It measures the writing, not the meaning
This is the part most people get wrong, so it is worth saying plainly. A detector cannot tell whether an argument is smart, whether an essay is honest, or whether a claim is true. It has no idea what your text is about. It only sees the shape of the words on the page.
Think of it like a music teacher who is deaf but can read sheet music. They cannot hear whether the song is beautiful. But they can look at the notes and say “this rhythm is very even and predictable” or “this one jumps around a lot.” An AI detector does the same thing with sentences. It reads the pattern, not the tune.
Everything a detector reports comes from that pattern. There are four things it tends to measure, in rough order of how much they matter: word predictability, sentence variety, phrasing, and formatting. Here is each one.
Signal 1: How predictable each word is
This is the big one. AI models write by choosing the most likely next word, over and over. So AI text is full of words that are easy to guess from the ones before them. A detector checks how surprising each word is. Low surprise looks like a machine.
Detectors have a technical name for this: perplexity. It sounds complicated, but it just means “how surprised is the model by this word?” If every word is the obvious next choice, perplexity is low, and the text reads as AI. If the writing keeps making unexpected choices, perplexity is high, and it reads as human.
Easy to predict, so it reads as AI:
“In conclusion, it is important to consider the many benefits and drawbacks of this approach.”
Harder to predict, so it reads as human:
“Anyway, the whole benefits-and-drawbacks thing kind of misses the point.”
Read those two out loud. In the first, you can almost finish each phrase before you reach it. In the second, “kind of misses the point” is a small swerve you did not see coming. That swerve is what a detector reads as human.
Signal 2: How much the rhythm varies
People mix long and short sentences without thinking about it. One sentence runs on for a while, packing in a couple of ideas and a clause or two, and then the next is three words. AI text tends to keep a steadier beat, with sentences of similar length marching one after another. A detector measures that variety, and low variety pushes the score toward AI.
The technical name here is burstiness. Human writing is bursty. It speeds up and slows down. It has a short punchy line, then a long winding one. AI writing is smoother and flatter, so it has low burstiness. The two words to remember are perplexity and burstiness, and together they carry most of the weight in any score.
Low variety (reads as AI)
“The city has many parks. The parks are popular with families. The families enjoy the open space. The space is well maintained.”
High variety (reads as human)
“The parks are packed on weekends. Families love them. On a good Saturday you can barely find a patch of grass, though the city keeps the place spotless somehow.”
Signal 3: Word choice and phrasing
Some detectors also weigh the words themselves. AI models lean on a recognizable set of transitions and hedges by default. You have probably noticed them: sentences that open with “In conclusion,” or “It is important to note that,” balanced “on one hand, on the other hand” structures, and words like “delve,” “leverage,” “furthermore,” and “moreover.”
No single word proves anything. Plenty of people write “furthermore.” But a page that stacks a dozen of these habits together starts to look machine-made, because that stack is exactly what a model produces when nobody tells it to sound like a person. This signal is softer and less reliable than the first two, so treat it as a nudge, not proof.
Signal 4: Punctuation and formatting
This is the weakest signal, but some tools still glance at it. AI output is often very tidy. Even paragraph lengths. Consistent use of the same punctuation. Neat bulleted lists with parallel structure. Curly quotes and dashes placed the same way every time.
Human drafts are messier. A stray double space, an inconsistent list, a sentence that trails off with a dash. That mess reads as human. It is a small signal and easy to fake in either direction, so no serious tool leans on it much, but it is part of the picture.
How the signals become one score
A detector does not just add these up by hand. It has looked at huge piles of writing, some human and some AI, and learned which mix of these patterns tends to come from a machine. When you paste in new text, it measures the same patterns and asks: does this look more like the AI pile or the human pile?
The answer comes back as a single number, usually a percent. That percent is a confidence, not a fact. It is the tool saying “based on the texture, I am this sure it is AI.” It is a guess built from surface patterns, and guesses can be wrong.
What a detector cannot see
Knowing what a detector measures also tells you what it is blind to. It cannot see:
- → Whether the writing is true, original, or any good.
- → Who actually typed it, or whether a person and a model worked on it together.
- → Whether AI text was edited by a human afterward, which usually erases the smoothness.
- → Intent. It cannot tell honest AI help from dishonest AI use.
All of those matter to a real decision, and none of them show up in the texture of the words. That gap is why a score should start a conversation, never end one.
Why the signals fail, in both directions
Once you know a detector only measures smoothness, its two failure modes make sense.
It flags real people. Plain, careful, formal human writing is smooth and even by nature. A student taught to write in a clean five-paragraph structure, a lawyer writing a contract, or a non-native English speaker using simple sentences all produce low-perplexity, low-burstiness text. The detector sees the texture and calls it AI, even though a person wrote every word. This is a false positive, and it is the most damaging mistake the tool can make.
It misses real AI. The same logic works in reverse. Because the signals are about texture, anyone can change the texture. Adding sentence variety, swapping in a few surprising words, and breaking up the even rhythm all lower the score. A few minutes of editing is usually enough for AI text to slip past. The tool is not judging authorship. It is judging texture, and texture and authorship are not the same thing.
The one thing to take away
An AI detector is a texture meter. It measures how smooth and predictable your writing is, gives that a score, and calls it a guess about authorship. That is genuinely useful as a first filter or a prompt to look closer. It is not proof, and it was never built to be. Read the score for what it is, and you will use it well.