An AI text detector is a tool that estimates how likely it is that a piece of writing was produced by a large language model rather than a person. That is the whole job, stated plainly. A detector does not "know" who wrote something the way a witness knows what they saw. It reads the text, measures a handful of statistical properties, compares those properties to what human and machine writing usually look like, and returns a probability or a label.
If you teach, administer, or parent in 2026, you have almost certainly seen one of these tools attached to a grade, a flag, or an awkward conversation. It is worth understanding what is actually happening under the hood, because the gap between "the tool said 92 percent AI" and "a student used AI" is wider than most dashboards admit. This piece walks through the core techniques, the rise of model-specific detectors, and the honest limits of the whole category.
The core idea: machine writing is statistically smoother
Human writing is bumpy. We reach for an odd word, double back, over-explain one point and under-explain the next. Large language models, by design, do something different. At every step they predict the most probable next token given everything before it. Even when you crank up the randomness, the output tends to hug high-probability paths. The result is text that is fluent, evenly paced, and statistically smoother than most human drafts.
Almost every detection technique is a way of measuring that smoothness. Two terms come up constantly:
Perplexity is a measure of how "surprised" a language model is by a piece of text. Low perplexity means the words are highly predictable; high perplexity means they are not. Machine-generated text usually has lower perplexity because it was literally optimized to be predictable.
Burstiness measures variation in that predictability across sentences. Humans write a long winding sentence, then a short one. Then a fragment. Models tend to keep a more uniform rhythm. Low burstiness is a soft signal of machine authorship.
A detector does not see "AI." It sees unusually low perplexity and unusually flat burstiness, and it translates that into a score. Keep that translation in mind, because it explains both why detectors work at all and why they misfire.
The main families of detection techniques
There are four broad approaches in use today, and most commercial tools blend several of them.
Statistical and feature-based methods. These are the oldest and most transparent. The tool extracts measurable features such as perplexity, burstiness, sentence-length variance, vocabulary richness, and punctuation patterns, then sets thresholds. They are fast and explainable, but they can be defeated by light editing and they struggle with short passages where there is not enough text to measure.
Trained classifier models. Here a separate machine learning model is trained on large collections of human and AI text and learns to tell them apart. This is the dominant approach in modern detectors. It catches patterns a human would never articulate, but it is only as good as its training data. Feed it writing from a newer model it never saw, or from a student whose natural style happens to be very clean, and accuracy drops.
Watermarking. Some AI developers embed a hidden statistical signal in generated text by subtly biasing which words the model picks. A matching detector can later read that signal. Watermarking is the most reliable technique in principle because it does not guess, it verifies. The catch is that it only works if the generator cooperates, the watermark survives editing, and the detector has the key. Most text students paste into an assignment carries no watermark at all.
Retrieval and similarity checks. This is closer to classic plagiarism detection. Instead of asking "does this look like AI," the system asks "have I seen this exact output before." It compares submitted text against a database of known AI generations or prior submissions. Useful as a supporting signal, useless against freshly generated, never-before-seen text.
Model-specific detectors: the Gemini detector question
A newer wrinkle is the rise of detectors marketed for a single model. You will see tools advertised as a "Gemini detector," a "ChatGPT detector," or a "Claude detector," each claiming to spot the fingerprint of one specific system.
The logic is real. Different models do leave slightly different statistical traces. They favor certain phrasings, certain transition words, certain sentence shapes. A classifier tuned narrowly on Gemini output can, in controlled tests, separate Gemini text from human text more sharply than a generalist tool.
But the marketing oversells it in two ways. First, the fingerprints are not stable. Every model update reshuffles the patterns, so a Gemini detector trained on last quarter's version may quietly degrade against this quarter's. Second, the moment a student edits the output, mixes models, or runs it through a paraphraser, the model-specific signal blurs into the general "this looks machine-smooth" signal that any detector is already using. A model-specific detector is best understood as a generalist detector with a narrower training set, not a forensic match. Treat the model name in the product title as a hint about what it was tuned on, not a guarantee about what it can prove.
Worked example: the same paragraph, two ways
Imagine a student submits a paragraph about photosynthesis. Version A was written quickly by a tired teenager: it has one run-on sentence, a misused semicolon, an oddly specific tangent about a houseplant, and a flat ending. Version B was generated by a model: balanced sentence lengths, tidy transitions ("Furthermore," "In conclusion"), no tangents, no friction.
A perplexity-and-burstiness detector will likely flag Version B and clear Version A, because Version A is bumpy and Version B is smooth. So far so good. Now imagine the student spends five minutes editing Version B, adding a contraction, breaking one sentence in half, and inserting a personal aside. The smoothness drops, the burstiness rises, and the score slides toward "human." Nothing about the authorship changed. Only the measurable surface did. This is the central fragility of the entire field, and no scoring dashboard fully escapes it.
Common misconceptions
"A high score is proof." It is not. A score is a probability estimate based on surface statistics, not evidence of an act. Detectors produce false positives, and those false positives fall hardest on writers whose natural style is clean and simple: multilingual students, younger writers, and anyone trained to write in a plain, structured way.
"Detectors read meaning." They do not understand the argument or check the facts. They measure form. A brilliant, original essay can score as AI if its prose happens to be smooth, and a shallow human-written one can pass.
"You can run it once and decide." Detector outputs vary with length, edit, and model version. The shorter the passage, the less reliable the result. A single number on a single run is the weakest possible basis for an accusation.
"A model-specific detector confirms which AI was used." As covered above, it suggests, it does not confirm. Model fingerprints drift and smear under editing.
Quick FAQ
Are AI detectors accurate? They are better than chance and useful as a screening signal, but they are not reliable enough to stand alone as proof. Accuracy depends heavily on text length, how much the text was edited, and whether the detector has seen the underlying model before.
Can AI detection be fooled? Yes, routinely. Paraphrasing tools, manual editing, and prompting a model to "write less formally" all degrade detection. This is an ongoing arms race, not a solved problem.
What should a teacher do with a flag? Treat it as a prompt to look closer, not a verdict. Compare the work to the student's known writing, look at draft history and revision trails, and have a conversation. The detector starts the inquiry; it should never end it.
Why do detectors disagree with each other? Because they use different techniques, thresholds, and training data. Two tools measuring different features will reasonably reach different scores on the same paragraph.
The honest summary is this: AI detection tools are useful instruments for measuring the statistical texture of writing, and they are genuinely improving. But they measure surface, not authorship, and the surface is easy to change. Use them as a flashlight, not a gavel.