The Two Key Goals of AI Governance

In my previous essays on “Polycentric Governance in the Algorithmic Age” and “AI Governance ‘on the Ground’ vs ‘on the Books,’” I explained how collaborative “soft law” efforts can go a long way toward improving accountability and responsibility among various emerging technology companies and individual innovators. Standards, codes, ethical guidelines, and multistakeholder collaborations create powerful social norms and expectations that are often equally or even more important than what laws and regulations might seek to accomplish.

The first phase of AI soft law development has been aspirational and focused on the formulation of values and best practices by soft law scholars, government officials, industry professionals, and various other stakeholder groups. Currently, and in years to come, the focus will increasingly shift to the implementation and enforcement of these values and best practices. The ultimate success of soft law mechanisms as a governance tool for AI will come down to how well aspirational goals get translated into concrete development practices. To reiterate, this involves the twin goals of:

(1) “baking in” or aligning AI design with widely-shared goals and values; and,

(2) keeping humans “in the loop” at critical stages of this process to ensure that they can continue to guide and occasionally realign those values and best practices as needed.

This is a useful way to the think about how to embed and align ethics, too. We essentially need the equivalent of transfer learning for ethical principles within AI systems as they evolve such that important values and principles are constantly embedded at each step of the process. Optimally, as algorithms and AI systems learn and develop new capabilities, the goal should be to ensure that the same guiding principles we have attempted to “bake in” remain and are extended. If AI systems can gain greater capacity to transfer and use knowledge it has learned from one task or application to another, by extension, it should be able to transfer and apply ethical principles and guidelines it has learned from one task or application to another.

Of course, human operators still need to be “in the loop” to correct for inevitable errors along the way. This does not mean the process is foolproof because not only will machines make errors, humans will as well. Moreover, as already noted, sometimes important values and best practices will be in tension with others and need to be balanced in ways that some parties won’t like. Nonetheless, the general framework of trained learning for AI ethics remains valuable.

Iterative amplification could be another way of thinking about how to gradually build safer AI systems over time. Paul Christiano, who runs the Alignment Research Center, a non-profit research organization whose mission is to align future machine learning systems with human interests, frame iterative amplification as follows:

Taken together, transfer learning and iterative amplification are essentially forms of learning by doing. As I’ve noted previously, it is a mistake to think of AI safety or algorithmic ethics as a static phenomenon that has an end point or single solution. Incessant and unexpected change is the new normal. That means many different strategies and much ongoing experimentation will be needed to address the challenges we confront today and the many others to come. The goal is to continuously assess and prioritize risks and then formulate and reformulate our toolkit of possible responses to those risks using the most practical and effective solutions available.

By intentionally eliciting problematic results from natural language processing models, and then taking steps to counter those results, red teaming represents the idea of ethical transfer learning and iterative amplification in action. However, Anthropic researchers correctly note that “[t]he research community lacks shared norms and best practices for how to release findings from red teaming,” and that “it would be better to have a neutral forum in which to discuss these issues.”

Luckily, there are many useful soft law mechanisms — some old, some new — that can address that problem and facilitate collaborative efforts. As I documented in earlier essays, many broad-based ethical guidelines already exist for AI development, and they are increasingly organized around a common set of values and best practices such as transparency, privacy, security, and non-discrimination. Again, professional associations like IEEE, ACM, ISO and others are particularly important coordinators in this regard. Industry trade associations and other NGOs also play a crucial role. The organizations and bodies need to work together to, in essence, align alignment efforts. That should include finding ways to better publicize red team research methods and results while identifying useful collective solutions to other common vulnerabilities that are identified.

The next step is ensuring that such values get translated into concrete guidelines and guardrails at the developer level. Marchant, Tournas, and Gutierrez highlight the growth of important internal measures that can help AI developers get serious about embedding ethics by design and ensuring that humans are kept in the loop along the way. In addition to the work done by professional bodies and trade associations, they identify many other important strategies to give shared norms and best practices real meaning:

[Note: This essay and several of the others referenced below are derived from a book I am finishing on the future of AI governance.]

Additional Reading:

Related posts:

A break to Turkey will please everyone’s likes and preferences, whether you want a quiet beach vacation or an action-packed journey loaded with excitement and thrills. Turkey’s coastline is one of…

Generating more money for your brand is crucial. A task that needs to be continuous. The purpose of this article is to outline some small things you can implement to amp up your revenue in a…

It is useful to start anything in clean mind. The same way, you can start docker. For that you have to stop all the running docker containers. Then you may proceed to remove all the docker containers…