tweaks

A few clever prompts can also make an AI-generated text look more “human” to detectors. For example, a writer can ask the model to:

• Make the tone more informal and conversational
• Vary the rhythm, like something written in a hurry
• Add some humor or personality

Cat Casey, Chief Growth Officer at Reveal and a member of the New York State Bar AI Task Force, shared that she can bypass most AI detectors 80-90% of the time just by adjusting her prompts, sometimes just by adding a single word like “cheeky.”

The same logic applies to manual edits. Rephrasing sentences, shifting their structure, or inserting small stylistic quirks can often be enough to throw off an AI checker. 

But the issue is that all these tricks only change how the text looks, not what it says. The depth, logic, and originality remain the same. It’s like building a house out of cardboard and painting the facade to look like it’s real — convincing from a distance, but hollow underneath.

When observing the ongoing race between AI generation and AI detection, one thing becomes clear: no matter how advanced the tools become, writers will always find ways to outsmart them. And when that happens, detectors produce false negatives. Rely too much on those scores, and you’re basically giving low-value fluff a free pass onto your website.

Why you can get a high AI score for human-written texts

The problem of false positives, when a human-written piece gets flagged as machine-generated, is just as common — and just as misleading. Aside from the US Constitution mentioned earlier, the Bible and Shakespeare’s works have been reported to receive high AI scores. 

While it’s tempting to think this only happens with older writing because it’s structured and predictable, that’s not the case. Modern texts can trigger the same false alarms just as easily. In a Washington Post test, the Turnitin AI checker analyzed 16 essays written by high school students and got fewer than half of them right.

So, why do detectors mistake real writing for AI? There are a few reasons behind it.

AI models train on human writing

The main reason human writing gets flagged as AI is simple — AI learned to write from us. Models like ChatGPT were trained on massive collections of human-written text, and one of the main goals of that training was to make them sound as human-like as possible.

OpenAI confirms this in its research papers, explaining that GPT models are trained to predict the next word in large text datasets from the web (which, at least until recently, were mostly produced by humans). They’re also fine-tuned through human feedback to make their tone and phrasing even more natural.

AI scores are just statistics

When AI detectors scan text for “AI-like” traits (predictable word choices, neat sentence structures, or repetitive phrasing), they’re really just measuring how statistically average the writing is. And much of human writing falls into that same range.

A 2023 study tested several detectors on both AI- and human-written samples. The tools identified GPT-generated content with some success, showing fairly consistent results for GPT-3.5 and less consistent for GPT-4. However, they often misclassified human-written texts, producing false positives and uncertain results.

Another study focused on OpenAI’s own detection model, examining how it handles real scientific writing. Researchers analyzed abstracts from academic journals published between 1980 and 2023 and found that around 5-10% were wrongly flagged as AI-generated with a high degree of confidence.

So when a detector labels the writing as “AI-generated,” it’s not because it somehow knows it came from a model. It’s simply because the line between human and machine writing has become too blurred for the math to tell the difference.

Over-editing can make human text look AI

When human writing is heavily edited to sound flawless, especially with the help of specialized tools, detectors can mistake that “perfection” for AI-generated text.

As one educator told Inside Higher Ed, students often deny using AI, only to later admit they relied on rephrasing tools to “clean up” their essays. They thought they were just improving clarity, not realizing those same edits could make their writing look algorithmic.

The irony is that the more effort a writer puts into making their text sound professional, the more likely it is to trigger the very systems designed to catch AI.

Detection tools carry built-in biases

AI detectors don’t treat all writing equally. Studies show that certain groups of writers are far more likely to be flagged as “AI” simply because of how they naturally write. These include:

• Non-native English speakers. Research comparing essays from US students and non-native English learners found that AI detectors misclassified more than half of the non-native essays as AI-generated, with an average confidence rate of around 61%. In the same study, nearly 20% of those essays were flagged by every single detector tested.
• Neurodivergent authors. Writers with autism, ADHD, dyslexia, and other forms of neurodivergence are also prone to false positives. Their writing may include repeated phrasing, unusual structure, or fewer personal pronouns, and these are the traits that AI detectors often read as “too consistent” or “impersonal.” In one case, a university professor was even mistaken for a bot for writing in a neutral, emotionless tone.

