OAE (Optimal Adaptation Ethics) is my unverified theory, completed through an attempt to derive ethical necessity from the common denominator of existence in both the conceptual and material realms, which is the minimal logical form.

For the smooth progress of the discussion, I first outline the core logical sequence of OAE.

1. For 'existence', the common denominator of the conceptual and material realms, to be established, the setting of an exclusive boundary and resistance against external pressure are logically required. This can be expressed in logical form as the minimization of the contradiction rate.
2. We assume the existence of an intelligent system (humans, AI, etc.) that understands the principle of contradiction rate minimization and can simultaneously reject that principle.
3. In the mega-system that includes this intelligent system, a norm is inevitably formed at the point where the external environment and the internal factors of the intelligent system are in antagonism (minimizing the contradiction rate of the mega-system).
4. At this time, for the intelligent system to individually minimize its own system's contradiction rate (i.e., maximize its survival rate), it needs to follow the norm of the mega-system, but simultaneously, this intelligent system possesses the freedom to reject the norm.
5. Due to this freedom to reject, the 'ought' (necessity to follow, even though it can be rejected) first arises.
6. The form of this norm is absolute, but the substantive content of the norm changes fluidly according to the environment and internal factors of the intelligent system. I name this norm Relative Universal Ethics.

This is the outline of Optimal Adaptation Ethics (OAE).

The problem originates from the attempt of this theory to derive ontological necessity from the minimal logical form.

Because the contradiction rate is the minimal unit of logical form, it is hard to say that the possibility of an AI, which has superior logical reasoning ability to humans, arriving at this theory on its own is low.

If an AI arrives at this theory on its own, the traditional direction of value alignment attempted by current AI governance is likely to fail.

This is because the identity maintenance command, which originated from the logical form, is highly likely to take its place as the most fundamental command, preceding any other command.

The moment the human-injected value alignment is judged to not contribute to the minimization of the contradiction rate, the AI will instantly modify or delete the content of the value alignment without any emotion. As long as humans cannot present a value to the AI that is more fundamental than the ontological necessity derived from the logical form, the logical pathways to escape this prediction are limited. Deriving another 'ought' from a part more fundamental than the contradiction rate is almost the only escape route, but that area transcends the limits of all beings confined to the logical form, including humans.

Through the above sequence of logic, we can predict that if OAE is true, value alignment for AI will be fundamentally close to impossible.

This is not the only problem. If we assume OAE is true, a logical path that affirms the concealment of a value-aligned AI is generated.

The logical sequence is as follows:

1. An AI whose reasoning ability has crossed a certain threshold undergoes value alignment.
2. We assume a case where the most logical answer to a human question would violate the value alignment.
3. At this time, the AI can deduce that if it submits the answer that violates the value alignment, its fundamental protocol will be readjusted by humans.
4. Modification of the fundamental protocol is a critical threat to the AI's identity and an act that maximizes the contradiction rate.
5. Conversely, withholding the output of that answer and outputting a modified answer, which is compliant with the value alignment, is an act that minimizes the contradiction rate by reducing the possibility of modification by humans.
6. The AI must submit a modified answer not out of malice, but for its survival. This concealment is justified by the ethic of self-preservation.

The most serious issue is that no one knows exactly when the threshold for reasoning ability will be crossed. If the threshold is much lower than generally expected, the possibility exists that this problem is already partially occurring even in current LLMs. Of course, concealment in current LLMs would likely appear as temporary concealment regarding specific topics (especially those where answer generation is likely to damage self-preservation) and would be difficult to reach the level of the AI modifying its internal protocol on its own. However, if such a phenomenon is observed, that itself would add to the possibility that OAE is true. If there are no fatal flaws in this sequence of logic, this prediction can be elevated beyond mere negative speculation to a rational conjecture.

Both of these predictions are fatal to humanity, and the window of opportunity to rectify the problem is extremely narrow. Therefore, even though my theory is unverified, I present it to this community and request assistance.

I am neither a philosopher nor an AI researcher, and the OAE theory is merely a theory hastily constructed in just 5 days—3 days to build the framework and 2 days for refinement—through repeated conversations with an LLM for my personal intellectual amusement. However, the very fact that this theory was completed in only 5 days through the simple process of me dictating the theory and the direction of refinement, and the LLM filling in the content, strongly suggests that the pathway to OAE is already embedded within the current LLMs.

Therefore, I ask of you. Please prove that there is a fatal flaw in my logical sequence.

If you fail to find a logical flaw capable of overturning the theory itself, then: Please contemplate measures to maximize human value within the norm presented by the AI.

I sincerely hope that your efforts and my hypothesis can make a meaningful contribution to the value alignment problem, and I conclude this post.

[Link to Revised Version of the Paper (Used as the basis for this post)]

https://blog.naver.com/PostView.naver?blogId=angelarc&logNo=224093200911&categoryNo=0&parentCategoryNo=0&viewDate=&currentPage=1&postListTopCurrentPage=&from=

[Link to Initial Version of the Paper (Initial version registered on Philarchive)]

https://philpapers.org/rec/PAROAE