As the realm of Artificial Intelligence (AI) expands, we find ourselves on the cusp of a new frontier—one where innovative technology promises profound advances with more general capabilities but also carries considerable risks. There is a lot we don’t understand so far; therefore it is imperative to navigate this landscape with a governance strategy that maximizes learning and balances innovation with safety. 

As we explore this frontier, metaphors can help us grasp the future and imagine what is not yet reality. They shape how we think about this new frontier of technology. Just as we’ve borrowed from Hollywood to envision what AI could be, let’s explore two fun and hopefully insightful ways to think about keeping AI exciting, useful, and safe: 

• The “Space Exploration Approach” where we see these new technologies as means to explore new worlds or capabilities that we’ve never seen before; and
• The “Snake-Filled Garden Approach” that pictures AI as a garden filled with both harmless and dangerous snakes.

Space Exploration Approach: Journey into the Unknown

This approach likens AI models to spacecraft venturing into unexplored celestial territories. Think of each new AI model as a spaceship. Before it “launches,” developers need to provide the plans, clear the airspace for launch, etc. And, after launch, just like NASA keeps us posted on space missions, developers should regularly update us on how their AI models are doing and what they are discovering. Here are a few ideas on how AI would be managed from this perspective:

• Registration: Treat frontier AI models like space missions, requiring developers to register system data for “launches” (new models) and report “new discoveries” (capabilities) that nobody has seen before. Important explorations happen early (before they become products); therefore, it seems wise that registration should occur before or during the training/development phase. 
• Updates: Mandate periodic updates on these “missions” (models and their use) and report any new capabilities.
• Assessment: Require independent third party audits when specific risk criteria are met for new capabilities. This gives governments the authority to request deeper evaluation, especially for powerful proprietary models that are otherwise subject to IP protection. Risk criteria for triggering these assessments will need to be updated periodically based on model advances, risk mitigation techniques and experience.
• Open Source & Community Engagement: During this ‘space exploration’ phase, it seems wise to focus on collecting data on newly discovered issues and capabilities. The emphasis should be on gathering information and facilitating learning, rather than hindering open innovation or allowing only well-funded companies to dominate the field. Moreover, regulators need to actively engage the developer communities to propose solutions for identifying new risks and mitigating misuse.

Snake-Filled Garden Approach: Risk Containment and Management

This approach treats AI models like gardens containing a mix of known and unknown venomous and non-venomous snakes. In this case, we need to run a set of standard tests to identify the known venomous “snakes” and then use a variety of methods to find and identify new dangerous “snakes.” When found we ring the alarm. If it’s harmless, we don’t make a fuss. A few concepts that would help govern AI from this perspective could be the following:

• Systematic Use-Case Testing: Apply a uniform set of tests for known venomous snakes (risky use cases, such as money laundering or medical diagnoses). This “snake search” can be applied to models developed as a “space exploration” all the way to a commercial product phase. Automated versions can provide lower-cost, scalable checks, but will miss some snakes.
• New Risk Detection: Encourage ongoing risk assessment and specify criteria for mandatory reporting, for example, it is essential to report a new venomous snake species (harmful capabilities), while a new non-venomous snake species (capabilities that are not harmful) may not require this step. These tests require more substantial funding and people involvement. It is not yet clear how to do this systematically, but it would likely require some sort of state-sponsored subsidy, clever test automation to improve scalability, or some sort of evaluation threshold (e.g., based on the size of the user base).
• Mitigation: Update deployment plans based on the results of these tests and assessments. Some models (likely including open-source versions) would need to be removed from the market due to newly discovered risks to users or society. It is not yet clear how to do this, but this would need to be figured out.

Maximizing Learning

The decision to adopt either approach depends on our expectations for what  AI is able to do as well as how much societal resources we’re willing to invest in uncovering currently unknown risks. Regardless of the approach, it’s crucial to establish methods for learning as we go. We need to maximize our learning at each step of the process and update what we do next based on what we learn. One way to do this is to phase in both voluntary and mandatory incident reporting mechanisms; these will enable data-driven decisions and help us measure real cases of AI-related harms that are currently hard to predict.

Navigating the uncharted territories of AI necessitates a governance model that is as dynamic and evolving as the technology itself. These metaphors, or potentially a combination of both, can help us work out how we want to balance the scales of innovation and safety as we forge ahead into this new frontier.

Want to Go Deeper?

For those interested in diving deeper into the governance of frontier technologies, the Center for Security and Emerging Technology (CSET) recently organized a roundtable discussing these risks. Additionally, emerging frameworks like Responsible Scaling Policies offer guidelines for frontier AI developers to identify and manage risks effectively.