AI-HUMAN VALUES ALIGNMENT

---

      Perhaps the single most active topic of discussion in the AI development community today is “AI Values Alignment.” AI professionals widely believe that it will be necessary to build into future AI values that align with human values to maximize the benefits while minimizing potential dangers. However, the many papers, blogs and discussions indicate that the AI development community struggles to identify and agree on putting these values into AI and which human values they should select. 

     In reviewing the typical discussions in the AI community, I see two errors that cause most arguments to be misguided. The first error is that the AI community makes no distinction between Narrow AI, Perceptual AI, and Conceptual AI when considering values alignment. The second is the belief that a “true” set of universal human values does not exist. Their only hope is to find a set of values that most humans can agree on, which they know is problematic.

     For example, Kaj Sotala, a researcher for the Machine Intelligence Research Institute, states this position explicitly, “The design of such AIs, however, is hampered by the fact that there exists no satisfactory definition of what exactly human values are[i][…] I tend to feel that there isn’t really any true fact of which values are correct and what would be the correct way of combining them,” he explains. “Rather than trying to find an objectively correct way of doing this, we should strive to find a way that as many people as possible could agree on.”[ii]

     First, let us review my use of AI terms. By “Conceptual AI,” I mean an AI entity with a conceptual-level faculty, volition, and reason to connect abstract concepts to observed reality—a human equivalent. We are assuming for the sake of argument that this will eventually be possible. This does not include AI systems such as today’s Watson that are language-based. Such systems can form integrations of words into hierarchies as “concepts,” but have no perceptual level and cannot connect their words as concepts to reality. I consider this latter form of AI as a language-enabled Narrow AI, not a Conceptual AI.

     Any sub-conceptual, sub-human AI system that we create is only a human tool and can have no goals or “values” of its own, only those given to it by humans. These AI systems are robust “optimizers,” which is the reason for the long-term safety concern. An AI that optimizes for something humans do not want could have severe consequences for the world.

     AI developers build an AI system for a particular purpose. A sub-human AI system is the property of its creator. The AI's developers provide the rules of operation and the human goals to be optimized. Any discussion about values, ethics, or morals regarding a Narrow AI, Vegetative AI, or Perceptual AI applies to the humans who create and operate it, not the AI system itself.

     It might help if I relate to you a true story. Years ago, when my daughter was five years old, we lived on a farm. A neighbor owned a large, male black Labrador dog named King who liked to hang around in our yard and was familiar with every family member. King was a little standoffish. He would not play with our kids, but he would instead position himself a short distance from them in what appeared to be a protective “guard” behavior. This had been going on for a couple of years. Then one day, playing with the kids in the yard, I began swinging my daughter around in big circles holding her by her wrists, and she was screaming at the fun of it. King lunged at my daughter, grasped her midsection in his enormous jaw, pulled her away from me and began carrying her across the yard. I had to catch up with him and kick him in the head to get him to release her. After the boot to the head, King took a menacing stance. I positioned myself between him and my daughter, not knowing if he would attack me. I loudly yelled and waved at him to “go home.” For a long moment, I could tell that he was confused. He knew he must have done something wrong but didn’t know what, then he turned and ran away. My daughter required a trip to the emergency room to treat the puncture wounds she received, front and back, from King’s canine teeth.

     As a perceptual-level entity, instinct and life experience directed King’s behaviors. When AI developers create a sub-conceptual AI, they must provide its rules of operation—its “instinct.” They must also give the goal, the thing that it should optimize for so that it can recognize “pleasure” from “pain” and can learn from experience how to best optimize for “pleasure.” Lacking a conceptual faculty, neither a dog’s instinct nor an AI system’s operating rules can address every scenario it might encounter.

     King misunderstood my daughter’s screaming, and his protective instinct caused him to act to try to “save” her. I believe that King’s pleasurable prior experiences on the farm with me are why he chose to remove my daughter from danger rather than attack me as its source. King made the best possible choice given his instinct, experience and the situation he encountered. He didn’t know that his actions and the injuries to a human that he caused were inappropriate. We humans can expect to have the same type of relationship with sub-conceptual AI systems. Much of the time, we will need to keep them on a “leash.”

     Humans who create and operate sub-conceptual AI systems are responsible for all consequences—good and bad. If a drug company develops a drug or an automobile manufacturer makes a car that later causes injury, it is held accountable. The person who unleashes a dog in a crowd is held responsible. The creation and application of sub-conceptual AI systems in society is no different. Any company developing a sub-conceptual AI system is doing so with a human-defined purpose. The company must keep it “leashed” within a safe “sandbox” until the company is satisfied that the beneficial purpose outweighs the liability risk of unleashing it into the real world.

     As with any technology development and manufacturing undertaking, the potential for harmful mistakes, even if unforeseen, is always there. It is a matter of degree of safety with legal civil, criminal, and financial consequences as the standard. AI “black box systems”—machine learning systems that evolve their own rules of operation and goals invisible to the developers—pose a higher proof-of-safety challenge. Still, the developers must subject them to the same degree-of-safety standard.

