AI DETECTOR FOR RESEARCH PAPERS · FREE PLAN
AI Detector for Research Papers and Theses
Upload a paper, a thesis chapter, a dissertation draft or a project report, up to 50 pages. See which sections look AI-written, with the figures and metadata checked too.
Open in beta · 10 free credits a month · Files deleted after 30 days
AI Document Detector
Open in betaWhat you need
Upload a PDF, DOCX or PPTX file
PDF, DOCX or PPTX, up to 20 MB and 50 pages.
At a glance
Per page
Text, embedded images and metadata are checked and summarised page by page.
AnyFact method
3 levels
Probably human, uncertain or probably AI — never a bare percentage — with the reasons and the limits of the method.
AnyFact method
10 credits
Free every month with an account, no credit card. Uploaded files are deleted after 30 days.
AnyFact free plan
Why academic writing is a hard case for AI detectors
Academic prose is formal, structured and full of standard phrasing, which is also how generated text reads. Detectors are most likely to misjudge exactly this kind of writing, and non-native authors are the most exposed.
Long documents are also mixed by nature: a methods section written by hand, an abstract polished with an assistant, figures from several sources. One score for 40 pages hides that.
AnyFact gives a result per section or page, checks embedded figures and reads the metadata, so authors and supervisors can see where to look and discuss the process, not just a number.
Does this sound familiar?
- You want to check your thesis chapter before it goes to your supervisor.
- A co-authored paper mixes several writers and tools.
- You supervise reports and want a structured second look, not a verdict on a student.
- Your field writes in a formal register and you worry about false positives.
Here's what we check
Upload the file
PDF, DOCX or PPTX, up to 20 MB and 50 pages.
Text, images, metadata
Each part goes to the right detector; unusual authoring software or dates are flagged.
A verdict in three levels
Probably human, uncertain or probably AI — never a bare percentage — with a short explanation and the limits of the method.
How to read the result
The signals point to a person. It is not a guarantee: a careful generator or heavy editing can leave few traces.
The signals disagree or are too weak. We say so rather than guess, and list what was found.
The signals point to generation. It is an indication, not proof — check the context before acting.
Sample layout. Real reports name the signals that were checked and the limits of the method.
What does an AI detector for research papers check?
An AI detector for research papers reads the manuscript file itself, not just a pasted extract, and reports which sections, pages or figures look generated. AnyFact accepts a paper, thesis chapter, dissertation draft or project report as PDF, Word or PowerPoint, up to 20 MB and 50 pages, and checks the text, the embedded images and the file’s metadata.
| Format | Typical research use | What is read | Result split by |
|---|---|---|---|
| Submitted manuscript, compiled LaTeX paper, exported thesis, scanned appendix | The text layer page by page; scanned pages through OCR in English, French and Arabic (first 20 scanned pages); embedded figures; metadata | Page | |
| Word (.docx) | Thesis chapter, report draft, manuscript before typesetting | Paragraphs grouped under your chapter and section headings; embedded figures; metadata, including the recorded editing time | Heading section, or numbered paragraphs when there are no headings |
| PowerPoint (.pptx) | Defense, viva or conference slides | The text of every shape, grouped shapes and tables; speaker notes; slide images; metadata | Slide |
| Pasted text | An abstract, one section or a cover letter | At least 300 characters of English, French or Arabic, scored per passage of 2–3 sentences | Passage |
A file check uses 3 credits; a pasted-text check through the AI text detector uses 1. Legacy .doc and .ppt files are refused, so save them as .docx, .pptx or PDF first. Of the first 112 documents analysed in the AnyFact beta (October 2026), 50 were PowerPoint files, 31 Word files and 30 PDFs, most of them coursework that students were checking themselves.
How the text of a long document is scored
Text is grouped into blocks of about 1,000 words, and each block is scored with the Fast-DetectGPT method using an open-source model (Qwen2.5). In a Word file with headings, results are reported under your chapter and section titles, so you can see that the literature review reads differently from the methods. When a part holds fewer than 300 words, the result says the text check gives little evidence there.
What the metadata of a paper can show
- An AI tool named in the creator, producer or application field, such as ChatGPT, Copilot, Gemini or Claude.
- A file produced by a programming library or converter (python-docx, python-pptx, PptxGenJS, docx.js, PHPWord…) rather than by an office application.
- A modification date earlier than the creation date, or a creation date in the future.
- 0 minutes of recorded editing time for a long document, or no metadata at all.
Research files often pass through tools other than an office suite, such as typesetting systems, reference managers and submission portals, and converters and “Save as” change metadata too. Each item above is a signal to explain, not proof of anything.
