ChatGPT Detectors in 2026: Can They Tell Your Writing From AI?

Tesify Team Avatar

·

Quick answer: ChatGPT detectors estimate whether a passage reads like text from ChatGPT or a similar model. They work best on long, unedited model output and worst on the texts students actually submit: edited drafts, mixed human and AI writing, translated passages and formal academic prose.

So can ChatGPT detectors tell your writing from AI? Sometimes, and not reliably enough to be treated as proof. Research has shown that paraphrasing defeats them, that they disagree with one another, and that they flag some human writers far more than others. If you write your own thesis, the useful question is not how to beat ChatGPT detectors but how to make your authorship visible, so that a detector result is never read on its own.

This guide explains how ChatGPT detectors decide, which kinds of text confuse them, what the evidence says, and what you can do as a thesis writer before and after a report lands on your supervisor’s desk. For a side-by-side look at the main tools, see our comparison of paid and no-cost AI detectors.

How ChatGPT detectors decide

ChatGPT detectors do not keep a copy of every answer ChatGPT has ever produced, and they cannot look up your text in OpenAI’s logs. They are statistical models trained to separate two kinds of writing. Broadly, three approaches exist:

  • Trained classifiers: a model learns from labelled examples of human and ChatGPT text and outputs a probability for new text.
  • Zero-shot statistics: the tool measures how predictable each word is to a language model; very predictable text is treated as more likely machine-written.
  • Watermark checks: a model can be built to embed a hidden statistical pattern in its output, which a matching checker can look for. This only works for text from a model that actually applies a watermark.

Each approach makes the same basic assumption: that machine text and human text differ in measurable ways. That assumption holds for raw model output on everyday topics. It weakens quickly once text is edited, translated, mixed or written in a formal register.

Laptop screen with abstract colored bars next to a printed draft covered in handwritten edits
A detector score is a probability, not a record of how a text was written.

Texts that confuse ChatGPT detectors

Type of text Typical detector problem
AI text that has been paraphrased Missed: the statistical signal is weakened or removed
Human text by non-native English writers False flags: simpler, more predictable vocabulary reads as machine-like
Heavily edited or proofread human text False flags: polishing removes the irregularities detectors expect from people
Mixed drafts (some AI-assisted sentences) Unclear: scores reflect the blend, not who wrote what
Formulaic academic sections False flags: methods, definitions and statistical reporting follow templates
Short passages Unstable: too little text for a confident estimate

A thesis contains several of these at once. A methods chapter written carefully by a second-language writer and then proofread is close to a worst case for a ChatGPT detector, even though every sentence is the student’s own.

What the research shows

Researchers at the University of Maryland stress-tested many kinds of detectors, including watermark-based ones, and showed that recursive paraphrasing can significantly reduce detection rates while keeping the text readable. Their conclusion was that reliable detection of AI-generated text cannot be guaranteed in practice when someone is actively trying to avoid it.

On the other side of the error, Liang and colleagues found that GPT detectors misclassified on average 61% of human-written TOEFL essays as AI-generated, a bias against non-native English writers. A European team that tested 12 public tools and two commercial systems found detection tools neither accurate nor reliable, with obfuscation making results much worse.

Even OpenAI, the company behind ChatGPT, withdrew its own AI text classifier in 2023 over its low rate of accuracy. Detectors have improved since, and vendors publish better figures for their current versions, but the core trade-off remains: catching more AI text means wrongly flagging more human text.

Our AI detection accuracy statistics collect the published error rates in one place.

Newer ChatGPT models and the moving target

Every time a new model is released, its writing style shifts, and detectors trained on older output can lag behind. Vendors retrain, but there is always a gap between a new model’s release and a detector’s update, and different vendors close that gap at different speeds. That is one reason two ChatGPT detectors can give the same passage very different scores, and why a score from a tool you used at home may not match the one your university sees.

For thesis writers, the practical lesson is simple: do not treat any single tool’s number as a stable fact about your text. Treat it as one tool’s opinion on one day.

Chapter by chapter: where ChatGPT detectors misfire most

Not every part of a thesis carries the same risk of a misleading score. Knowing where false flags cluster helps you prepare the right evidence and read any report calmly.

  • Abstract. Short, dense and written to a formula, abstracts are a common source of high scores. Keep the drafts that show how you cut a longer summary down to the word limit.
  • Introduction and literature review. Summaries of published work use the cautious, generic phrasing that detectors associate with models. Your reading notes and annotated sources are the evidence here: they show which papers you read and what you took from each.
  • Methodology. Descriptions of sampling, instruments and procedures follow established templates. An ethics application, a data management plan or a pilot-study report written months earlier shows the same content developing over time.
  • Results. Statistical reporting in APA style is close to fill-in-the-blank. Your analysis output files and scripts connect every sentence to your own data.
  • Discussion and conclusion. These are usually the most personal chapters, so they are less often flagged. When they are, your supervisor’s comments on earlier drafts are the strongest record.

The pattern is clear: the more conventional a section, the more a detector struggles, and the more useful your working files become.

