Is ChatGPT Detectable? What Detectors Actually Catch

Most teachers, journals, and recruiters now run a checker on every long-form submission. The score that comes back is not a verdict, but it is a flag, and a flag is enough to start a difficult conversation. The safest move is to understand how detection actually works before you paste anything into a submission box.

What is ChatGPT detection?

ChatGPT detection is the practice of using statistical or classifier-based software to estimate whether a piece of text was produced by a large language model. The major tools score perplexity (how predictable the next word is) and burstiness (how much sentence-to-sentence rhythm varies) against patterns trained on millions of human and AI samples.

These scores do not prove a passage is AI-written. They produce a probability, often between 0% and 100%, and the threshold for “flagged” is set by the platform. Turnitin, for example, only displays scores above its internal cutoff and shows the raw percentage to the instructor.

Detectors are tuned for ChatGPT specifically because it has the largest install base. GPT-4o and GPT-5 outputs sit inside almost every major training set, which is why those models tend to flag at higher rates than smaller open-source models on the same prompt.

How do AI detectors actually flag ChatGPT?

Detectors look for two main signals.

The first is perplexity. Human writers pick unexpected words and phrasings, while ChatGPT tends to pick the statistically most likely next token. Long stretches of low-perplexity text raise the AI score.

The second is burstiness. People write in uneven bursts: a long sentence, a short one, an aside, a fragment. Models default to a smoother cadence. According to GPTZero’s documentation, both signals are weighted together rather than read in isolation.

Some detectors layer a classifier on top. Copyleaks and Originality.ai train models on labeled human-vs-AI corpora and return a confidence score per paragraph. Grammarly’s checker uses a similar classifier inside its Authorship suite, which the team explains in Grammarly’s blog on ChatGPT detection.

Which detectors catch ChatGPT today?

Six detectors come up in almost every academic, publishing, and recruiting workflow. Each one weighs different signals, which is why scores rarely match across tools.

For deeper testing notes on individual tools, read our breakdowns of GPTZero accuracy and Turnitin AI detection.

Humanize Your Draft Before You Submit

Can Turnitin detect ChatGPT (and ChatGPT 5)?

Yes, Turnitin still flags ChatGPT output, including text generated by GPT-4o and GPT-5. Turnitin’s AI writing indicator scores submitted papers from 0% to 100%, and the company says its model is retrained as new versions of ChatGPT ship.

The catch is the threshold. According to Turnitin’s published guidance, scores below 20% are not displayed because the false-positive risk is too high at that range. So a paper with a real but low AI percentage may show 0% to the instructor, while a long passage of raw GPT output can land in the 70% to 90% range.

GPT-5 specifically tightens word choice and rhythm, which lowers perplexity in some passages and raises it in others. Detectors built around 2023 and 2024 GPT signals have already retrained, but third-party benchmarks vary. If your school uses Turnitin, assume your raw output can be flagged and plan around it.

Does paraphrasing make ChatGPT undetectable?

Light paraphrasing does not erase the signal. According to Compilatio’s analysis, simple synonym swaps and word reorders leave perplexity and burstiness mostly intact, especially over passages longer than 600 words.

Heavy paraphrasing, which means rewriting sentence structure, inserting your own examples, and merging sources, does lower the score, sometimes below the flag threshold. It also moves the work back to you, which is the point. If you have done that much rewriting, the result is closer to your draft than ChatGPT’s.

What does not work: running output through a “spinner” that swaps words at random. Modern detectors are tuned for that pattern and tend to flag it harder than untouched ChatGPT output. Read more on the limits of paste-and-edit shortcuts in our guide on pasting essays into ChatGPT.

Why do ChatGPT detectors get false positives on human writing?

False positives happen when human prose looks statistically smooth. According to GPTZero, ESL students, neurodivergent writers, and anyone trained on formulaic templates such as five-paragraph essays, legal memos, and technical reports tend to produce lower-perplexity text by default.

