Asking an AI how to teach an AI 'ethical and/or moral behavior'

Note to the Reader: Back in 2014 or so, following my reading of an article in American Scientist (by Brian Hayes) concerning 'new strategies for the Prisoner's Dilemma' -- a strategic cooperation/defection thought experiment derived from Game Theory -- it occurred to me that some of these 'new strategies' (e.g., 'super-cooperators') were rather 'old' and rather congruent with the biblical Golden Rule (aka 'The Law of Reciprocity', or, do unto other as you would have them do unto you ; and variants). Over the next several years, I began to develop this idea incrementally and I endeavored to merge the two ethical-moral 'strategies' (i.e., the Iterated Prisoner's Dilemma [IPD] game and The Golden Rule) in the context of 'teaching an AI to be kind', that is, teaching an Artificial Intelligence (embodied in a robot, let's say) ethical (or 'moral') behavior towards human beings. This idea would eventually take the form of a lengthy essay (an earlier draft of which was published here on LI for a time, before I removed it ahead of my last submission) which was received with 'mixed reviews' by at least four critics (from two different publication venues/platforms, one draft submitted in 2017/18, and another in 2023). These were mostly negative criticisms (but with one critic giving me positive feedback and criticism). Having had one mixed review already, I decided to 'test out' my central concept before submitting it (to the second would-be publisher's reviewers) by 'employing' chatGPT and giving it only the minimum necessary prompts (i.e., my summary description of the concept). The following is the (unedited) result. I was pleasantly surprised to see that -- with a few 'broad brush strokes' and simplified/generalized additions -- the chatbot output largely reflected my original conceptual vision. The 'nuts and bolts' of this didactic vision (the 'training' of the AI) were mostly laid out in my essay, though clumsily edited. I will leave any additional critiques to you, my dear readers. -- M. A. R.

A ChatGPT Response to an ‘Ethical AI’ Prompt (2023)

PROMPT (by M. A. Ricciardi):

Using the Iterated Prisoner's Dilemma from Game Theory, and combining it with The Golden Rule (or 'Law of Reciprocity'), show how these two ideas can be used to teach an AI how to behave ethically or morally.

ChatGPT OUTPUT [Note: the following chat bot output has not been edited in anyway]):

Combining the Iterated Prisoner's Dilemma from game theory with the Golden Rule (Law of Reciprocity) can be a valuable approach to teach an AI how to behave ethically or morally. The Iterated Prisoner's Dilemma provides a framework for making decisions in repeated interactions, while the Golden Rule emphasizes treating others as you would like to be treated. Here's how you can integrate these concepts to teach ethical behavior to an AI:

1.

Define the Iterated Prisoner's Dilemma:

o In the Iterated Prisoner's Dilemma, two players repeatedly engage in a game where they can choose to cooperate or defect.

o Each player faces a choice: cooperate (C) or defect (D).

o The payoffs are set up such that mutual cooperation leads to a higher reward than mutual defection, but one player can exploit the other if they consistently choose to defect.

2.

Implement the Golden Rule:

o Introduce the concept of the Golden Rule: "Treat others as you would like to be treated." This principle encourages cooperation and mutual benefit.

3.

Reward Structure:

o Set up a reward structure that reflects the principles of the Golden Rule. Players are rewarded for mutual cooperation and penalized for mutual defection. This reflects the idea that ethical behavior benefits both parties.

4.

Repeated Interactions:

o Encourage the AI to engage in repeated interactions with other agents or entities. This mimics real-world scenarios where ethical behavior is essential for maintaining relationships and achieving long-term benefits.

5.

Learning Algorithm:

o Employ a learning algorithm, such as reinforcement learning, to allow the AI to adapt its strategy over time.

o Use rewards and penalties based on the outcomes of interactions to guide the AI's decision-making process.

6.

Reciprocity:

o Integrate the concept of reciprocity into the AI's decision-making process. When the AI encounters other agents, it should consider not only its own interests but also the interests of the other party.

o The AI should make decisions that maximize the collective benefit, even if it means making short-term sacrifices.

7.

Continuous Learning:

o Enable the AI to continuously learn and update its strategies based on the outcomes of previous interactions.

o Over time, the AI should become better at recognizing when cooperation is beneficial and when it's necessary to stand firm against exploitation.

8.

Feedback Mechanism:

o Implement a feedback mechanism that allows the AI to assess the ethical implications of its decisions.

o Encourage the AI to self-reflect on its behavior and adapt its strategies to align with ethical principles.

By combining the Iterated Prisoner's Dilemma with the Golden Rule and using reinforcement learning, an AI can learn to behave ethically or morally in various situations. Over time, it should develop a strategy that prioritizes cooperation, reciprocity, and the well-being of all parties involved, reflecting the principles of ethical behavior and the Golden Rule.