Quick answer: An AI detector is software that estimates how likely it is that a passage was produced by a language model such as ChatGPT. It does not identify an author, and it cannot prove that you did or did not write something.

For a thesis, an AI detector score is best read as a prompt for a conversation, not as a verdict. Independent studies have found that detectors miss obfuscated AI text, flag some human writing, and are especially unreliable on writing by non-native English speakers. If you wrote your own chapters, your strongest protection is not a lower score but a clear record of how the work was produced: notes, drafts, version history and a supervisor who has seen the thesis develop.

This guide explains what an AI detector actually measures, what the published evidence says about its accuracy, how universities tend to use the reports on theses and dissertations, and what to do before you submit so that a detector result never becomes the only evidence in the room.

What an AI detector actually measures

Most AI detectors are classifiers. They are trained on large sets of human-written and machine-generated text and learn statistical differences between the two. When you paste in a chapter, the tool does not compare it with a database of ChatGPT outputs. Instead, it scores features of the text itself: how predictable each next word is, how uniform the sentence lengths and structures are, and how closely the passage resembles the machine-generated examples the model was trained on.

Two ideas come up again and again in detector explanations:

• Predictability (often called perplexity): language models tend to choose likely next words, so text with very predictable word choices looks more machine-like.
• Variation (often called burstiness): human writing usually mixes long and short sentences and shifts rhythm more than model output does.

The problem for thesis writers is obvious once you see the method. Academic prose is deliberately predictable. You use standard phrases, defined terms, a formal register and repeated structures such as “This chapter presents” or “The findings indicate”. A careful methods section can look statistically “smooth” even when every word was typed by you.

What the research says about AI detector accuracy

Detector vendors publish their own accuracy figures, but independent testing gives a more cautious picture. A team led by Debora Weber-Wulff tested 12 publicly available detection tools and two commercial systems used in academia, including Turnitin, and concluded that the available detection tools are neither accurate nor reliable, with a bias towards classifying output as human-written and much weaker performance once text was paraphrased or machine-translated.

A second widely cited study, by Liang and colleagues at Stanford, tested seven GPT detectors on essays written by people. The detectors misclassified on average 61% of TOEFL essays written by non-native English speakers as AI-generated, while essays by US eighth-graders were classified far more accurately. The authors linked the bias to lower lexical variety in non-native writing, which is exactly the feature many detectors treat as a machine signal.

Even the company behind ChatGPT gave up on its own tool. OpenAI withdrew its AI text classifier in July 2023 because of its low rate of accuracy. Some universities have drawn similar conclusions about commercial tools: Vanderbilt University explained in 2023 why it was disabling Turnitin’s AI detector, noting that even a small false-positive rate applied to tens of thousands of papers a year would wrongly flag hundreds of students.

For a longer overview of published error rates, see our roundup of AI detection accuracy statistics.

Why a thesis is a special case for AI detectors

Most published tests use short essays. A thesis or dissertation is different in ways that make a single headline score even less meaningful:

• Length: a 15,000 to 80,000-word document is scored in segments. A handful of flagged paragraphs in a long document can look alarming in a report without saying anything about the rest.
• Formulaic sections: literature summaries, methods descriptions, definitions of terms and statistical reporting follow conventions. APA-style results sentences such as “An independent-samples t-test showed” are close to templates.
• Edited text: theses go through supervisor comments, editing and proofreading. Heavily edited text drifts towards the polished uniformity detectors associate with models.
• Second-language writers: a large share of graduate students write in English as an additional language, the group the Stanford study found to be most at risk of false flags.

None of this means detectors are useless. It means a score on a thesis needs context: which passages were flagged, what kind of writing they are, and what other evidence exists about how the chapter was written.

How universities use AI detector reports

Policies vary widely, so read your own institution’s academic integrity policy and your programme handbook before you submit. In practice you will usually meet one of three approaches:

Across these models, most published guidance treats a detector result as insufficient on its own to prove misconduct. What matters is the overall evidence. That is good news if you wrote your thesis yourself, because you can produce that evidence.

What an AI detector score cannot tell anyone

It helps to be precise about the limits, especially if you ever need to discuss a report:

• A score is a probability estimate from a model, not a measurement of authorship.
• Different detectors routinely give very different scores for the same passage.
• A score cannot distinguish between text you wrote, text you wrote and then edited heavily, and text a tool generated.
• A low score does not prove human authorship either; paraphrased model output can pass.
• The score says nothing about whether the ideas, analysis and data are your own, which is what examiners actually assess.

