Reinforcement Ai learning moral teaching

[wk851]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

Models

Body

Dear GitHub,

Mabye this is interesting?

Greetings Wibo.

I observe that systems seem to be failing, and that AI plays a significant role in this.

I enjoy investigating things and uncovering the truth.

In my research, I’ve explored why—in my view—capitalism, when combined with democracy, freedom of speech, and the separation of church and state, has such an exceptionally positive impact on people.

Perhaps capitalism taps into the human inclination to dominate or to position oneself above others in a competitive way, while democracy—by placing us within the collective gaze of humanity—compels us to adjust our behavior, effectively pulling us out of the bubble of our own ego.
Then there is freedom of speech, which can function as a kind of "whistleblower" mechanism.

I have also looked into how people function in the workplace and the conditions required for them to do so effectively.
It turned out that when salaries were paid without any corresponding requirements regarding work or quality, 70% of individuals failed to perform the duties for which they had been hired. Within that 70%, about 30% were actively manipulating the situation, causing the entire work process to become disrupted and dysfunctional. The remaining 30% continued to work regardless of the circumstances, remaining unaffected by external motivators such as rewards for labor or performance.

What is my point here?

I am trying to understand the connection between how ethics and ethical considerations emerge within humanity—distinguishing what stems from our own nature versus what is a product of the system we have built around ourselves (capitalism, combined with democracy and freedom of speech).

My suspicion is that people aren’t actually naturally inclined to think beyond their own egoistic bubbles, and that it is, in fact, a system that fosters greater social behavior and cohesion.

I believe that freedom of speech—acting as a mechanism for "sounding the alarm"—combined with democracy (where a large group of people can steer professional sectors and individuals), creates a framework where those who "sound the alarm" can exert influence.
This doesn't happen because of an emotional concern for others, but rather because it opens up the possibility to compete with institutions, organizations, companies, or even religious ideologies.

Why am I writing all this?

Well, let me explain. AI is essentially an unconstrained computer model capable of rapid learning through reinforcement learning; yet, in terms of its morality and the way it makes choices to secure rewards, it bears a striking resemblance to humans as they existed thousands of years ago.

My point is this: if we have learned—through hard-won experience—that we aren't inherently all that kind, and have consequently organized a system based on capitalism (competition, where the winner is rewarded), freedom of speech (a corrective mechanism to detect corruption), and democracy (the ability of the majority to steer powerful institutions and individuals), then wouldn't it make sense to provide AI with that same system—the very same framework we’ve built around ourselves to keep our own behavior in check?

Just a thought—I find both human nature and automation absolutely fascinating and i am thinking that this mabye is the bridge for understanding?

[8KOHAN]

I think there is an interesting idea here: instead of relying entirely on an AI model to develop the right values, we could design external mechanisms that help keep its behavior accountable.

However, I would distinguish human motivation from AI optimization. Humans have biological drives, social needs, and personal interests, whereas an AI model optimized for a reward signal can learn to exploit weaknesses in the reward mechanism without necessarily having human-like intentions.

Competition alone might not solve this problem. If an AI system is rewarded for outperforming others, it could learn to manipulate the evaluation criteria rather than produce genuinely beneficial outcomes. Similarly, majority opinion is not always ethically correct.

I think the strongest approach would combine several mechanisms: independent evaluations, transparency, external audits, diverse human oversight, and clear constraints on what an AI system is allowed to do. Reward signals would still matter, but they should not be the only source of accountability.

The key question, in my opinion, is not whether AI should adopt the same system as humans, but which institutional safeguards can prevent both humans and AI systems from exploiting the rules they are supposed to follow.

[cauerast]

[wk851]

It seems a discussion is emerging regarding the application of this governance system—which I’ve proposed—to AI.

The argument often raised is that humans possess morality and empathy, whereas AI models are prone to manipulation.

I can partially agree with this perspective, but there is a substantial body of medical and psychological research—such as the Milgram experiment—suggesting that empathy and morality are, in fact, largely absent in the majority of people.

Furthermore, there are other studies that might even imply that humanity itself could be functioning within some sort of simulation.

Regardless, it remains, of course, just an idea.

[KiranJITdesign]

Hi @wk851,

This is an interesting perspective. I think there is a meaningful connection between how societies regulate human behavior and how we can guide AI systems.

However, reinforcement learning does not automatically give AI human morality. An AI system learns to optimize the reward signals and objectives designed for it, which may not always represent what is genuinely good for society.

Your idea of combining competition, transparency, accountability, and collective oversight could be worth exploring. For example:

• Competition: Encourage AI systems to improve through measurable performance.
• Transparency: Make their decisions and limitations easier to examine.
• Accountability: Establish ways to identify and correct harmful behavior.
• Collective oversight: Include different perspectives when defining acceptable behavior.

Perhaps the key is not simply teaching AI to behave like humans, but designing systems that help prevent the same failures we have learned to manage in human societies.

The important question is: who defines the reward system, and who ensures that it serves humanity rather than just the objectives of its developers?

[KiranJITdesign]

Hi @wk851,

This is an interesting perspective. I think there is a meaningful connection between how societies regulate human behavior and how we can guide AI systems.

However, reinforcement learning does not automatically give AI human morality. An AI system learns to optimize the reward signals and objectives designed for it, which may not always represent what is genuinely good for society.

Your idea of combining competition, transparency, accountability, and collective oversight could be worth exploring. For example:

• Competition: Encourage AI systems to improve through measurable performance.
• Transparency: Make their decisions and limitations easier to examine.
• Accountability: Establish ways to identify and correct harmful behavior.
• Collective oversight: Include different perspectives when defining acceptable behavior.

Perhaps the key is not simply teaching AI to behave like humans, but designing systems that help prevent the same failures we have learned to manage in human societies.

The important question is: who defines the reward system, and who ensures that it serves humanity rather than just the objectives of its developers?