What Is AI Governance
AI Governance is a set of policies, procedures and frameworks. The intention is to ensure that AI systems are safe and accountable. It’s to manage risks like biases, privacy and security. At it’s core AI governance components include ethical guidelines, risk management, compliance and operational control.
Governance in Practice
AI governance casts a wide net. While there are organizations making strides in this area, it’s a slow process. This is more the political arm of AI and as such, it tends to react slowly. By the time certain rules or laws are passed, they may no longer even be relevant. Technology has moved on. In addition, the practice of building in and verifying the rules or laws have been followed, takes man hours away from building out functionality. A developer may need to spend as much as half their work day to insure compliance. When you have an organization spending this much time on verification, it costs. The costs are in both man hours and project length.
Governance Wrap Up
To start I would say I am a proponent of AI governance. It’s a start. It may be slow and costly to implement, but the net cast by governance is a good road map in the development of AI. Yes it’s slow, it can be costly to verify. Yes it has not yet achieved global acceptance, which means AI development in silos occurs today. Keep in mind that AI is currently in it’s wild west stage. It’s exciting, given the possibilities, but it’s also very dangerous if bad actors implement this technology for their own benefit. So what can we do to speed things up? I would propose building pillars of AI governance into the framework itself.
AI Architecture – The Implementation of Governance
AI governance can and should be the intellectual body watch dogs of AI development, but I propose we build the pillars into the AI architecture itself. Ethical guidelines, risk management, compliance and operational control can all be addressed at the architectural layer itself. Should that occur, developers are free to focus on implementation, not compliance. Compliance would be carried out by the underlying architecture. Having said that, this too is not without it’s own set on unique problems.
AI Architecture – Implementing Governance Issues
First, just from a pure technology viewpoint, the more layers of governance or rules implemented will have an impact on performance. The speed at which AI can process. For now I think this is a necessary trade off. Having said that, processor speed and memory continues to advance rapidly, not to mention quantum computers currently being explored in the labs. Speed of processing is important, but I would argue the implementation of ethics, risk management and operational control are even more important.
AI Architecture Implementation
Here’s where things get dicey. This is no simple task to carry out, nor is it purely a question on technology, the how of the implementation. Whether this is coded in or burned into the chip level, there are still big questions to be answered. For example, the treatment of women in one country can be different from other countries. The strive for military power or economic competitive advantage also carries its own definitions of what is ethical or not. I would propose, for a phase one implementation, to address the universal needs all countries have. Food, water, clothing, housing, medical care, education, environment and fairness. Risk management and operational control would also be added to this new layer. Taking those as the immediate need would insure we’re proceeding in a way that’s beneficial for all. The projects undertaken need to address a current, human specific need.
AI Governance vs Architecture
To wrap up this short article, AI Governance vs Architecture, I believe both are paramount steps to moving forward. There are roadblocks for both, most specifically, global acceptance. This won’t be easy, but nothing that has so much potential rarely is. This is not simply a technical issue to be solved, it’s laying the groundwork for a foundation that will impact future implementations.
