How Do AI Detectors Work? Inside an AI Detection Tool

An AI detection tool estimates how likely a piece of text is to have been written by a language model like ChatGPT, Claude or Gemini. It does this by measuring patterns in the writing, such as how predictable each word is and how much the sentences vary, then comparing those patterns with what it learned from large sets of human and AI text. The result is a probability, not a fact about how the text was made.

That last point matters more than anything else in this guide. Once you understand what a detector is actually measuring, it’s much easier to read its score sensibly.

What does an AI detection tool actually measure?

AI text detection starts from one simple idea. A language model writes by picking the next word from a list of likely options, one word at a time. Even with some randomness added, it tends to choose words that are statistically safe.

People don’t write that way. We pick odd words, change rhythm mid-paragraph, add an aside, or use a phrase that only makes sense in our own context. A detector tries to measure that difference.

What a detector does not do is look up your text in a database of AI answers. That’s how plagiarism checkers work. An AI detector reads the passage itself and asks, “does this look like the kind of text a model would produce?”

What is perplexity?

Perplexity is a measure of how surprised a language model is by a piece of text. If the model can easily guess each next word, perplexity is low. If the text keeps choosing words the model didn’t expect, perplexity is high.

AI-generated text usually has low perplexity because it was made by a model picking likely words. Human writing tends to have higher perplexity, with more unusual word choices and turns of phrase.

You can see this idea in action in GLTR, an early research tool from the MIT-IBM Watson AI Lab and HarvardNLP. It colored each word by how highly GPT-2 ranked it as the next word, so you could literally see whether a passage was mostly “safe” choices. We cover it in our post on the GPT-2 Output Detector and GLTR.

The weakness is obvious once you say it out loud. Simple, formal or formulaic human writing is also predictable. That’s one reason a Stanford study published in Patterns found detectors flagged essays by non-native English writers far more often than essays by native speakers.

What is burstiness?

Burstiness describes how much a text varies. Most often it’s measured as variation in sentence length, though some tools also look at how perplexity changes from sentence to sentence.

Human writers are “bursty.” A long, winding sentence is followed by a short one. Then a fragment. AI text often settles into a steady rhythm where most sentences land in a similar length range and follow a similar shape.

GPTZero popularized perplexity and burstiness as detection signals. Its own explainer now notes that, since autumn 2023, it no longer relies on those two measures for its detection and has moved to a deep-learning model. That shift is typical of the whole field.

How do classifier models detect AI?

Most current AI content detection is done by a classifier: a machine learning model trained on many examples labeled “human” and “AI.” It learns whatever features separate the two groups, which can include perplexity-like signals, word choice, punctuation habits, sentence structure and much more.

An early public example was OpenAI’s GPT-2 Output Detector, a RoBERTa model fine-tuned to tell GPT-2 output apart from human web text. Its model card on Hugging Face reports roughly 95% accuracy on GPT-2 text, but also warns against using it to judge ChatGPT text for misconduct cases.

That warning shows the main limit of classifiers. They’re only as good as their training data. A classifier trained on one family of models can miss text from a newer one. The RAID benchmark (ACL 2024), which tested 12 detectors on over 6 million generations, found that detectors were easily fooled by adversarial attacks, changes in sampling settings and models they hadn’t seen.

There’s also a “zero-shot” family of methods that needs no training at all. DetectGPT, from Stanford researchers, checks whether a passage sits at a spot where a model’s probability scores curve in a telltale way, using only the model’s own word probabilities.

How does AI watermarking work?

Watermarking takes a different approach. Instead of guessing after the fact, the AI company hides a statistical signal in the text while it is being generated.

Google’s SynthID Text is the best-known example. According to Google’s developer documentation, it works by nudging the probabilities of the words a model picks using a secret key. A matching detector can later check whether those nudges are present. Google says the watermark survives light edits and cropping but gets weaker when text is heavily rewritten or translated, and works less well on factual answers where there’s little room to vary wording.

