Data Provenance Framework Enables Compliance for Generative AI Datasets, Tackling Exponential Growth

The rapid expansion of generative artificial intelligence relies heavily on large, openly available datasets, yet the ethical and legal foundations of these resources often receive insufficient attention. Matyas Bohacek and Ignacio Vilanova Echavarri, from Imperial College London, and their colleagues tackle this critical issue by introducing the Compliance Rating Scheme, a novel framework for assessing the transparency, accountability, and security of generative AI datasets. This research establishes a method for tracking the origin and legitimacy of data, information frequently lost as datasets are shared and modified online. By releasing an open-source Python library built on data provenance technology, the team provides a practical tool for both evaluating existing datasets and guiding the responsible creation of new ones, ultimately promoting more trustworthy and ethical development within the field of artificial intelligence.

Generative AI and Large Language Model Surveys

A comprehensive collection of research papers and articles concerning generative AI and large language models has been assembled, categorized for clarity. This includes surveys providing broad overviews of the field, as well as focused studies on specific models and techniques. Several surveys explore the landscape of generative AI, including comprehensive analyses of large language models and their capabilities. Dedicated research examines text-to-image, text-to-video, and audio diffusion models, detailing the advancements in these areas. Further studies investigate the emergent abilities of large language models and their planning capabilities, proposing new benchmarks for evaluation.

Research also addresses critical ethical and legal considerations surrounding generative AI, including data provenance, privacy, and copyright. Investigations explore the challenges of ensuring responsible AI development, particularly concerning the sourcing and use of training data. Studies highlight the need for frameworks to assess dataset integrity and compliance with ethical and legal principles.

Dataset Integrity Evaluation Using Compliance Rating Scheme

To address ethical and legal concerns surrounding large datasets used in generative artificial intelligence, scientists engineered the Compliance Rating Scheme (CRS), a framework for evaluating dataset integrity. This work pioneers a method for assessing compliance with principles of transparency, accountability, and security, scrutinizing data origins and construction. Researchers developed an open-source Python library, built around data provenance technology, to implement the CRS and integrate it into AI training pipelines. The system delivers a comprehensive assessment by examining the data’s journey from creation to use in AI models, identifying potential ethical or legal violations. Scientists harnessed data provenance to build a system capable of flagging data scraped without consent, recognizing that manual inspection of billions of data points is impractical. This innovative library provides a crucial layer of verification, mitigating risks and promoting responsible AI development.

Dataset Compliance Ratings for Generative AI

Scientists have developed the Compliance Rating Scheme (CRS), a new framework for evaluating the ethical and legal compliance of datasets used in generative artificial intelligence. Recognizing that large-scale datasets are often created using unclear practices, the team focused on data provenance and accountability. The work introduces a proactive and reactive system, capable of assessing existing datasets and guiding the responsible creation of new ones. Experiments revealed a significant issue with current dataset practices, demonstrating that nearly 50% of popular AI training datasets include data sourced without the consent of its creators, potentially violating copyright.

This highlights a critical vulnerability in the AI ecosystem, where manual inspection of vast datasets is virtually impossible. The team formulated four principles for accountable, license-compliant datasets and conceptualized the CRS as a tool to evaluate compliance. They then developed DatasetSentinel, an open-source Python library designed to integrate these principles into dataset processing and training pipelines. Recognizing that the growth of GAI relies heavily on large datasets often created with unclear practices, researchers designed this scheme to assess compliance with principles of transparency, accountability, and security. Alongside the framework, they released an open-source Python library that integrates data provenance technology, enabling both evaluation of existing datasets and responsible construction of new ones. This work addresses a critical gap in the field, moving beyond the development of AI models to focus on the foundations upon which they are built.

By providing a means to assess and improve dataset quality, the researchers aim to foster more responsible AI development practices and encourage greater awareness of ethical and legal considerations surrounding data. The library’s dual functionality, both reactively assessing existing data and proactively guiding new data collection, offers a practical tool for implementing these principles. The authors acknowledge that the library’s effectiveness is currently limited by its dependence on existing data provenance protocols and support for specific data types. Expanding support for emerging data types will be an ongoing challenge, and broader discussion within the AI community is needed to promote a shift in values, recognizing that technical solutions alone are insufficient to address unsustainable dataset practices.