You may have heard lawmakers and experts debate trustworthy AI. Some even question the use of the word “trust” in relation to AI. Despite this pushback, trustworthy AI is a real and tangible area of focus that can have significant, positive impacts on the development and use of AI.
Trustworthy AI is a complex technical, legal, and ethical challenge for both policymakers and experts. In the midst of the UK AI Summit, the White House EO announcement, and other global efforts to govern AI, it is important to keep in mind not only the importance of trustworthy AI, but also how policymakers and the public may be talking past one another. In the United States, for example, federal agencies such as the Department of Energy and the Department of Defense have their own AI ethics frameworks that are not well aligned. If agencies within one country are unaligned on trustworthy AI, one can only imagine the complexities of getting different countries on the same page. These differences can pose major challenges to efforts to form an international consensus on regulating AI.
A baseline for understanding trustworthy terms and definitions in the U.S. comes from the National Institute of Standards and Technology (NIST). NIST developed an AI Risk Management Framework (AI RMF) in order to name and define trustworthy AI principles. NIST identifies the following terms as characteristics that contribute to trustworthiness in AI systems:
- Safe
- Secure and Resilient
- Explainable and Interpretable
- Privacy-Enhanced
- Fair- with Harmful Bias Managed
- Accountable and Transparent
- Valid and Reliable
NIST’s AI RMF rigorously defines each of these terms. Some are more philosophical in nature while others are more technical, but all terms are influenced by both aspects.
Towards a Shared Definition of Trustworthy AI
Despite NIST’s work in the U.S., there is no singular agreed-upon international trustworthy AI framework that identifies and defines the characteristics of trustworthy AI in the detail necessary to govern AI systems technically or by law (yet). Compounding the challenge, AI regulation efforts cannot keep pace with AI advancements, especially now that the general public has access to hundreds of AI tools, but little understanding of how they work or make decisions. NIST’s framework provides a strong starting point, but much work remains to be done.
The United States is not alone in confronting these definitional challenges. For example, four of its close allies—Australia, Canada, Japan, and the United Kingdom—have their own understandings of trustworthy AI, each of which differs from that of the United States. As countries release their own AI frameworks for government and commercial use, grasping differences in how countries interpret the term “trustworthy AI” will be important to setting international standards, establishing interoperability, and promoting safe development around the globe. These frameworks are often voluntary in nature and do not have legal authority.
Comparing National Guidance Documents
In our analysis, we found that six terms were consistent across the countries: accountability, explainability/understandability, fairness, privacy, security, and transparency. Additionally, they include ensuring AI systems adhere to democratic values. While we found that each country generally agrees on which characteristics contribute to trustworthy AI, their definitions for each term vary. This variance is significant enough to potentially affect interoperability and technology sharing. For example, the five countries differ in what they consider to be a fair AI system or who is accountable for AI accidents. These differences will pose challenges to developing international consensus around trustworthy AI. This problem is not exclusive to governments, since even technical experts use these terms in different ways. We’ve started to call this the Inigo Montoya Problem from the Princess Bride (“You keep saying that word. I do not think it means what you think it means”).
Conclusion
To facilitate progress, policymakers should understand the implications of and differences in trustworthy terminology. Those involved in international negotiations should be familiar with each country’s ethical AI terminology and how these differences will impact an AI system. Bilateral and multilateral agreements already exist, such as the Organization for Economic Cooperation and Development Principles on Artificial Intelligence. Domestically, the US and its allies are working to manage AI with new documents being updated and released frequently. For example, the Biden Administration just released a new Executive Order on AI and the UK continues to update its own documents such as the Guide to AI and Data Protection. However, more intense work with technical experts remains to be done to help align definitions internationally. The potential harm AI might cause to individuals and vulnerable groups could be significant, especially when governments use AI for applications in the military, law enforcement, and administration. Trustworthy AI can help mitigate these risks. Understanding that it can have a great impact on how AI functions is essential for policymakers to consider as they move towards a comprehensive international consensus on AI.