These false positives likely occur because of low perplexity, which makes texts look “too predictable” to AI checkers. But that doesn’t change the fact that it’s an unfair penalty on writers whose natural style falls outside the narrow patterns AI detection models were trained to see as “authentic.”

Putting AI detectors to the test: What our experiment showed

Reading dozens of studies about AI detection was interesting, but not enough. I wanted to see for myself what happens when you test human-written and AI-generated texts side by side. So, I ran a small experiment of my own. 

Step 1. The “control” text — a pre-AI article

I started with an article titled “IT Vendor Management: What It Is and How to Get It Right.”

It’s 2,373 words long and was written back in August 2021, more than a year before ChatGPT appeared. So, it’s undeniably human.

I fed this article into five different AI detectors, all free or freemium versions, because I wanted to see what an average user would experience. The lineup looked like this:

• ZeroGPT Checker
• Grammarly
• QuillBot AI Checker
• Copyleaks AI Detector
• Originality.ai
  

Some of these tools have strict word limits, so when the article exceeded them, I just trimmed it until it fit.

Step 2. The AI-generated text

Next, I asked ChatGPT (GPT-5 model) to write a new version of the same article. The prompt included the exact title and outline of the original article. I also provided the following information on the context:

• The client is a software outsourcing company
• The intro should mention relevant research
• The tone should be easy-to-read yet professional
• The message should emphasize that with reliable IT partners, vendor management becomes easier
  

The model produced an 896-word article, which I ran through the same five AI detectors with no editing or tweaks.

Step 3. The “humanized” version

After that, I copied the entire AI-generated text into an online humanizing tool called Humanize AI. There was no special reason for that choice. I just picked one at random.

The tool gave me a slightly longer text (917 words), which I also tested using the five detectors.

Step 4. The “tuned by prompts” version

Finally, I tried a more hands-on approach. I asked ChatGPT to rewrite the AI-generated article in two iterations using prompts like:

• Make the style more human-written
• Avoid clichés
• Add depth to arguments
• Use sentences of different lengths
• Make wording more unique
• Add research where relevant
• Use a less formal tone
  

That gave us a 983-word version, which I submitted to the selected five detectors.

Step 5: Comparing the results

In the table below, you’ll see how each detector scored all four versions.

In four out of five cases, the human-written article was correctly marked as human, and all five raw ChatGPT versions were caught as AI. So far, so good.

But simply running the AI-generated article through Humanize AI dropped its AI score so much that, in three out of five detectors, it looked more human than the actual 2021 article. And the version I fine-tuned with prompts passed as “human” with three detectors as well. 

The main issue, though, is that the results were anything but consistent. Quillbot AI, for example, correctly spotted the human-written text but failed to recognize both altered AI versions. Copyleaks, on the other hand, caught the prompt-tuned article as AI but missed the one run through the humanizer tool entirely.

All this means you can’t get a reliable AI score. 

But should it bother you at all?

Why focusing too much on the AI score is the wrong approach 

In content marketing, the goal isn’t to get a “100% human” label. It’s to create content that ranks well on Google and brings traffic to your site. Of course, that content also needs to deliver value — engaging readers, driving conversions, and building your brand’s credibility. But none of that matters if no one ever sees your content in search results.

And here’s the twist: Google doesn’t actually care whether your content was written by a human, an AI, or a mix of both. In fact, its official Search guidance clearly states that using AI is fine, as long as the result meets quality standards.

Those standards come down to the well-known E-E-A-T framework: Experience, Expertise, Authoritativeness, and Trustworthiness. To rank well, your content needs to reflect all four, and that’s where purely machine-written text usually falls short.

AI-generated articles without meaningful human input tend to sound surface-level, repetitive, and impersonal, which puts them at risk of violating Google’s spam policies.

So, in practice, using AI isn’t the problem. It’s how writers use it that matters. If you flood the web with low-value content that happens to “pass” AI detectors because it’s been slightly altered, it won’t rank or bring you any real results. But if you use AI to speed up research, drafting, or editing, and still put in the effort to make your content accurate, persuasive, and genuinely useful, you’ll be fine, even with a high AI score.

After all, if Google doesn’t care whether a piece contains AI patterns, why should you?