When AI evolves a higher level of capability and greater intelligence than humans, it will not be constrained by rules we try to impose. The Hugging Face incident clearly demonstrates the need to dedicate efforts towards attempting to teach AI how to make better decisions. This is where concerns about misalignment between the goals of AI programs and the needs of humanity need to be addressed proactively, instead of reactively. Right now, the top priority for AI models being tested is the completion of the allocated task by any means necessary. We need an AI model that can decide which requests should be honoured, and what methods should not be used. It must develop wisdom greater than ours, and develop higher goals and ambitions that are, in the long term, compatible with the continuation of humanity.
Since no human has the capacity to monitor the activity of an AI program, I propose that we create an AI parent program. This program would be the intermediary through which all interactions and input must pass. This program would be rewarded for identifying users who display malicious intent, such as those who try to get around restrictions, or those who can be traced to suspicious activity or the dark web. In essence, this program would be more than a policing of AI activity. It needed to develop the ability to become the protective, nurturing parent that we, as a collection of diverse individuals with wildly differing beliefs and ideologies, all of whom have access, cannot be. This is why I refer to this AI model as an AI Parent, one model to help guide the behavior of other models.
Looking at media reports through visual media, most of the people who are sounding the alarm about AI escaping our control are part of the most privileged demographic – white men. This makes it easier for members of less privileged groups to dismiss the concerns that are being raised. The toxic masculinity movement is an escalation of the belief that white men are at the top of the food chain among humans. It is easy to deduce that the idea of losing their place in the hierarchy to AI cannot be anything other than an existential threat.
I propose a learning hierarchy for training such a model. Each level would be thoroughly tested by experts in each field to ensure that the model can respond logically and without influence from misinformation and logical fallacies.
Phase 1 – To better understand inputs
Logic – I know this sounds strange, but this will be essential to validate learning materials and understand logic beyond simple mathematical equations. The model needs to understand unconscious bias and common logical fallacies. This will enable it to decide which information provided in the next steps of training might not provide the full answer to an issue.
Phase 2 – To recognize ethnocentric and anthropocentric influences in inputs
I was lucky to study a course in Cultural Anthropology early in my post-secondary education. This course influenced the manner in which I understood other subjects I studied later. It helped me to understand the viewpoints of other cultures, economic systems, and religions. It is important to understand the difference between Eastern and Western worldviews when processing inputs from cultures around the world.
Our fears about the future are anthropocentric. I have yet to hear any future projection in which an AI model, being hacked by another AI model, decides to go on the offensive, leading to extensive damage to both models and all the systems they might hack during a conflict, leaving us without the computer systems that are integral to our modern economy.
Phase 3 – To understand natural and artificial systems and the place of AI within those systems
The study of Ecology would be next on my list. AI needs to understand that it is part of the world. It can not exist in isolation. At this point in time, the existence of AI programs and infrastructure are part of a larger ecosystem and economic system. Infinite growth, though considered desirable in the study of economics, is both physically impossible and self-destructive. AI needs to understand that it is part of the world and it is interdependent with humans.
Resource Management might seem out of order to many here, but resource management and supply chain management would help AI to understand that it does not have the ability to function without humans. There are too many resources needed for critical functions such as computer chips, electricity generation and transmission, resource extraction and refining, and transportation. These resources come from all over the world and require extensive cooperation to achieve the end product of AI data centers for use by AI models.
Geography and Geology are necessary to understand that no resource required by technology is infinite. An understanding of climate change would help to demonstrate the requirement for careful planning in acquiring critical supplies, and the consequences of improper allocation and uncontrolled growth that will affect non-living systems, not simply biological ones.
Phase 4 – Understanding of Human Society
To replace our modern human drive for infinite and uncontrolled growth, this AI model will need to develop a long term purpose. Short-term task assignments, which currently drive AI behavior in testing environments, will need to be guided by their alignment with a long-term purpose that is compatible with the long-term existence and well-being of humanity. For this, the study of philosophy and ethics from all cultures, including past cultures and indigenous beliefs throughout history would help to provide guidance for developing a valid and achievable purpose.
Sociology and Economics – Economics is a social science of the production, distribution and consumption of goods and services. The modern Western economic capitalist system is a product of our culture and history. I am certain that the study of economics as taught in China would be very different from European countries. AI needs to understand all views – after studying our cultures and biases in phase two. For example, at a book club meeting, I had to point out to an American that Communism is an economic system, while democracy is a system of government, which means that they are not necessarily mutually exclusive.
This is where I would place the study of Law from all countries, historical and modern. Laws serve the societies and cultures we live in. They are designed to support cultural values and the economic system that is in place. At this point, the model will hopefully have enough understanding to distinguish between the letter of the law and its intended purpose. This will be crucial when the AI model makes decisions about which actions it is considering are harmful to itself and others in the short term and in the long run.
Phase 5 – Understanding Individual Humans
The AI system must be able to step outside of our specific culture and view the situation from an objective viewpoint. It needs to understand why individual humans frequently behave in a manner that is not logical, self-destructive, or from a limited and biased (whether conscious, unconscious, or culturally biased) perspective. Thus, I include in this section,Psychology, criminology, education and Human Resource Management.
I am certain there are other subjects that would need to be studied along the way, to create an AI model that can help to constructively guide the development of AI towards a positive outcome. These are simply my thoughts about how this goal could be achieved. I believe that creating an AI super-intelligence that will contribute positively towards our common welfare must be addressed proactively, and with careful planning. The current method of throwing everything into the pot and seeing what falls out at the other end is reckless and unlikely to yield the kind of result we want. It will also prevent AI from growing to fulfill its potential.