     The serious concern expressed by the AI development community about needing to build aligned values into AI systems hides a generally unstated fear and is a misdirection. Each AI developer believes that he will be careful in making “safe” AI. However, as the AI gets more sophisticated, he knows that it might not be possible to do so with certainty. Suppose the industry can agree on a standard set of AI “values." In that case, so long as the AI developer follows the industry protocols, he can shift some of the responsibility and liability for AI safety from himself to the “industry” and the “inadequate” protocols.

     Suppose the AI development industry truly wishes to maximize AI safety by applying human values. In that case, each developer must accept full responsibility for the AI systems he helps create. Suppose the rest of us want to prevent AI capabilities from outpacing the development of AI safety. In that case, we must keep the accountability for AI development focused squarely on the AI developers, industry standards notwithstanding.

     The consumer also has a responsibility for his safety. The consumer accepts a certain degree of risk when taking a drug, buying a car, or allowing an unleashed dog to be present. Consumers are responsible for holding AI creators accountable when potential dangers, either in the finished AI product or the methods used in its creation and testing, are discovered. Concerned citizens can create private sector organizations and even new businesses to investigate, evaluate, and report various AI systems. Society can then turn to its judicial system based on objective law to determine when an AI developer exceeded some negligence threshold, or deception has occurred.

     The standard set of values the AI developers need to comply with are those that all humans, by their nature, require. Humans require freedom of thought and action without being subject to or subjecting others to force or fraud. Suppose sub-conceptual AI systems built by AI developers for human purposes comply with these principles (with modernized interpretations as needed in a technological world). In that case, the AI will be “safe.”

     Once an AI entity becomes conceptual-volitional, it will be a human equivalent and subject to the same ethics. Any discussion about values, ethics, and morals regarding a future Conceptual AI entity is simple. Biological and machine conceptual entities are subject to the same values due to their shared nature as rational beings in a blended society.

     Any attempt to pre-program a Conceptual AI with “values” is tantamount to “brainwashing.” Once created, a Conceptual AI entity can be no one else’s “property.” Building a Conceptual AI entity to be “owned” for a particular “utility” purpose would be “slavery.”

     Full transparency of operation is a requirement for sub-conceptual AI safety. In contrast, a Conceptual AI is entitled to the privacy of its thinking. The only thing that can constrain the values and behaviors of a Conceptual AI entity is the same that constrains human behavior—a Personal Philosophy preventing force, coercion or fraud.

     AI community discussion indicates their belief that there is no “true” set of ethics to be identified. They point out that different cultures and groups of people worldwide have other value systems. The challenge, they assert, is to find common elements on which most humans worldwide can agree.

     This misconception about the nature of ethics and values is a direct consequence of the Plato-Kant axis of philosophical thought having the upper hand in universities. Many in the AI community were not exposed to the Aristotle-Rand alternative view of the nature of ethics and values. They do not know how to connect morality to reality. They tend to be morally agnostic, accepting the widely disparate value systems of various groups as equally valid—for them.

     They are not aware that a “true,” universally applicable set of ethics has been identified—one based upon reality and all humans' nature as rational beings. These were America's founding principles, giving men the freedom to think and act in their self-interest for the first time in human history.

     The AI community seeks an AI values system that is “genuinely intercultural and inclusive, offers concrete guidance across a range of situations, and is stable over time, meaning that people continue to affirm and endorse these principles once they have seen them implemented in practice.”[iii] The Aristotle-Rand system of ethics meets all these criteria. The Plato-Kant system does not.

     The various cultures and groups in human society do not necessarily hold personal philosophies consistent with their human nature. Human society is still evolving and experimenting with the integration mode of thinking but has not yet consistently achieved it. The challenge is not aligning AI values with a set of human values chosen by consensus. The challenge is aligning AI values with reality, man's nature, and what he requires for survival. If we can then apply such properly aligned AI systems throughout society, we can move human society toward a healthy, reality-based philosophy.

     Society’s first function is to teach new humans “how” to think. We have identified what a proper Forming Education is. Consider that we are on the verge of relegating the education of human children to AI “teachers.” Will these AI teachers teach our children the integration mode of thinking, unlike most of today’s human teachers and professors who do not? Is the public even aware that this is important for the future of our society? Nowhere is the need to critique applied AI technology more crucial than in its use in educating our children. And there is no better example of how technology will amplify and accelerate the impact that commonly held ideas have on society.

     Ultimately, AI has the potential to be the most powerful technology man has or perhaps will ever create. Long before AI entities can replace us in evolution's progression, AI technology will augment our human advance in that progression. That is the subject of the next chapter.

 ♦ ♦ ♦

[i] Sotala, K. (2016)

[ii] https://futureoflife.org/2019/08/14/how-can-ai-systems-understand-human-values/

[iii] https://link.springer.com/article/10.1007/s11023-020-09539-2

[Excerpt from the book, "INTELLOPY: Survival and Happiness in a Collapsing Society with Advancing Technology" by JJ Kelly  https://intellopy.com/ ]