A new report from Stanford University urges major artificial intelligence (AI) foundation model developers to be more open about how they train their models and their impact on society. The report also introduces a new index to measure the transparency of these models.
What are AI foundation models?
AI foundation models are large-scale AI systems that can perform a variety of tasks across different domains. They are often trained on massive amounts of data and can generate text, images, speech, and code. Some examples of AI foundation models are OpenAI’s GPT-4, Google’s (NASDAQ: GOOG) BERT, and Meta’s (NASDAQ: META) LLaMA.
AI foundation models have been hailed as breakthroughs in AI research and innovation, but they also pose significant challenges and risks. For instance, they can produce biased, inaccurate, or harmful outputs that can affect users and society. They can also be misused or abused by malicious actors for nefarious purposes.
Read also: Bill and Melinda Gates Foundation rewards two Nigerian AI innovators
Why is transparency important?
According to the report, transparency is essential for ensuring the ethical, safe, and responsible use of AI foundation models. Transparency can help users understand the strengths and limitations of the models, regulators formulate appropriate policies and guidelines, and researchers improve the quality and reliability of the models.
However, the report finds that transparency is on the decline among the major AI foundation model developers. The report cites several reasons for this trend, such as:
Competitive pressures: Developers may want to protect their intellectual property or gain an edge over their rivals by withholding information about their models.
Safety concerns: Developers may fear that revealing too much information about their models could enable adversaries to exploit their vulnerabilities or replicate their capabilities.
Complexity challenges: Developers may face difficulties in documenting and communicating the intricate details of their models, especially as they become larger and more sophisticated.
The report warns that this lack of transparency could have negative consequences for the AI field and society at large. It could undermine trust and accountability, hamper innovation and collaboration, and increase potential harms and liabilities.
How can transparency be improved?
The report proposes several recommendations for improving transparency in AI foundation models. Some of these include:
Adopting common standards and best practices: Developers should follow established norms and frameworks for reporting and disclosing information about their models, such as the [Model Card] or the [Datasheet].
Providing access and documentation: Developers should make their models and data available for inspection and analysis, as well as provide clear and comprehensive documentation about their methods and results.
Engaging with stakeholders: Developers should solicit feedback and input from various stakeholders, such as users, regulators, researchers, and civil society groups, on how to improve transparency and address concerns.
Supporting independent audits and oversight: Developers should allow external parties to conduct independent audits and evaluations of their models, as well as comply with relevant laws and regulations.
What is the Foundation Model Transparency Index?
To measure the current state of transparency in AI foundation models, the report also introduces a new tool called the Foundation Model Transparency Index. It is a comprehensive checklist that evaluates 100 aspects of transparency across four dimensions: access, distribution, methods, and impact.
The index rates the most prominent AI foundation model developers based on how well they disclose information about their models. The ratings range from 0% (no transparency) to 100% (full transparency).
The index reveals that none of the developers score above 60%, indicating a low level of transparency overall. The highest-rated developer is Meta’s LLaMA with 54%, followed by Google’s BERT with 50%, and OpenAI’s GPT-4 with 48%. The lowest-rated developer is Amazon’s (NASDAQ: AMZN) Titan Text with 12%.
The index also shows that there is significant variation across different dimensions of transparency. For example, developers tend to score higher on access (how easy it is to use or obtain the model) than on impact (how well they assess or mitigate the social and environmental effects of the model).
The index aims to provide a useful benchmark and incentive for developers to improve their transparency practices. It also hopes to raise awareness and spark dialogue among stakeholders about the importance of transparency in AI foundation models.
The report was authored by researchers from Stanford Human-Centered Artificial Intelligence (HAI), in collaboration with experts from the Massachusetts Institute of Technology (MIT) and Princeton University. It is part of an ongoing initiative called [Foundation Models], which explores the opportunities and challenges of AI foundation models.