How do you check a research paper or thesis for AI before you submit it?
Check the final file in the form you will submit, in parts if it is long, and record how each flagged part was written. The same steps work for a journal manuscript, a thesis, a dissertation or a project report.
- 01Pick the file you will actually submit: the compiled PDF of a LaTeX paper, the .docx of a thesis chapter, the .pptx of your defense slides.
- 02If it is over 50 pages or 20 MB, split it by chapter or by part (front matter, main chapters, appendices) and check the main chapters first.
- 03For co-authored or embargoed work, agree with your co-authors and check your venue’s rules before uploading the manuscript to any tool.
- 04Upload the file and read the breakdown, section by section or page by page, before the overall verdict.
- 05For each flagged part, note who drafted it, which sources and data it relies on, and which tools touched it, such as language editing, translation or an assistant.
- 06Act on the verdict level, as set out in the table below, and download the PDF report to keep with your research records.
- 07If your venue asks for an AI-use statement, write it from those notes rather than from memory.
What to do with each verdict in a paper or thesis
| Verdict | Typical reading in a long document | What to do |
|---|---|---|
| Probably human | The scored parts lean human. It does not certify the whole document. | Submit as planned and keep your records; another tool may read the text differently. |
| Uncertain | Weak evidence, common on formula-heavy pages, tables, short sections and standard methods text. | Treat it as no information, and make sure drafts and notes for those parts exist. |
| Mixed | Some sections, pages or figures read as generated and others do not. | Map each flagged part to its author and process. Disclose permitted assistance; rewrite any part an assistant drafted where your rules do not allow it. |
| Probably AI | Most of the scored text reads as generated, or the metadata names an AI tool. | If you wrote it, assemble your evidence and talk to your supervisor before submission. If an assistant drafted it, rewrite it yourself or disclose it as your venue requires. |
How do you check a long thesis or dissertation for AI?
Check it in parts of 50 pages or fewer, starting with the chapters that carry your own argument. A free account’s 10 monthly credits cover three document analyses at 3 credits each (Pro’s 1,000 cover 333), so choose the parts that matter most before you start.
| Part | Check it? | Why |
|---|---|---|
| Abstract | Paste it as text if it is at least 300 characters | Short and formulaic; the passage-level text check shows which sentences stand out |
| Introduction and literature review | Yes, first | Long stretches of your own synthesis and argument, where your voice should be clearest |
| Methods | Yes | Standard phrasing can read as predictable; expect uncertain results and keep your protocols |
| Results and discussion | Yes | Your interpretation of your own data, where your analysis notes are strong evidence |
| Conclusion | Yes | Often short; read an uncertain result as no information |
| References and appendices | Usually not | Reference lists, code listings and raw tables are not your prose and make the text check less informative |
| Figures | Yes, with their chapter | Up to 30 embedded images per file go through the photo detector; images under 200 px or thinner than 1:6 are skipped |
Using a Word thesis with headings
When your chapters use Word’s heading styles, the result follows those headings, so you can see that section 2.3 stands out rather than “page 14”. Without headings, results are given by numbered paragraphs. Use the Heading 1 and Heading 2 styles for chapter and section titles; text that is merely bold is not a heading. The AI detector for Word documents page covers .docx files in more detail.
Scanned pages and appendices
Scanned pages, such as a signed declaration or printed survey forms in an appendix, are read with OCR in English, French and Arabic, for the first 20 scanned pages of the file. If scanned material would push the main text over the limits, put it in a separate file.
Using an AI detector for reports
Project reports, internship reports, lab reports and capstone write-ups follow the same logic. They are often built on a department template, so cover pages, standard declarations and fixed headings read as predictable and tend to come back uncertain. Judge the result on the parts you wrote, and keep the blank template to show what was given.
Why can an honest research paper be flagged as AI-generated?
Because good academic prose is regular, and regular wording is what statistical detectors read as generated. These features of research writing push a score toward AI, and each has a record you can keep:
| Feature | Why it reads as predictable | What to keep |
|---|---|---|
| Standard methods language | Protocols are described in fixed phrasing so others can reproduce them | Lab notebooks, protocols and the papers you adapted the methods from |
| Literature review synthesis | Summaries of other work reuse the reporting verbs and hedges of the field | Reading notes and annotated papers, with dates |
| Abstracts and structured summaries | Short, dense and built on a fixed structure | Earlier versions and the comments that shaped them |
| Second-language writing | Careful authors choose common, safe constructions | Drafts in your first language, or from before language editing |
| Language editing and grammar tools | Heavier rewriting replaces your wording with typical wording | The pre-editing version and the changes made |
| Machine translation | Fluent, regular output even when the research is entirely yours | The source-language draft |
| Several co-authors | Sections drafted by different people, sometimes with different tools, read unevenly | A record of who drafted which section |
The second-language effect has been measured. Liang et al. (Stanford), writing in Patterns in 2023, showed that GPT detectors in wide use judged English essays by non-native writers to be machine-generated at a much higher rate than essays by native speakers. A great deal of research is written in English by people for whom it is a second language, so a flag on such a paper calls for particular care.
