Foundation models like ChatGPT are changing society with their remarkable capabilities, serious risks, rapid implementation, unprecedented adoption and ongoing controversy. Meanwhile, the European Union (EU) is finalizing its AI Act as the world's first comprehensive legislation to regulate AI, and just now the European Parliament passed a draft of the law with 499 votes in favor, 28 against and 93 abstentions. The law includes explicit obligations for foundation model providers such as OpenAI and Google.
In this post, we assess the extent to which large foundation model providers currently meet these design requirements and find that they largely do not. Foundation model providers rarely provide sufficient information about the data, computational power, and implementation of their models, as well as the key features of the models themselves. This generally does not meet the design requirements to describe the use of copyrighted training materials, specify the hardware and emissions used during training, and how they evaluate and test models. As a result, we recommend that policymakers make transparency a first priority, informed by the requirements of the AI Act. The assessment shows that foundation model providers can currently comply with the AI Act, and that disclosure regarding the development, use and performance of foundation models would improve transparency across the ecosystem.
To fully comply with the requirements of the EU AI Act, foundation model providers should take the following steps:
1. Improved transparency:
Providers should provide adequate information on the training data, hardware, emissions and evaluation methods used for their models. This will ensure better accountability and regulatory compliance.
2. Copyright-related issues:
Providers should provide clarity on the use of copyrighted training materials and take the measures to reduce the risk of copyright infringement. Legislators and regulators should provide guidance on how copyright relates to the training process and output of generative models.
3. Energy consumption and emissions reporting:
Reporting of energy consumption, emissions and measures to reduce emissions should become standard practice for foundation model providers. Clear guidelines and metrics should be developed to accurately measure the energy requirements of training foundation models and make reporting these costs more reliable.
4. Risk management and evaluation:
Foundation model providers should make a thorough assessment of the potential risks their models pose, both in terms of malicious use and unintended damage. They should be transparent about the measures they take to mitigate these risks and evaluate the effectiveness of these measures. Guidelines for evaluation standards should be developed to assess the performance of foundation models in a consistent and reliable manner.
5. Release strategies:
Providers should consider their release strategies and their impact on transparency and accountability. Both open and restricted releases have advantages and disadvantages, but it is important that foundation model providers are aware of the consequences of their choices. It is essential that policymakers consider the different release strategies when formulating regulations to ensure that there is adequate accountability and transparency across the ecosystem.
The evaluation of foundation model providers' compliance with the EU AI Act shows that there is significant room for improvement in transparency and accountability. The implementation and enforcement of the AI Act will bring positive change to the foundation model ecosystem. It is essential that foundation model providers take action to establish industry standards that improve transparency, and that policymakers take action to ensure that sufficient transparency is the foundation of this overall technology. This review is just the beginning of a broader initiative to assess and improve transparency of foundation model providers to complement our efforts in holistic assessment, ecosystem documentation, standards development and policy recommendations.
source: Stanford University