Stanford researchers have documented this bias in their own evaluations of detection tools, and the result is consistent: human writers can be wrongly flagged, especially on short submissions where the detector has too few tokens to read rhythm.

That is why every responsible detector publishes a “do not act on the score alone” note. Grammarly is explicit on this point: their checker is meant to inform a conversation, not to accuse a writer. For a longer view of how flagging interacts with plagiarism, see our piece on whether ChatGPT plagiarizes.

What should you do before submitting AI-assisted writing?

Treat AI as a research and drafting input, not a finished product. The checklist below works for students, freelancers, and in-house writers.

1. Check the policy. Schools, journals, and clients increasingly publish AI-use rules. Read them first.
2. Save your draft history. Google Docs, Notion, and Microsoft Word all keep version logs that prove you wrote and edited the text.
3. Keep your sources in your own words. Quote directly when you must, paraphrase your own arguments, and note the sources in plain text.
4. Run a screening pass. If your school or platform allows a self-check (Turnitin Draft Coach, GPTZero free tier), use it before submission.
5. Rewrite the AI passages by hand. Replace stock phrasings, add your own examples, and remove anything you cannot defend in a viva or a follow-up call.
6. Disclose if asked. If a teacher or editor asks how you used AI, answer honestly. The penalty for lying is almost always higher than the penalty for honest disclosure.

For a fuller workflow on revising raw GPT output, read our guide on how to make ChatGPT sound human.

Where does Word Spinner fit in this workflow?

Word Spinner is a humanizer, not a detector-evader. It rewrites AI-style sentences for clarity, rhythm, and tone after you have already supplied your own argument and sources. We do not promise detector-proof output, because no honest tool can.

Use Word Spinner the same way you would a copy editor. Paste in your edited draft, take the suggested rewrite, then read it line by line to make sure the meaning and the citations stay intact. If the topic is high-stakes (a graded essay, a peer-reviewed paper, a paid client deliverable), keep your version history and your own annotations alongside the rewritten copy.

You can also pair Word Spinner with our free ChatGPT checker to spot which sentences still read as AI before you submit.

Try Word Spinner Free

Frequently Asked Questions

Can teachers tell if you used ChatGPT?

Sometimes. Teachers spot ChatGPT by tone shifts, generic examples, missing or fabricated citations, and AI checker scores from tools like Turnitin and GPTZero. None of these is conclusive on its own, which is why most schools require a conversation or a draft-history check before any academic action.

Is 40% AI detection bad?

It depends on the tool and the threshold. A 40% Turnitin AI score will display to the instructor and prompt a follow-up question. Some detectors set their flag line at 25% and others at 60%, so the same percentage can mean very different things. Treat any non-zero score as a reason to revise, not a verdict.

Can you get caught for using ChatGPT?

Yes. Detectors flag a percentage of AI usage, but most consequences come from the surrounding evidence: missing draft history, inconsistent style, fabricated sources, or an inability to discuss the paper in person. If you used AI but cannot defend the writing as your own work, the risk goes up sharply.

Does ChatGPT 5 change anything for detectors?

Modestly. GPT-5 produces tighter, more varied prose than earlier versions, which trims some easy detection signals. Turnitin, GPTZero, and Copyleaks have all retrained against newer ChatGPT outputs, so flag rates have not collapsed. Expect detectability to remain a moving target across model upgrades.

Will an “undetectable AI” tool keep me safe?

No. “Undetectable AI” tools rewrite ChatGPT output to evade scoring, and detectors retrain against them quickly. They also shift the academic and ethical risk onto you, since you are the one submitting the work as your own. Use a humanizer for clarity, not as proof of authorship.

Should I tell my professor I used ChatGPT?

If your school’s policy allows AI assistance and asks for disclosure, yes. Honest disclosure usually leads to a lower penalty than concealment. If the policy bans AI entirely, talk to your professor about resubmitting in your own words before a checker flags it.