Before you submit: an authorship checklist

Instead of chasing a number, put yourself in a position where your process speaks for itself. Work through this list for every chapter:

1. Keep version history on. Write in a tool that records versions, such as Word with OneDrive, Google Docs or Overleaf. Do not paste a finished chapter into a new file at the end.
2. Keep your working notes. Reading notes, coding memos, analysis logs and outlines show how ideas developed.
3. Save supervisor feedback and your responses. Comment threads and tracked changes are strong evidence of an evolving draft.
4. Record any AI use your policy allows. If you used a grammar tool or asked a chatbot to explain a statistical test, note what, when and why, and declare it if your university requires a statement.
5. Be ready to talk about every section. You should be able to explain why you chose each source, design decision and analysis. That is what a viva or defense tests.
6. Check your similarity report separately. Similarity and AI detection are different checks; our guide to what counts as a good Turnitin similarity score covers the first one.

Should you run your own thesis through an AI detector?

Many students do, and it is understandable. If you choose to, treat the result as information, not as a target:

• Use it to spot passages that read as generic, then improve them for the reader: add your specific data, your reasoning and concrete examples from your study.
• Do not rewrite clear sentences into awkward ones just to lower a score. Examiners grade clarity, and contorted prose hurts you more than a detector does.
• Never upload unpublished data, participant quotes or confidential material to a tool whose terms allow it to store or reuse submissions. Check the privacy terms first.
• Remember that your university may use a different detector, so your score and theirs can disagree.

Our comparison of paid and no-cost AI detectors covers how the main tools differ, and if you are worried about a false flag specifically, read how to reduce AI detection false positives on your thesis without gaming anything.

If your thesis is flagged by an AI detector

Stay calm and stay factual. A flag is the start of a process, not its end. Typically you should:

1. Ask which passages were flagged and which tool and version produced the report.
2. Read your institution’s procedure so you know your rights, timelines and whether you can bring a representative.
3. Gather your evidence: version history, dated drafts, notes, data files, supervisor emails and any declared AI use.
4. Offer to walk through the flagged sections and explain how you wrote them and why you made each choice.
5. If you are a non-native English speaker, it is reasonable to point to the published research on detector bias, such as the Stanford study above.

Most cases that rest on a detector score alone are resolved once the student can show how the work was produced. The earlier you start keeping that record, the easier this becomes.

AI detectors versus similarity checkers

Students often confuse the two. A similarity checker compares your text against databases of published work, web pages and earlier submissions and shows matching passages with their sources. An AI detector does not look for matches at all; it estimates whether the style of the text resembles model output. A thesis can have a low similarity score and a high AI score, or the reverse. They answer different questions, so read each report on its own terms.

Declaring permitted AI use in a thesis

More universities now ask graduate students to state whether and how they used AI tools. Even where no statement is required, a short, honest declaration removes ambiguity if an AI detector report ever raises questions. A useful declaration answers four things:

• Which tool: the name and, where relevant, the version or date range you used it.
• For what task: for example, checking grammar in the final draft, explaining a statistical concept, or suggesting search terms for a database.
• What you did with the output: whether you accepted, edited or rejected the suggestions, and how you verified anything factual.
• What the tool did not do: for example, that it did not generate text, analysis or references that appear in the thesis.

Keep the statement factual and short, and follow any template your graduate school provides. If you are unsure whether a particular use is allowed, ask your supervisor in writing before you rely on it, and keep the reply. That written answer is worth far more than any detector score if your process is ever questioned.

Where Tesify fits

The best answer to any authorship question is a thesis whose structure and development you can show step by step. Start structuring your thesis in Tesify: it helps you structure your thesis chapter by chapter, and 9,000+ students have written 15,000+ chapters with it. Every word stays 100% written by you.

Frequently asked questions (FAQ)

Can an AI detector prove I used ChatGPT?

No. An AI detector produces a probability estimate based on writing style. It cannot identify the author of a text, and published studies show both false positives and false negatives.

Why was my own writing flagged by an AI detector?

Formal, formulaic or heavily edited academic prose can look statistically similar to model output. Research has also found that detectors flag writing by non-native English speakers more often.

Do universities check theses with AI detectors?

Some do and some have switched detectors off. Check your university’s academic integrity policy and your graduate school’s thesis submission guidance for the rules that apply to you.

What AI detector score is acceptable for a thesis?

There is no universal threshold. Institutions that use detectors generally treat the score as one signal to be read with other evidence, not as a pass mark.

How can I prove I wrote my thesis myself?

Keep version history, dated drafts, reading and analysis notes, data files and supervisor feedback, and be ready to explain every decision in your thesis in your own words.

Is an AI detector the same as a plagiarism checker?

No. A similarity checker looks for matching text in existing sources, while an AI detector estimates whether the writing style resembles machine-generated text.