Here’s where SynthID Text stands as of November 2026:

• Gemini text is watermarked. Google DeepMind’s SynthID page says SynthID watermarks text generated by the Gemini app and web experience.
• The text watermark is open source. Developers can add it to their own models through Hugging Face Transformers and detect it with their own keys.
• Public checks focus on other media. Google announced a SynthID Detector portal in May 2025, with text listed as coming later. Its current SynthID page describes public checks for images, video and audio, not text.

The bigger limit is that a watermark only helps for models that use one, and only for whoever holds the detector. Text from a model without a watermark looks like any other text.

How do the main AI text detection methods compare?

Method	What it looks at	Strength	Weak spot
Perplexity	How predictable each word is to a language model	Easy to explain	Flags plain or formal human writing
Burstiness	Variation in sentence length and rhythm	Catches flat, uniform text	Easy to change with light editing
Classifier models	Many learned features from labeled examples	Usually the most accurate	Depends on training data; can miss new models
Zero-shot methods (e.g. DetectGPT)	Probability patterns from a model itself	No training set needed	Works best when you can access the source model
Watermarking (e.g. SynthID Text)	Hidden signal added during generation	Strong evidence when present	Only works for watermarked models and key holders

Why do different AI detectors give different results?

If you paste the same paragraph into three tools and get three different scores, nothing is broken. That’s normal, and there are clear reasons for it:

1. Different training data. Each classifier learned from its own mix of human and AI samples, so each has its own blind spots.
2. Different models covered. A tool trained heavily on one chatbot’s output may be weaker on another’s.
3. Different thresholds. One tool might call 50% “likely AI” while another waits for 80%. Some tools are tuned to avoid false positives and others to catch more AI text.
4. Different units of analysis. Some score the whole document, others score each sentence and then average.
5. Text length. Short passages give a detector very little to work with, so scores swing more.
6. Editing. Mixed text, where a person rewrote parts of an AI draft, sits in a gray zone that tools handle differently.

We go deeper on the research behind this in Are AI detectors accurate?, and compare specific tools in our roundup of the best free AI detectors.

How does our AI detector show sentence-level results?

Our free AI detector is built to show its reasoning, not just a number. When you check a passage, you get:

• An overall AI percentage and a verdict: likely human-written, mixed, or likely AI-generated.
• Sentence highlighting. Each sentence is shaded as “Likely AI,” “Possibly AI” or human, and you can hover over a sentence to see its individual AI likelihood.
• Signals. A short list of the specific observations that drove the score, such as stock phrases or very uniform sentence rhythm.
• Writing stats. Sentence count, average sentence length, a burstiness figure (variation in sentence length) and vocabulary diversity.

Sentence-level results are the useful part. A single score of 45% tells you very little. Seeing that the intro and conclusion are highlighted while the middle reads as human tells you exactly where to look.

Like every detector, ours can be wrong, especially on short, formal or non-native writing. Treat it as one piece of evidence. If you want to work through a full review, our guide on how to check if text is AI generated walks through the steps.

How should you read an AI detection score?

A few habits make any score more useful:

• Read it as a probability. “80% AI” means the tool is fairly confident the text matches AI patterns. It doesn’t mean 80% of the words were written by AI.
• Check length first. Very short text is unreliable on almost every tool.
• Look at which sentences were flagged. Clusters of flagged sentences say more than one overall number.
• Use more than one tool. Agreement between different methods is more meaningful than one high score.
• Add context. Drafts, version history and a conversation with the writer will always tell you more than software.

If your own writing keeps getting flagged because it reads stiff or repetitive, editing for varied rhythm and more specific detail usually helps. Our AI humanizer can help you revise a draft’s flow while keeping your meaning, and our guide on how to humanize AI text covers doing it by hand. Results on any given detector are never guaranteed, though, and students should always follow their school’s AI policy.

Want to see how a detector reads your text? Paste a passage into our free AI detector and look at the sentence highlights, not just the headline score. You can check plan limits on our pricing page.