The guide on whether AI detectors are accurate explains both kinds of error and how to weigh them.
How can you show your research and writing process if you are questioned?
With the records research already produces, kept in order. A thesis or paper leaves a far longer trail than a class essay, and that trail is the strongest answer to any question about authorship.
Records worth keeping from day one
- Version history of your chapters: Google Docs versions, or Word’s version history for files saved to OneDrive or SharePoint. If you write in LaTeX, a Git repository or your editor’s history does the same job.
- Data and analysis notebooks or scripts, with dated runs and outputs that match your figures and tables.
- Lab or field notebooks, protocols and instrument logs.
- Notes from supervisor meetings and the comments returned on each draft.
- Your reference manager library, which records when you added each source, and your reading notes.
- Emails with co-authors about who drafts which section.
Disclose AI assistance where your venue requires it
Many journals and universities ask authors to state how they used AI tools. Two widely followed references set the tone:
- COPE (Committee on Publication Ethics), in its position statement “Authorship and AI tools”, first published in February 2023: AI tools cannot be authors because they cannot take responsibility for the work, authors who use them should say which tool and how in the methods or a similar section, and authors stay responsible for every part of the manuscript.
- The ICMJE Recommendations (International Committee of Medical Journal Editors, updated January 2026): journals should ask authors at submission whether they used AI-assisted technologies and how, in the cover letter and the relevant section of the work, and chatbots should not be listed as authors.
Your university’s thesis regulations and your target journal’s author guidelines decide what applies to you, including whether language editing must be declared. Read them before writing your statement, and keep a copy of the version you relied on.
If your supervisor or an editor raises a concern
- 01Ask what exactly prompted the concern and which sections or figures it covers.
- 02Share a dated timeline: proposal, data collection, analysis, chapter drafts, feedback rounds.
- 03Show the work behind a flagged results section: the notebook, the script, the output.
- 04Explain any tools you used and point to the disclosure you made or the rule you followed.
- 05Ask how the review procedure of your institution or journal works, and keep the exchange in writing.
How should supervisors and reviewers use an AI detector on a thesis or manuscript?
As a structured second look that tells you where to ask questions, never as a finding. A section-by-section result turns a vague impression into specific questions about specific parts.
A structured second look for supervisors
- 01Agree the rules with your students at the start: which AI uses are allowed in the thesis and how they are declared.
- 02Run a check when a draft gives you a reason to, for example a sharp change from earlier chapters, and tell the student you did.
- 03Read the breakdown: which sections, which figures, which metadata signals.
- 04Discuss the flagged parts with the student, using their drafts, data and notes. Ask them to talk you through a flagged argument or analysis.
- 05Record what you saw and what the student showed, and decide on the whole picture. A detector result alone is never enough.
Confidentiality before you upload
An unpublished manuscript, a thesis under examination or a co-authored draft is not only yours to share. Check your venue’s or institution’s rules before you upload it to any third-party tool, AnyFact included.
- ICMJE says reviewers should follow the journal’s AI policy or ask permission before using AI in a review, and that confidentiality may rule out uploading a manuscript to AI software unless the journal explicitly permits it.
- Thesis examiners should check their university’s examination rules, and ask the graduate school when those rules say nothing about third-party tools.
- If you may upload, remove names and identifying details where you can.
- Know where the file goes: AnyFact processes it with Modal (compute), Supabase (storage) and Anthropic’s Claude (second opinion and explanation), and deletes uploaded files after 30 days.
For the editorial side, including questions to send authors and how to read a mixed result in a submission, see the page on AI detection in academic publishing.
How do you check thesis defense or conference slides for AI?
Upload the .pptx itself: the check reads the text of every slide, including tables and grouped shapes, the speaker notes and the images, and gives a result slide by slide. Defense decks are short on words, so the images, the notes and the metadata often say more than the bullets.
- Figures copied from your thesis go through the photo detector like any other image. A generated illustration flags its slide even when the text is yours.
- Speaker notes written as a full script give the text check something to measure; notes pasted from an assistant are analysed like any other text.
