🔎
Content-recall testing
Surface verbatim memorization through next-passage prediction and direct probing of source texts.
01 — Overview
When models reproduce protected text, how do you prove it?
Large language models increasingly reproduce copyrighted passages and paraphrase protected sources — raising urgent legal, ethical, and scientific questions about how to evidence such misuse. This tutorial reframes copyright infringement and plagiarism as an evidence-discovery process rather than a binary classification task, and equips attendees with practical, reproducible forensic tools to audit models even under black-box access.
Across three hours we cover two complementary halves. Part I — Copyright Infringement introduces the legal framing and walks through Copyright Detective, an interactive forensic system unifying content-recall testing, paraphrase-level similarity analysis, persuasive-jailbreak probing, and unlearning verification. Part II — Plagiarism Detection builds on “Do Language Models Plagiarize?” to examine verbatim, paraphrase, and idea-level plagiarism in model outputs — and how to measure each. Attendees leave able to run these audits themselves.
What you’ll learn
🧬
Paraphrase-level similarity
Catch leakage that survives rewording using ROUGE, semantic, Jaccard, and MinHash signals.
🔓
Persuasive-jailbreak probing
Stress-test refusal mechanisms with Ethos, Alliance-Building, and Reciprocity strategies.
🧠
Unlearning verification
Probe whether “forgotten” content is truly erased via representational drift analysis.
📑
Plagiarism taxonomy
Distinguish verbatim, paraphrase, and idea-level plagiarism in generated text.
⚖️
Evidence-first auditing
Frame infringement as discoverable evidence and produce defensible, reproducible audit reports.
02 — Schedule
A three-hour, hands-on program
Tentative agenda · all times in Korea Standard Time (UTC+9) · subject to change.
-
1:30Denghui Zhang
Evidencing copyright infringement
From the legal perspective on LLM copyright to evidencing infringement — content-recall testing, paraphrase-level similarity, and persuasive-jailbreak probing — plus copyright mitigation and other open copyright issues.
-
2:15
Break & hands-on setup
-
2:25Guangwei Zhang · Hands-on
Hands-on copyright challenge
Get hands-on with Copyright Detective — collect model responses and audit verbatim and paraphrase-level copyright leakage yourself, live.
-
3:25
Coffee break
-
3:40Dongwon Lee
Plagiarism detection in LLMs
From “Do Language Models Plagiarize?” — measuring verbatim, paraphrase, and idea-level plagiarism in model outputs.
-
4:10
Q&A & wrap-up
Open problems, responsible deployment, and where the research goes next.
-
4:25
End of tutorial
03 — Venue & date
Jeju Island, South Korea
- Conference
- ACM SIGKDD / KDD 2026
- Conference dates
- August 9–13, 2026
- Tutorial session
- Tuesday, August 11, 2026 · 1:30 PM – 4:30 PM
- Venue
- International Convention Center Jeju (ICC Jeju)
Seogwipo-si, Jeju-do, South Korea - Format
- Hands-on tutorial · 3 hours · in person
- Time zone
- Korea Standard Time (UTC+9)
04 — Related materials
Slides, code, demos & the proposal
05 — Speakers
Your tutors
Denghui Zhang
Assistant Professor
Stevens Institute of Technology
Copyright Detective
Guangwei Zhang
Algorithm Engineer
Pine AI
Copyright Detective
Dongwon Lee
Professor
The Pennsylvania State University
Plagiarism & trustworthy AI
06 — References
Read more
-
Zhang, G., Zhu, J., Qian, C., Gong, N., Mihalcea, R., Xu, Z., He, J., Ma, J., Huang, Y., Xiao, C., Li, B., Abbasi, A., Lee, D., Ji, H., & Zhang, D. (2026). Copyright Detective: A Forensic System to Evidence LLMs Flickering Copyright Leakage Risks. arXiv preprint arXiv:2602.05252 [cs.CL].
arxiv.org/abs/2602.05252 ↗ -
Lee, J., Le, T., Chen, J., & Lee, D. (2023). Do Language Models Plagiarize? In Proceedings of the ACM Web Conference 2023 (WWW ’23), pp. 3637–3647.
doi.org/10.1145/3543507.3583199 ↗
Cite the system
@misc{zhang2026copyrightdetective,
title = {Copyright Detective: A Forensic System to Evidence LLMs
Flickering Copyright Leakage Risks},
author = {Guangwei Zhang and Jianing Zhu and Cheng Qian and Neil Gong
and Rada Mihalcea and Zhaozhuo Xu and Jingrui He and Jiaqi Ma
and Chaowei Xiao and Bo Li and Ahmed Abbasi and Dongwon Lee
and Heng Ji and Denghui Zhang},
year = {2026},
eprint = {2602.05252},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2602.05252}
}