Foundation models have emerged as transformative digital technologies, introducing new capabilities and risks that have captured unprecedented public attention. However, the current foundation model ecosystem lacks transparency, mirroring issues faced by earlier digital technologies like social media platforms. The 2023 Foundation Model Transparency Index revealed that major developers scored an average of only 37 out of 100 points for transparency. This opacity presents significant challenges to understanding and governing these powerful AI systems. As foundation models continue to evolve and impact society, there is a growing need for standardized, comprehensive transparency practices. Governments worldwide are beginning to address this issue through various legislative and regulatory initiatives, aiming to mandate public reporting and increase accountability in the AI industry.
Existing attempts to address transparency challenges in AI have primarily focused on model evaluations and documentation frameworks. Model evaluations aim to clarify strengths and weaknesses but often lack broader societal context. Documentation approaches, such as data sheets and model cards, provide more comprehensive information by posing open-ended questions about dataset creation, model development, and limitations. Ecosystem cards have been introduced specifically for foundation models, emphasizing the importance of tracking relationships between datasets, models, and applications.
These methods, however, face limitations in standardization and completeness. For instance, the Llama 2 model card, while addressing many high-level categories, omits several lower-level questions from the original model card framework. In addition to this, reproducibility checklists required by AI conferences have attempted to enforce some transparency standards. Despite these efforts, the current landscape of AI transparency remains fragmented and inconsistent, highlighting the need for a more structured and comprehensive approach to foundation model transparency reporting.
Researchers from Stanford University, Massachusetts Institute of Technology, and Princeton University propose Foundation Model Transparency Reports, which offer a structured approach to address the transparency challenges in the AI industry. These reports are designed to be published periodically by foundation model developers, providing essential information in a standardized format. This method is built upon recommendations from the G7’s voluntary code of conduct and the White House’s voluntary commitments, while also incorporating the 100 transparency indicators defined in the Foundation Model Transparency Index.
The proposed approach aims to consolidate crucial information, making it easily accessible to stakeholders and facilitating analysis and comparison across different developers. The transparency reports go beyond current government policies by specifying a precise schema for information disclosure, covering the entire supply chain of foundation models. By implementing these reporting practices, developers can establish stronger norms of transparency in the AI ecosystem, potentially improving compliance with various jurisdictions and reducing the overall compliance burden. The methodology also includes examples of report entries based on publicly available information, setting a clear precedent for future transparency efforts in the foundation model industry.
Foundation Model Transparency Reports are designed based on six key principles derived from the strengths and weaknesses of social media transparency reporting. These principles aim to create a more comprehensive and standardized approach to transparency in the AI industry. The first three principles build on the strengths of existing social media transparency reports: (1) Consolidation of information into a centralized location, providing stakeholders with a single, predictable source for relevant data. (2) Structured reporting that addresses specific queries, typically organized into four top-level sections, setting clear expectations for the report’s content. (3) Extensive contextualization of information to ensure proper interpretation by diverse stakeholders with varying levels of expertise.
The remaining three principles address the shortcomings of current social media transparency practices: (4) Independent specification of information to be included, preventing selective reporting by platforms. (5) Full standardization of both form and content, enabling easy comparison and aggregation of data across different platforms. (6) Clear specification of methodologies for computing statistics to avoid misinterpretation and ensure consistency in reporting. These principles aim to create a more robust and meaningful transparency framework for foundation models.
Building upon these principles, Foundation Model Transparency Reports incorporate indicators derived from the Foundation Model Transparency Index. This approach ensures a comprehensive coverage of the foundation model ecosystem, addressing various aspects of the supply chain. The reports are designed to provide specific, standardized information that allows for meaningful comparisons across different developers and models.
The structure of these reports is carefully crafted to balance detail with accessibility. They typically include sections that cover key areas such as model development, training data, model architecture, performance metrics, and deployment practices. Each section contains clearly defined indicators that developers must report on, ensuring consistency and comparability.
To facilitate implementation, the methodology includes examples of how developers can report information related to these indicators. These examples serve as templates, demonstrating the level of detail and format expected in the reports. By providing such guidance, the Framework Model Transparency Reports aim to establish a uniform standard for transparency in the AI industry, making it easier for stakeholders to access, interpret, and analyze crucial information about foundation models.
The Foundation Model Transparency Reports are designed to align with existing and emerging government policies, facilitating compliance across different jurisdictions. The methodology tracks six major policies, including the EU AI Act and the US Executive Order on AI, mapping the report’s indicators to specific requirements within these regulations.
This alignment serves multiple purposes. First, it incentivizes foundation model developers to adopt the transparency reporting framework, as much of the information disclosed will also satisfy legal requirements. Second, it provides a clear picture of how different jurisdictions prioritize various aspects of AI transparency, highlighting potential gaps or overlaps in regulatory approaches.
However, the analysis reveals a relatively low level of alignment between current government policies and the comprehensive set of indicators proposed in the transparency reports. This discrepancy underscores the lack of granularity in many governmental transparency requirements for AI. By offering a more detailed and standardized reporting structure, the Foundation Model Transparency Reports aim to not only meet but exceed current regulatory standards, potentially influencing future policy development in the field of AI governance.
To illustrate the practical implementation of Foundation Model Transparency Reports, the researchers constructed example entries drawing from the practices of nine major foundation model developers. This approach was necessitated by the current lackluster transparency practices across the industry, as revealed by the 2023 Foundation Model Transparency Index (FMTI).
The example report focuses on 82 out of 100 indicators where at least one developer demonstrated some level of transparency. For each indicator, the researchers selected the developer whose practices best-exemplified transparency, resulting in a composite report that showcases a variety of best practices across different aspects of foundation model development and deployment.
This exercise revealed several key insights:
1. There are still 18 indicators where no major developer currently provides transparent information, particularly in areas related to labor and usage statistics.
2. Even for the 82 indicators with some level of disclosure, there is significant room for improvement in terms of contextualization and methodological clarity.
3. The lack of a common conceptual framework among developers leads to inconsistencies in how information is reported, particularly regarding data pipelines and labor involvement.
4. For many indicators, it remains unclear whether the disclosed information is comprehensive or partial.
These findings underscore the need for more standardized and comprehensive transparency practices in the foundation model ecosystem, highlighting areas where developers can establish meaningful precedents and improve their reporting methodologies.
Transparency in foundation model development serves multiple crucial functions, from enhancing public accountability to improving risk management. As the field evolves, establishing robust norms and industry standards for transparency becomes increasingly important. Different aspects of transparency cater to specific societal objectives and stakeholder groups. Transparency in data, labor practices, computing usage, evaluations, and usage statistics directly informs the understanding of model biases, labor conditions, development costs, capabilities, risks, and economic impact. By fostering a culture of openness, the AI community can collectively address challenges, sharpen understanding, and ultimately improve the societal impact of foundation models.
Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. If you like our work, you will love our newsletter..
Don’t Forget to join our 52k+ ML SubReddit
Asjad is an intern consultant at Marktechpost. He is persuing B.Tech in mechanical engineering at the Indian Institute of Technology, Kharagpur. Asjad is a Machine learning and deep learning enthusiast who is always researching the applications of machine learning in healthcare.