Why a lower score is the wrong goal

It is tempting to keep rewriting until a tool reports a low number. That approach has three problems. First, the next tool, or the next version of the same tool, may disagree. Second, contorting clear academic sentences to look less predictable makes your thesis harder to read and can cost marks for clarity. Third, a pattern of last-minute rewrites can itself look suspicious if your work is ever reviewed, because it breaks the steady development your version history would otherwise show.

A better goal is a thesis that is unmistakably yours in substance: built on your question, your data, your interpretation and your field’s debates, with a record of how you got there. That is what examiners assess, and it is the evidence any fair review will look for.

Mixed authorship: the case detectors were not built for

Real student writing is rarely all-human or all-machine. You might have asked a chatbot to explain a statistical test, used a grammar tool on your final draft, or had a sentence suggested by your word processor. Many universities permit some of these uses and ban others. ChatGPT detectors cannot tell which kind of help shaped a passage. They produce a single estimate for the text as it stands.

That is why the most important protection is a clear policy check and an honest record. Before you use any tool, read what your university allows. If a use is permitted, note it. If you are unsure, ask your supervisor in writing. A declaration made at the time is far more credible than an explanation offered after a report.

Tutor and student discussing a printed draft with revision marks at a table
Talking through drafts and revisions shows authorship better than any score.

How to make your authorship visible

  1. Write where history is recorded. Word with cloud saving, Google Docs and Overleaf keep version histories that show a chapter growing over weeks.
  2. Keep the scaffolding. Outlines, reading notes, coding memos, analysis scripts and rough drafts are evidence of the thinking behind the prose.
  3. Use your own data and examples. Passages anchored in your participants, your dataset and your field context are both stronger writing and harder to mistake for generic output.
  4. Keep supervisor feedback threads. Comments and your responses show an evolving argument.
  5. Declare permitted AI use. Follow your graduate school’s format, or write a short factual statement of what tool you used and for what.
  6. Know your thesis cold. Being able to explain every choice in a meeting or defense is the most persuasive evidence of all.

If a ChatGPT detector flags your work

Ask for the report and the name of the tool. Read your institution’s academic integrity procedure so you understand the steps and your rights. Then prepare a short, organized evidence file: version history screenshots or exports, dated drafts, notes, data files and any AI-use declaration. Offer to talk through the flagged passages. If you write in English as an additional language, it is reasonable to mention the published research on detector bias.

Avoid two reactions that make things worse: rewriting the flagged passages before anyone has discussed them, and running the text through a so-called humanizer tool. Both can look like an attempt to hide something. If false flags are a recurring worry, our guide on reducing AI detection false positives on your thesis covers legitimate, transparent steps.

A note on privacy before you test your own text

Many students paste chapters into online ChatGPT detectors to see what happens. Before you do, check what the service does with submitted text. Some tools store submissions, use them to improve their models, or keep them in a database that could later match against your final thesis. Unpublished findings, participant quotes, interview transcripts and confidential partner data should never be uploaded to a service whose terms you have not read. If you want to test a passage, use a short extract that contains no sensitive material, and remember that the result tells you about that tool on that day, nothing more.

Do supervisors rely on ChatGPT detectors?

Practice varies by university and by individual. Many supervisors rely more on what they know about a student’s work: how the draft changed over time, whether the writing matches the student’s voice in meetings and emails, and whether the student can explain the analysis. Some institutions have switched AI detection off entirely; others use it as a screening signal that must be followed by a human review. Few published policies treat a detector score as sufficient evidence by itself.

That is reassuring, but it also places responsibility on you. The relationship with your supervisor, built through regular drafts and honest conversations, is your strongest protection against any misunderstanding.

Where Tesify fits

A thesis built section by section, with a visible structure from the start, makes your process easy to show. Start structuring your thesis in Tesify: it helps you structure your thesis chapter by chapter, and 50,000+ students have written 230,000+ chapters with it. Every word stays 100% written by you.

Frequently asked questions (FAQ)

Can ChatGPT detectors detect edited AI text?

Often not. Research on paraphrasing attacks shows that editing or rewording model output can significantly reduce detection rates, which is one reason a low score proves little.

Can ChatGPT detectors flag text I wrote myself?

Yes. Formal, formulaic or heavily proofread writing, and writing by non-native English speakers, can be flagged as AI-generated even when it is entirely human.

Do different ChatGPT detectors give the same result?

No. Tools use different models and training data, and they update at different times, so scores for the same passage can differ widely.

Is using ChatGPT for grammar checking allowed in a thesis?

It depends on your university’s policy. Some allow limited language support if declared; others do not. Check the policy and ask your supervisor in writing if unsure.

Can OpenAI tell if text came from ChatGPT?

OpenAI withdrew its public AI text classifier in 2023 because of its low accuracy, and there is no public tool that checks text against ChatGPT conversations.

What is the best evidence that I wrote my thesis?

Version history, dated drafts, notes, data files, supervisor feedback threads and your ability to explain every part of the work in your own words.

Write your thesis with AI

Structure, draft, cite, and format your thesis faster with Tesify’s AI writing tools, automatic bibliography, and plagiarism checker. Free to start, no credit card required.

Tesify Team Avatar

Leave a Reply

Your email address will not be published. Required fields are marked *