analiza aiqbench
This preprint uses a membership inference attack on 34 licensed but pay-walled O'Reilly books to test whether OpenAI models appear to recognize content they were not licensed to train on. The GPT-4o result (AUROC 0.82, CI 0.60-0.96) is suggestive but rests on a small sample, a wide confidence interval, and models that differ in more than training data. Readers should treat it as early evidence that model provenance and licensing questions may be empirically testable, not as proof of unauthorized access.
Beyond Public Access in LLM Pre-Training Data
arXiv:2505.00020v3 Announce Type: replace-cross Abstract: Using a legally obtained dataset of 34 copyrighted O'Reilly Media books (13,962 paragraphs), we apply the DE-COP membership inference attack to test whether OpenAI models show recognition of non-public (pay-walled) content they were not licensed to train on. Our main contribution is examining non-public books, which allows us to consider potential access violations and not just training inclusion. Our results suggest that GPT-4o, OpenAI's more recent and capable model in our sample, exhibits patterns consistent with recognition of non-public books, with an AUROC score of 0.82 (95% bootstrapped CI: 0.60-0.96) -- where above 50% is better than random chance. GPT-3.5 Turbo (trained two years prior) and the smaller GPT-4o Mini both score near chance recognition on non-public data (0.54 and 0.56, respectively). Results are robust to independent paraphrasers (Gemma and Qwen) and leave-one-book-out sensitivity analyses. We also test both GPT-4o and GPT-4o mini, with the same October 2023 training cutoff date, to ensure the model is not simply identifying language shifts over time. These findings support disclosure requirements for pre-training data provenance and formal licensing markets for training content. Results remain preliminary given the small book sample, possible contamination of post-cutoff controls, and differences across models in size, architecture, and training composition.
Abstract:Using a legally obtained dataset of 34 copyrighted O'Reilly Media books (13,962 paragraphs), we apply the DE-COP membership inference attack to test whether OpenAI models show recognition of non-public (pay-walled) content they were not licensed to train on. Our main contribution is examining non-public books, which allows us to consider potential access violations and not just training inclusion. Our results suggest that GPT-4o, OpenAI's more recent and capable model in our sample, exhibits patterns consistent with recognition of non-public books, with an AUROC score of 0.82 (95% bootstrapped CI: 0.60-0.96) -- where above 50% is better than random chance. GPT-3.5 Turbo (trained two years prior) and the smaller GPT-4o Mini both score near chance recognition on non-public data (0.54 and 0.56, respectively). Results are robust to independent paraphrasers (Gemma and Qwen) and leave-one-book-out sensitivity analyses. We also test both GPT-4o and GPT-4o mini, with the same October 2023 training cutoff date, to ensure the model is not simply identifying language shifts over time. These findings support disclosure requirements for pre-training data provenance and formal licensing markets for training content. Results remain preliminary given the small book sample, possible contamination of post-cutoff controls, and differences across models in size, architecture, and training composition.
38 pages, 6 figures, 12 tables. Updated to reflect the published version in AI and Ethics
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2505.00020 [cs.CL]
(or arXiv:2505.00020v3 [cs.CL] for this version)
https://doi.org/10.1007/s43681-026-01438-z DOI(s) linking to related resources
From: Ilan Strauss Dr. [view email] [v1] Thu, 24 Apr 2025 15:49:59 UTC (499 KB) [v2] Wed, 6 May 2026 17:41:54 UTC (237 KB) [v3] Fri, 30 Oct 2026 16:20:37 UTC (461 KB)
Text integral sub licența deschisă a publicației · 272 cuvinte
Subiecte conexe
Adapting Linear-Time Architectures for Tabular In-Context Learning
arXiv · 30 sept. 2026
Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models
arXiv · 21 oct. 2026
Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection
arXiv · 3 nov. 2026
Does Pretraining Teach Chemical Language Models Chemistry?
arXiv · 14 oct. 2026