Value Specification. What specific values and whose values should be embedded in AI? Human values are extremely difficult to formalize.
Regarding AI helping humanity resolve its moral development, we can count on the expectation that AI will guarantee us greater objectivity and, by its very nature, linked to physical phenomena, free us from biases not only cultural but also those inherent to our nature.
We all know that human behavior is determined by the evolutionary requirements of Homo sapiens as a prehistoric hunter-gatherer. AI should also know this, since it has the capacity to observe us in our biological context.
As a “cultural animal,” Homo sapiens has the ability to control its instincts, repressing antisocial ones (aggression, tribalism, irrationalism, etc.) and promoting, as superstimuli, prosocial ones (empathy, caring, altruism, etc.). This objective vision seems more within reach of AI than of today’s civilized human beings, inevitably burdened by prejudices.
The post is not saying that AI will help humanity in its ethical evolution. Quite the contrary — first, we need to teach ethics to AI.
When I write about the need for ethical education and development of people, I mean a process in which people, by becoming familiar with various ethical views, dilemmas, and situations, become wiser themselves and capable of collectively teaching AI less biased and more robust values.
I understand that you assume that AI is already or will become morally objective in the future. But this is not the case. Modern AI does not think like humans; it only imitates the thinking process, being complex pattern recognition systems. Therefore, they are far from objective.
Current AI is trained on huge amounts of our texts: books, articles, posts, and comments. It absorbs everything in this data. If the data contains prejudice, hatred, or logical errors, AI will learn them too, perceiving them as the “objective” norm. A striking example: if a model is trained on racist texts, it will reproduce and even rationalize racism in its responses, defending this point of view as if it were an indisputable truth.
This example shows how critically important it is who trains AI and on what data. Moreover, all people have different values, and it is not at all obvious to AI which of them are “better” or “more correct.” This is the essence of one of the key philosophical problems — Value Specification.
I agree that the ideal AI you describe could indeed help humanity in its ethical evolution in the future. But before that happens, we still need to create such an AI without straying from ethical principles. And this post is dedicated to how we can do that — through ethical co-evolution, rather than passively waiting for a savior.
if a model is trained on racist texts, it will reproduce and even rationalize racism in its responses, defending this point of view as if it were an indisputable truth.
But this isn’t a very brilliant intelligence. If a human being with a high IQ is raised in an irrationalist environment (racism, theism, tribalism, sexism, etc.), they usually have the capacity for logical rationality to understand both the inconsistency of such approaches… and the psychological and cultural origins that gave rise to them.
It’s assumed that AI will far surpass human intelligence, so they must be perfectly familiar with the biases, heuristics, and contradictions of human reason… just as human ethologists understand animal behavior.
Regarding AI helping humanity resolve its moral development, we can count on the expectation that AI will guarantee us greater objectivity and, by its very nature, linked to physical phenomena, free us from biases not only cultural but also those inherent to our nature.
We all know that human behavior is determined by the evolutionary requirements of Homo sapiens as a prehistoric hunter-gatherer. AI should also know this, since it has the capacity to observe us in our biological context.
As a “cultural animal,” Homo sapiens has the ability to control its instincts, repressing antisocial ones (aggression, tribalism, irrationalism, etc.) and promoting, as superstimuli, prosocial ones (empathy, caring, altruism, etc.). This objective vision seems more within reach of AI than of today’s civilized human beings, inevitably burdened by prejudices.
How will an AI that is built, trained, and curated by humans along every step of it’s development end up being “objective”?
The post is not saying that AI will help humanity in its ethical evolution. Quite the contrary — first, we need to teach ethics to AI.
When I write about the need for ethical education and development of people, I mean a process in which people, by becoming familiar with various ethical views, dilemmas, and situations, become wiser themselves and capable of collectively teaching AI less biased and more robust values.
I understand that you assume that AI is already or will become morally objective in the future. But this is not the case. Modern AI does not think like humans; it only imitates the thinking process, being complex pattern recognition systems. Therefore, they are far from objective.
Current AI is trained on huge amounts of our texts: books, articles, posts, and comments. It absorbs everything in this data. If the data contains prejudice, hatred, or logical errors, AI will learn them too, perceiving them as the “objective” norm. A striking example: if a model is trained on racist texts, it will reproduce and even rationalize racism in its responses, defending this point of view as if it were an indisputable truth.
This example shows how critically important it is who trains AI and on what data. Moreover, all people have different values, and it is not at all obvious to AI which of them are “better” or “more correct.” This is the essence of one of the key philosophical problems — Value Specification.
I agree that the ideal AI you describe could indeed help humanity in its ethical evolution in the future. But before that happens, we still need to create such an AI without straying from ethical principles. And this post is dedicated to how we can do that — through ethical co-evolution, rather than passively waiting for a savior.
But this isn’t a very brilliant intelligence. If a human being with a high IQ is raised in an irrationalist environment (racism, theism, tribalism, sexism, etc.), they usually have the capacity for logical rationality to understand both the inconsistency of such approaches… and the psychological and cultural origins that gave rise to them.
It’s assumed that AI will far surpass human intelligence, so they must be perfectly familiar with the biases, heuristics, and contradictions of human reason… just as human ethologists understand animal behavior.
Do you believe that AI will be able to become superintelligent and then superethical and moral on its own, without any targeted efforts on our part?