What is greCAPTCHA?
greCAPTCHA is an empirically-validated and open-source platform to automatically create and evaluate assessments about claimed authors' manuscripts. It supports four workflows depending who uploads the manuscript (the test creator or taker) and who pays for LLM inference costs. It is also highly customizable, supports the printing of question sets, and batch processing of multiple manuscripts.
As a proof of concept, greCAPTCHA co-author Nihar B. Shah, in his capacity as an Editor-in-Chief of the Transactions on Machine Learning Research (TMLR), interviewed the authors of 10 papers submitted to TMLR and which were slated for desk rejection. He reported the results in a post on TMLR. Only 1 of the 10 papers' authors was able to answer substantial questions about their own paper! greCAPTCHA aims to scale these interviews, allowing anyone to reliably assess authors' understanding of their own work.
Why use greCAPTCHA?
Conferences, journals, funders, schools, and universities are struggling with a surge of potentially AI-generated submissions from ostensibly human authors, who may not have exercised sufficient human oversight for their manuscripts. Institutions evaluating submissions can no longer reliably credit expertise based solely on authors' names on submitted work.
greCAPTCHA was empirically validated with 31 academic participants. It measures the capacity to verify: the knowledge and reasoning required to critically assess the claims, methods, and evidence underlying one's specific contributions to a manuscript. It has two dimensions:
Comprehension
Understanding the relevant content and its basis.
Justification
Explaining and evaluating the choices underlying the work.
The name borrows from the Graduate Record Examinations (GRE) used in university admissions, and alludes to CAPTCHAs, which tell human users apart from bots. Here the aim is to distinguish authors who present AI-generated research as their own with no human contribution or oversight from those who contributed to, or at least oversaw, the research.
How does greCAPTCHA work?
- 1
Structure your workflow
Test creators can upload source material, or allow test takers to upload their own. Test creators can pay all API costs, or have test takers pay.
- 2
Create question prompts
Create prompts to generate personalized questions about source material.
- 3
Share the link
Share a public link with your test takers.
- 4
Take the assessment
Test takers upload their source material and provide an API key (if the test creator hasn't already), and take their personalized test.
- 5
Review the report
Test creators and takers both receive a graded report immediately upon test completion.
How was greCAPTCHA evaluated?
In a one-hour in-person session, 31 researchers answered eight questions (15 minutes per paper) about one of their own papers and about an unfamiliar paper chosen by the research team. The unfamiliar paper was either in or outside their field. A semi-structured interview followed.
Scores
High authorship separation. Scores separate own papers from unfamiliar papers with an AUC of 0.90 (1.0 is perfect separation, 0.5 is chance). Excluding the multiple-choice questions raises it to 0.934.
Multiple-choice questions are too easy. The multiple-choice questions barely separated the two (AUC 0.595): its answers can be found by searching the manuscript, so it mostly measures attention.
No statistically detectable field effect. There was no statistically significant difference between in-field and out-of-field unfamiliar papers (p = 0.437), suggesting greCAPTCHA is more than a “field detector”.
Participants' views
Assesses capacity to verify. Most saw it as a valid way to assess capacity to verify, and 12 noted it would be hard to prepare for without having written the paper.
Effective questions target less consequential knowledge. Deep questions about specific details gave the best evidence, more than questions about overall motivation or significance.
Assessment can promote learning and preparation. Questions prompted reflection and new insights; some suggested it as preparation for presentations.
Grading requires careful calibration. The expected level of detail should be clear, and equally correct alternative answers should get credit.
Support for deployment, but not agreement on its role. Most supported deployment; human oversight and ways to contest results were broadly requested.
Frequently asked questions
Who is greCAPTCHA for?
Anyone who needs evidence that people understand work submitted in their name. The report can inform peer review, job screening, admissions, grant proposals, and course projects. For example, a conference or journal can create question prompts and have authors take a test on their own manuscripts, or an instructor can test students on shared course material or on their own project reports.
Who provides the source material?
Both the test creator or taker can. The test creator can upload source material that every test taker is asked about, such as course readings, or allow each test taker to upload their own, such as a manuscript or project report.
What does it cost, and who pays?
Accounts are free. Generating and grading questions calls language models through OpenRouter, so it requires an OpenRouter API key. The test creator can pay all API costs with their own key, or have each test taker provide a key and pay for their own test.
Do I have to write my own questions?
No. You can use the default set of question prompts, or customize them and create your own to suit your needs.
When are results available?
Immediately. Both the test creator and the test taker receive a graded report as soon as the test is completed.
Does a high score prove that the work is correct, or that no AI was used?
No. greCAPTCHA measures capacity to verify: the knowledge and reasoning needed to critically assess one's own contributions. That capacity is a prerequisite for verification, but it does not establish that verification occurred or that the work is correct.
How well does it work?
In a study with 31 researchers, scores separated their own papers from unfamiliar papers with an AUC of 0.90 (0.934 without the multiple-choice questions, which proved too easy).
Bibtex citation
@misc{payan2026grecaptcha,
title = {{greCAPTCHA}: Assessing Understanding as Evidence of Research Authorship Under Generative {AI}},
author = {Payan, Justin and Gyevn{\'a}r, B{\'a}lint and Kasirzadeh, Atoosa and Shah, Nihar B.},
year = {2026},
eprint = {2609.20481},
archivePrefix = {arXiv},
primaryClass = {cs.DL},
url = {https://arxiv.org/abs/2609.20481},
note = {Justin Payan and B{\'a}lint Gyevn{\'a}r contributed equally.}
}