Is My Vision-Language Data in Your AI? Membership Inference Test (MINT) Demo 2

Users can now check if their data was used to train AI models: MINT Demo 2 provides a practical web platform that allows people to test whether specific images or text they own were included in training datasets for popular AI systems, achieving detection accuracy rates up to 90%.

This addresses a major transparency gap in AI development: The framework helps answer a fundamental question that has been difficult for the public to verify—what data companies actually used to build their AI models—which is increasingly important as AI regulations emerge that require disclosure of training data sources.

The tool works across multiple AI types: The platform can audit different kinds of AI systems including face recognition models, large language models (LLMs), and image/text databases, making it broadly applicable rather than limited to one type of technology.

Different audit approaches for different situations: MINT offers multiple technical architectures (MINT, aMINT, and gMINT) depending on how much information is available about the model being tested, making it flexible for real-world scenarios where companies may have varying levels of transparency.

This supports compliance with new AI regulations: By providing a concrete tool for detecting training data membership, MINT Demo 2 helps organizations and regulators verify compliance with emerging AI laws that increasingly require transparency about what data was used to train commercial AI systems.