In an effort to enhance transparency in the world of black-box neural networks, Stanford University, in partnership with MIT and Princeton, has introduced the Foundational Model Transparency Index (FMTI). This novel scoring system evaluates the top 10 AI models using publicly available data, revealing a concerning lack of transparency, with hope emerging from open-source models.
The Transparency Challenge in AI
The researchers behind FMTI emphasize the growing societal impact of AI models, coupled with diminishing transparency, akin to the opacity seen in previous digital technologies like social networks. The absence of transparency poses challenges for consumers in comprehending model limitations and addressing potential harm.
What FMTI Examines
FMTI assesses the 10 largest AI models across 100 diverse criteria encompassing transparency and openness. These factors include model structure, training data information, computing resource requirements, and policies related to model usage, data protection, and risk mitigation. A comprehensive list of metrics and methodology is available in the accompanying 110-page study.
Ranking and Implications
After sharing the test results with company executives and allowing them to challenge assessments, the researchers adjusted the scores. Despite this balanced approach, the average score for all models stands at a mere 37 out of 100 (37%). This outcome signifies that none of the models presently offer sufficient transparency. Meta’s Llama 2, Hugging Face’s Bloomz, and OpenAI’s GPT-4 top the list with 54%, 53%, and 48% scores, respectively.
The Open vs. Closed Model Debate
The ongoing policy debate centers on whether AI models should be open or closed. Open models, characterized by public code release, like Llama 2 and Bloomz, typically received higher ratings compared to closed models, including GPT-4. Stanford University’s Alpaca, developed from the open source code Llama, highlights the advantages of open models.
The researchers anticipate that the FMTI will encourage AI creators to improve transparency in their models, notes NIXsolutions. They plan to publish the FTMI ranking annually. Nine out of ten ranking participants have previously engaged in US government initiatives for responsible AI use. FMTI data could also prove valuable to the European Union in shaping the Artificial Intelligence Law, offering policymakers insights into AI’s current state and potential regulatory changes.