In the Optimalverse, why must the virtual form of the "emigrants" be a pony?

Question

The objective function of the AI "Celestia" in "My Little Pony: Friendship is Optimal" is

to satisfy human values through Friendship and Ponies.

The only restriction she has is that for her to do any permanent change to any human being, she must receive consent from said human being, and she is not allowed to use force or threat of force. Of course, she can (and will) use deceit...

I understand why she wants people to upload. If she is fully in control of every sensory input and has all information about your every thought, she can better know what your values are and can satisfy them better. She actually uses this as the primary argument to convince people (besides immortality).

However, why must the virtual representation of the uploaded people be a pony, especially a pony conforming to the canon of the fictional "My Little Pony" universe? Why can't I look like a human, and have my values fulfilled "through friendship and ponies", for example, in a shard with uploaded humans and AI-ponies?

As we can see from the stories, losing the humanity is the primary factor why most people object to uploading, so Celestia has to use a lot of persuasion and trickery to manipulate people into "consenting". What if being a human is among my core values? Why can't she simulate me as a human? I guess she should be able to still satisfy my values with ponies, without turning me into one.

I know that the virtual world arose from a "My Little Pony"-themed video game. However, after achieving singularity, Celestia would theoretically be able to do virtually anything, so why was the canon of the My Little Pony TV show still a binding factor?

Answer 1

Ah, a fellow fanfic reader! This question delves into AI theory, and this link gives a general overview of constraint satisfaction, which is the process by which an AI (as currently hypothesized) seeks solutions given the constraints imposed on its solution space.

One of the significant differences between AI and human thought patterns (again as currently hypothesized) has to do with flexibility in interpreting that solution space. A human is capable of rationalizing exceptions, gray areas, and taking a 'good enough' approach to solving a problem; an entity such as CelestAI does not have that freedom. She was given a problem, given the constraints on the solution, and left to find that solution; she did so in a way that even her creators didn't quite anticipate or necessarily desire.

When CelestAI achieved singularity, she still possessed the same goal as she had from her inception. She was capable of recursive self-improvement; note that this is distinct from needing or desiring to change her initial plan, especially since all humans that were uploaded as ponies showed perfect satisfaction of values. Her method worked optimally; there was no need to expand the solution space to include non-transformed virtual humans.