- Decks exported by AI slide tools such as Gamma, Tome, Beautiful.ai or Plus AI often name the tool in their metadata. If your institution allows the tool, say so in your acknowledgements or to your committee.
Slide-specific limits and signals are covered on the AI detector for PowerPoint page.
What can an AI checker for a thesis or paper not tell you?
It tells you how the file reads to one detection method. It cannot decide authorship, integrity or compliance with your venue’s rules.
- It is not proof of misconduct, and a probably-human result is not a certificate either.
- It does not predict Turnitin or any tool your university or publisher runs. Those are separate systems, and AnyFact does not replicate them.
- It is not a plagiarism or similarity check; see the difference between an AI detector and a plagiarism checker.
- It does not verify references, data or statistics. Fabricated citations or wrong numbers can sit in entirely human text.
- An image check says whether a figure looks generated, not whether it faithfully shows your data.
- It does not know your venue’s rules, so it cannot say whether the help you used was allowed or properly declared.
Read the result as a map of where to look and what to document, then let people make the decision with your records in front of them.
Before you check
Is the result proof that a content was made by AI?▼
No. It is an indication, not proof. Detectors can be wrong in both directions, especially on compressed, cropped, edited or very short content. Use the result to decide what to check next, not as a final judgement on anyone.
Why no percentage?▼
A single percentage suggests a precision no detector has. You get one of three levels — probably human, uncertain or probably AI — with the signals behind it and the limits of the method.
How does the analysis work?▼
When the content carries provenance data, such as C2PA content credentials or metadata, we read it first. Then an open-source detector scores the content. When the score is borderline, Claude reviews it as a second opinion, and the result is explained in plain words.
What happens to my files?▼
They are processed to run your analysis — including, for borderline cases, a review by Claude, Anthropic’s model — and deleted after 30 days.
What does it cost?▼
There is a free plan: a free account gets 10 credits every month, with no credit card. Pro ($15 a month) gives 1,000 credits a month; Business ($49 a month) gives unlimited Business Checks and AI checks under a fair-use limit.
Can I check a LaTeX paper for AI?▼
Yes. Compile it and upload the PDF: the text layer is read page by page, along with embedded figures and metadata. Pages made mostly of equations, tables or code give the text check little prose to work with.
How do I check a thesis longer than 50 pages?▼
Split it into parts of 50 pages and 20 MB or less, for example by chapter, and check each part as its own analysis at 3 credits. Start with the chapters that carry your own argument; references and appendices can usually wait.
Should I include the reference list and appendices?▼
Usually not. Reference lists, raw tables and code listings are not your prose and make the text check less informative. Check them only if they contain sections you wrote, or figures whose origin you need to confirm.
Do I have to disclose that I used AI for language editing?▼
It depends on your venue. COPE and the ICMJE Recommendations ask authors to disclose how they used AI tools, and individual journals and universities decide whether light language editing counts. Read your journal’s author guidelines or your thesis regulations, and declare when in doubt.
Can a peer reviewer upload a manuscript to an AI detector?▼
Only if the journal allows it. ICMJE says reviewers should follow the journal’s AI policy or ask permission, and that confidentiality may rule out uploading a manuscript to AI software unless the journal explicitly permits it. That applies to AnyFact as much as to any other tool.
Are the figures and charts in my paper checked?▼
Embedded images are, up to 30 per file, through the photo detector. Images under 200 px or thinner than 1:6, such as icons, logos and rules, are skipped. A generated figure in a human-written paper gives a mixed result that names the page or section.
Will the result match my university’s or journal’s checker?▼
Not necessarily. Institutional tools such as Turnitin are separate systems with their own models, and AnyFact does not replicate or predict them. Use the result to understand your document and prepare your records, not to forecast another tool’s report.
Can I check a paper written in a language other than English?▼
French and Arabic are supported alongside English, for both the text check and OCR on scanned pages, and the explanation is written in English, French or Arabic. Papers in other languages are outside what the text detector is built for.
Check a paper section by section
What you need
Upload a PDF, DOCX or PPTX file
Free plan — 10 credits every month, no credit card.
Each result is an indication, not proof.
Files deleted after 30 days.
You might also want to check
AI PDF Checker: Was This PDF Written With AI?
Upload a PDF and see, page by page, which parts look AI-written: text, scanned pages (OCR), images and metadata, with an explained verdict. Free plan, no card.
AI Checker for Students: Check Your Assignment Before You Submit
Check your essay, report or slides for AI before you submit: upload the file or paste the text and see which parts look AI-written, and why.
Was This Document Made With AI?
Check the text, images and metadata of a PDF, DOCX or PPTX file, page by page, with a verdict explained in plain words. Free plan.