Transparency Through Documentation: A Pathway to Safer AI

Across the globe, people are using generative AI systems, built on foundation models, everyday to help with homework, act as customer service representatives, and create images based on user requests. Although the foundation models these systems are based upon are impressive, their complexity often creates uncertainty and ambiguity for everyday people, civil society, and/or regulators. That is why proper and responsible documentation of the development and deployment of these systems is crucial. 

Documentation acts as a roadmap, guiding developers and providing insight into these complex systems. For policymakers, comprehensive documentation can assist in enforcing compliance and accountability, ensuring that AI systems adhere to legal and ethical standards. Civil society organizations can benefit from proper documentation, as well, to bridge the information asymmetry between AI developers and the public, enabling better advocacy and oversight. It can also enable application developers to build user-facing services safely and responsibly. 

So what exactly is documentation? And what is its key function? 

AI documentation is a collection of artifacts such as written documents and processes that provide information on how a system is built, works, and operates. It can include instructions for how the AI system is meant to be used, information about the development process and key features, and provide regulators a way to monitor these systems post-deployment. 

Despite the importance of AI documentation, current industry practices are inconsistent, incomplete, and lack standardization.

Although efforts to create policy for the development of safe AI are underway, such as the EU AI Act Code of Practices and the NIST Artificial Intelligence Risk Management Framework, without global alignment and collaboration, the industry is at risk of developing a fragmented understanding of good practices. 

Partnership on AI has been working to understand the current landscape of global frameworks’ documentation requirements and gauging current industry documentation practices. It is essential for policymakers and industry leaders to come together to develop standardized documentation practices to create an ecosystem of transparency and accountability. 

The Importance of Documentation and Navigating its Challenges

To understand the importance of transparency within complex systems we must first understand the systems themselves. Many of the AI systems we interact with today are built on foundation models. Foundation models are large deep learning neural networks, such as large language models like Claude 2, Llama 2, and GPT-4o which are trained on large datasets and are able to perform a wide variety of tasks such as summarizing text  and generating text and images. 

However, with vast capabilities comes great responsibility. 

In order to ensure these ubiquitous systems are functioning properly and working safely for the millions of users interacting with them on a daily basis, we need to have transparency into how the systems work and how they can become tools for exacerbating harm. Documentation is essential to creating safe AI systems, including:

• Supporting Accountability, Auditing, and Safety Research: Documentation helps enable scrutiny of black-box systems and helps regulators review how a system has been developed, how it works and the impacts of its use. This helps regulators ensure AI systems are performing as intended without causing harm.
• Transparency and Reproducibility: Good documentation allows developers, researchers, civil society organizations, and regular people to verify the safety of these systems. It ensures that modifications to the systems are detectable to others and replicable.

There are a myriad of challenges to documenting AI systems due to the volume and complexity of the data used in training the foundation models. The opaque nature of these systems makes it difficult to track changes and updates, which is critical to maintaining safety standards.

A lack of standardized documentation practices makes it challenging for regulators, developers, and civil society organizations to assess AI safety across the lifecycle of these systems.

Creating documentation has challenges, as users’ private information and some security flaws should not be publicly exposed, but these issues can be addressed through a committed and collaborative approach to developing standards across the industry and globe that balance transparency, safety, and confidentiality.

Global Interoperability and a Path Forward

To develop safer AI systems, documentation practices need to be standardized globally and across the industry. International efforts to establish a standardized process should include interoperable frameworks that facilitate the sharing of critical safety information across the globe. It is also important to have collaboration between industry stakeholders and civil society in setting documentation standards to ensure that the needs of all stakeholders are being met, including vulnerable groups, such as those of marginalized communities.

Over the past year, PAI has been working on understanding this landscape to determine a path forward for standardizing documentation practices. Later this year, we will be releasing two reports, the first of which will outline the current landscape of leading national and international documentation policy frameworks, and the second of which will detail the current landscape of documentation practices across the industry. The first will explore whether current policy frameworks for foundation model documentation are interoperable and what challenges to interoperability exist that can impact the development of best practices and accountability globally. It also proposes recommendations to help overcome these challenges. The second report on post-deployment documentation will look to build towards the future. It will provide a basis for future policy and guidance development by highlighting the value of documentation on foundation models after they are deployed while looking at the current progress of the field in disclosing this information.

The importance of documentation in AI cannot be overstated, and PAI will continue to lead in advocating for global alignment on standardized practices to develop safer AI systems that work for everyone. To stay up to date on our progress in this space sign up for our newsletter.