GPT detection tools estimate whether text was written by a large language model such as ChatGPT by measuring how predictable the wording is and comparing it with patterns learned from large sets of human and AI writing. They can be useful signals, but they're probabilistic: they make mistakes in both directions, and their reliability drops on short, edited or formulaic text.
ZeroTell publishes this blog and makes an AI detector, so we have firsthand reasons to be honest about what detection can and can't do.
Why GPT text is detectable at all
Language models write by predicting likely next words. That makes their output statistically smoother than most human writing. People make odd word choices, change rhythm, include specific details, and wander a bit. Models tend to choose safe, high-probability words and keep a steady pace.
Detectors exploit that difference. They don't "recognize" ChatGPT the way you'd recognize a friend's handwriting. They measure how much the text looks like what a model would produce.
The main detection methods
Perplexity and burstiness
Two ideas came up early in public discussions of AI detection:
- Perplexity measures how surprised a language model is by the text. Low perplexity means the words were very predictable, which is more typical of AI.
- Burstiness describes variation across sentences. Human writing tends to mix long and short, simple and complex. AI writing is often more uniform.
These are intuitive, but on their own they're blunt. A plain, clear human writer can have low perplexity too.
Trained classifiers
Most modern detectors use classifiers: models trained on large collections of labeled human and AI text. The classifier learns many subtle features at once, including word choice, structure, transitions and rhythm, and outputs a probability.
Classifiers are generally stronger than single metrics, but they depend heavily on their training data. If they haven't seen much writing from a certain field, language background, or newer AI model, accuracy drops.
Sentence-level scoring
Many tools score each sentence or segment, then combine those into a document score. That's why you see highlights. It also explains why mixed documents, part human and part AI, get mid-range scores that are hard to interpret.
Watermarking
A watermark is a hidden statistical pattern a model deliberately builds into its output, detectable by someone with the key. It's a promising idea because it doesn't rely on guessing. OpenAI has publicly said it researched text watermarking, and other companies have explored similar ideas, but you can't currently run ChatGPT text through a public watermark checker. Watermarks also weaken when text is heavily edited or translated.
| Method | How it works | Strength | Weakness |
|---|---|---|---|
| Perplexity | Measures predictability | Simple, explainable | Flags plain human writing |
| Burstiness | Measures variation | Captures rhythm | Easy to distort with editing |
| Classifier | Learns patterns from examples | Most accurate in practice | Depends on training data |
| Watermark | Detects a pattern built in by the model maker | Strong if available | Not broadly available for checking |
A cautionary history
In early 2023, OpenAI released its own AI text classifier. Within months it withdrew the tool, saying it wasn't accurate enough. That's a useful reminder: even the company that built ChatGPT found detection hard. Today's detectors are more refined, but the underlying challenge, separating overlapping styles of writing, hasn't disappeared.
What a detection score means
When a tool says "78% AI," people often read that as "78% of this was written by AI" or "78% chance this person cheated." Neither is quite right. Depending on the tool, the number usually means one of:
- the model's confidence that the text is AI-generated, or
- the share of sentences it classifies as likely AI.
Either way, it's an estimate from a pattern-matching model. It doesn't know who wrote the text, what tools they used, or what your course allows. For more on reading scores, see what an AI score means.
Where GPT detection struggles
Being clear about weak spots helps everyone use detectors fairly.
- Short text. A few sentences don't give the model enough signal.
- Formulaic genres. Lab reports, abstracts, legal summaries and cover letters follow templates that look predictable.
- Second-language writers. Simpler vocabulary and careful grammar can resemble AI patterns. Educators and researchers have raised fairness concerns about this.
- Edited or mixed text. When a person revises AI text or adds AI paragraphs to their own, scores land in a murky middle.
- New models. Detectors trained on older AI writing may lag behind newer models.
- Translation. Text written in one language and machine-translated can read as AI-like.
None of this means detectors are useless. It means they're one input among several.
Two sentences, two readings
Here's a small illustration of why predictability matters. Compare:
"Regular exercise offers numerous benefits for both physical and mental health."
"I started running in March mostly to get out of the apartment, and by June I noticed I was sleeping through the night again."
The first sentence is exactly what you'd expect to follow the topic "exercise." Every word is the likely choice. A detector would find it very predictable, and in fact many humans write sentences just like it. The second is full of choices a model wouldn't be likely to make on its own: a month, an apartment, a specific side effect.
That's also why a single sentence is never enough to judge. Plenty of honest essays contain lines like the first one. Detectors look for a pattern sustained across many sentences, and even then, they're estimating.
How to use GPT detection responsibly
If you're a teacher
- Use detection as a prompt to look closer, not as a verdict.
- Compare with the student's other work and in-class writing.
- Ask about process. Drafts, notes and version history tell you more than any score.
- Know the tool's limits for short texts and non-native writers.
- State your AI policy clearly so students know what's allowed.
If you're a student
- Know your course's AI rules.
- Write in tools with version history so you can show your process.
- Check your draft if you're worried, and treat flagged sentences as feedback on style, such as overly generic phrasing.
- Don't try to game scores. Running text through tools to hide AI use is treated as misconduct in many policies.
If you're an editor or publisher
- Combine detection with fact-checking, since AI text often contains invented details.
- Set clear contributor guidelines on AI use and disclosure.
- Look for substance: original reporting, specific expertise, real examples.
Detection versus good writing
There's a side effect worth mentioning. Text that detectors flag is often text that human readers find bland: generic claims, predictable transitions, no specifics. When writers respond by adding real examples, clear opinions, and varied sentences, the writing gets better whether or not anyone runs a detector.
That's the honest role of a humanizer as well: improving readability and voice in your own work, within the rules you write under. It isn't a way to make someone else's work yours.
Key takeaways
- GPT detection estimates how much text resembles language model output, using predictability measures and trained classifiers.
- Watermarking could help in the future, but you can't currently check ChatGPT text for one.
- Scores are probabilities, not proof, and they're weaker on short, formulaic, edited or second-language writing.
- Teachers should pair scores with process evidence; students should keep drafts and follow their course's AI policy.
- Writing with specifics and a real point of view is both better writing and less likely to be misread.
To see detection in action, run a paragraph you wrote and a paragraph an AI wrote through the same checker, and compare which sentences get highlighted and why.