LLM-generated Code and Open Source License Compliance: How Big Is the Problem?
summary
Using a 35× larger reference set than Xu et al. — osskb.org (≈27 TB) queried via scanoss.py — the study finds LLM outputs match existing open‑source code at rates noticeably above the prior ~0.3–3% benchmarks, including near‑verbatim reproductions (example: a C++ Winnowing implementation). The work uses the Winnowing algorithm as a fast preliminary indicator, introduces a WM density metric that localizes snippet matches without needing full original files, and recommends verifying provenance, adding attribution, and performing license‑compatibility checks when matches appear.
AI-written summary · about
chapters (10)
- 0:00LLM code license risks
- 2:33Why research matters now
- 5:47Plagiarism risks in AI code
- 9:48Large-scale experiment design
- 13:25Winnowing algorithm for matches
- 15:35Filtering false positives
- 18:45Results and prior study
- 21:13Next steps and questions
- 23:44Legal and ethical implications
- 28:27Q&A
related talks
33mPractical Evaluation of LLMs for Code
42mOpen Source LLMs — Powering AI Accessibility and Innovation
16mDeploying LLM on prem: let's review licenses and highlight open source technologies
01h 49mAI engineering with open access LLMs that lie, curse, and steal
49mLLM—Search Engine Killer?
35mOpen-source AI: why it matters and how to get started
40mLLM for Coding, the State and lnitiative
40mInstructLab: Applying Open Source Methods to Building and Trai...