A new study from the AI Disclosures Project has raised questions about the data OpenAI uses to train its large language models (LLMs). The research indicates the GPT-4o model from OpenAI demonstrates a “strong recognition” of paywalled and copyrighted data from O’Reilly Media books.
The AI Disclosures Project, led by technologist Tim O’Reilly and economist Ilan Strauss, aims to address the potentially harmful societal impacts of AI’s commercialisation by advocating for improved corporate and technological transparency. The project’s working paper highlights the lack of disclosure in AI, drawing parallels with financial disclosure standards and their role in fostering robust securities markets.
The study used a legally-obtained dataset of 34 copyrighted O’Reilly Media books to investigate whether LLMs from OpenAI were trained on copyrighted data without consent. The researchers applied the DE-COP membership inference attack method to determine if the models could differentiate between human-authored O’Reilly texts and paraphrased LLM versions.
Key findings from the report include:
- GPT-4o shows “strong recognition” of paywalled O’Reilly book content, with an AUROC score of 82%. In contrast, OpenAI’s earlier model, GPT-3.5 Turbo, does not show the same level of recognition (AUROC score just above 50%)
- GPT-4o exhibits stronger recognition of non-public O’Reilly book content compared to publicly accessible samples (82% vs 64% AUROC scores respectively)
- GPT-3.5 Turbo shows greater relative recognition of publicly accessible O’Reilly book samples than non-public ones (64% vs 54% AUROC scores)
- GPT-4o Mini, a smaller model, showed no knowledge of public or non-public O’Reilly Media content when tested (AUROC approximately 50%)
The researchers suggest that access violations may have occurred via the LibGen database, as all of the O’Reilly books tested were found there. They also acknowledge that newer LLMs have an improved ability to distinguish between human-authored and machine-generated language, which does not reduce the method’s ability to classify data.
The study highlights the potential for “temporal bias” in the results, due to language changes over time. To account for this, the researchers tested two models (GPT-4o and GPT-4o Mini) trained on data from the same period.
The report notes that while the evidence is specific to OpenAI and O’Reilly Media books, it likely reflects a systemic issue around the use of copyrighted data. It argues that uncompensated training data usage could lead to a decline in the internet’s content quality and diversity, as revenue streams for professional content creation diminish.
The AI Disclosures Project emphasises the need for stronger accountability in AI companies’ model pre-training processes. They suggest that liability provisions that incentivise improved corporate transparency in disclosing data provenance may be an important step towards facilitating commercial markets for training data licensing and remuneration.
The EU AI Act’s disclosure requirements could help trigger a positive disclosure-standards cycle if properly specified and enforced. Ensuring that IP holders know when their work has been used in model training is seen as a crucial step towards establishing AI markets for content creator data.
Despite evidence that AI companies may be obtaining data illegally for model training, a market is emerging in which AI model developers pay for content through licensing deals. Companies like Defined.ai facilitate the purchasing of training data, obtaining consent from data providers and stripping out personally identifiable information.
The report concludes by stating that using 34 proprietary O’Reilly Media books, the study provides empirical evidence that OpenAI likely trained GPT-4o on non-public, copyrighted data.
(Image by Sergei Tokmakov)
See also: Anthropic provides insights into the ‘AI biology’ of Claude
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.
Explore other upcoming enterprise technology events and webinars powered by TechForge here.
Read the full article here