A major AI training data set contains millions of examples of personal data

Research reveals that vast quantities of validated personal identity documents, including passports, credit cards, and sensitive resumes, have been scraped from the public web. These datasets, often used to train generative AI models, inadvertently contain highly private information such as government identifiers, home addresses, and demographic details. This confirms that virtually any data published online is likely captured in these massive collections, posing significant privacy risks regardless of user intent. The primary concern centers on DataComp CommonPool, a widely used dataset intended for academic research but licensed for commercial use. Because it derives from the same sources as major models like Stable Diffusion, it shares similar privacy vulnerabilities. Given the dataset’s massive download count, numerous downstream commercial models are likely trained on this same compromised data, effectively duplicating and scaling these privacy breaches across the AI industry without transparency or consent. This situation is critically relevant to open data because it challenges the assumption that public data is safe for open sharing or reuse. It demonstrates that open web data inevitably contains non-public personal information, making simple availability insufficient for ethical use. Developers and policymakers must recognize that open data frameworks need robust privacy safeguards and validation processes to prevent the accidental proliferation of sensitive personal identities in AI training pipelines.

Source: technologyreview.com
Published on 2025-08-01

A major AI training data set contains millions of examples of personal data

Related news