Stages of Artificial intelligence's mental development to full personhood

Question

First post here so please go easy on me. My knowledge of software and programming is limited.

Question: What kind of progression/stages would we see if an AI started as a highly advanced computer program that would naturally increase in complexity, slowly becoming self aware and capable of independent thoughts and feelings. Keep in mind they are specifically designed to go through this and the chances of homicidal, haywire or rebellious AI is essentially zero, these life forms are very commonly used by those who can pay.

I am aware that computing time allows for basically instant growth and development. In this case the actual work done is within the crystal structure which start very small and is not as efficient as actual computers but is where "they" exist. The development of the consciousness is directly tied to the growth of the crystal. While I am for hard sci-fi this is a softer plot driven part of it.

Context: I run TTRPGs and am trying out writing for the first time. It's a (relatively) hard sci-fi setting with intelligent, self aware crystalline creatures. They are motionless and need to be integrated into computers and machinery to do anything. They sell their "children" as bug proof, hard working, pacifistic systems for personal, ship, or station management. Over time and as people help grow the crystal it becomes more cognizant and thus able to better manage things with less and less user intervention. Essentially from the start is begins taking over all system related work, maintenance, and management. If allowed to grow large enough it will eventually become a fully fledged individual able to handle massive scale projects with it's own unique insight and personality.

What I am looking for:

1. Comparable milestones to a human infants mental development
2. How the crew and users input and saved data might help mental development in comparison to a human child's interaction with family.
3. Ways or examples of partial self aware AI trying to express its want's and opinions through (safely) messing with the systems it controls much like a baby that cannot speak yet.
4. Anything else I may not have foreseen.

Answer 1

Personally, I like the 7-Level AI model. Growth can be described as ascending in levels. Your crystalline creatures start at level 1, and with time, experience, manual improvement, etc, they can rise up through the levels (usually) eventually capping out at level 4 or 5.

Here are the levels

1) Rule-based: These types of AI are simple programming written by humans (or sapients generally). They can be very complex, but are basically just big collections of if-then statements. Software engineers can easily create, modify, troubleshoot, and understand these. This is where you have aircraft autopilots, self-checkout machines, smart TVs, etc.

2) Context-aware: Also sometimes called "Expert Systems" these are AIs that have rudimentary learning capabilities and can improve with data. They are extremely competent in a very narrow application and can deal with novel scenarios unlike rule-based AI (so long as the problem is within their domain). Often, this is where "neural nets" start being used by software developers which can then be "trained" to accomplish a specific task. Unlike in a rule-based system, the exact way an expert-system reaches a conclusion is inscrutable to even the smartest human. Examples of this include things like person-recognition security cameras, modern smart vacuum cleaners, customer service chat-bots, or medical image analysis systems.

3) Domain-specific: This is the level where the AI can start reliably passing for (pretending at) human-level intelligence and even exceeding it so long as it stays in the specific domain that it's been trained/designed for. A "Level 3" chatbot can, for example, pass for completely human in a Turing test. This is where the cutting edge of today's technology is and includes things like self-driving cars, supercomputers playing GO, or advanced medical diagnostics AI. This level of AI completely fails at tasks outside of its domain.

Speculation Border! (levels below this are speculated)

4) Reasoning Systems: Human-like general intelligence that can use some sort of reasoning process to solve novel problems. Arguably has emotions and other things that we associate with being sapient, however Level 4 AI is still missing several key components like the ability to creatively problem-solve (as opposed to a systematic problem-solving approach), self-developed initiative (needs to be told what to do), and is still missing some of the more abstract elements that we associate with being human. An AI of this level would be able to perform most jobs we have today very well, and even optimize them, but never propose systematic changes. For example, they could run a fast-food restaurant with extreme efficiency but never decide to build an expansion or create a new menu-item without external input. Their self-awareness is debatable.

5) True AGI (Aritifical General Intelligence: Self-aware and human-like AI. Can do everything a human can do including learning, being creative, philosophizing, etc.

6) Superhuman AGI (ASI): Basically a better AGI that can beat human experts in all domains. Can solve complex issues and execute large projects that would otherwise require multiple humans to manage. In sci-fi, this is where most Ship, facility, or City-level AIs are located.

7) Singularity: An intelligence that is capable of improving itself at a super-linear rate (the rate of self-improvement gets faster as time progresses). Here be dragons and Gods.

EDIT:

Should be noted that AI won't ascend through most of these levels without (likely invasive and intensive) external help. A rule-based system will never reach level 2 (the smart TV will never develop new menu items or adapt to a new video codec without an update), even if it operates for thousands of millions of years and I find it unlikely that a level 3 could ascend to level 4 or beyond without being granted the missing, more abstract, elements.

It is conceivable that a very well-designed Level 2 system eventually becomes skilled enough to reach level 3 as they can learn based on experience, and the ascension from 5 to 6 could also be possible as it is in large part a question of scope.

Answer 2

We don't know.

Since we cannot currently build AIs that are convincingly persons, and we're a long way from being able to do so, it is not currently possible to say what the stages in their development might be. There have been some attempts to portray the process in fiction, but those are based on the inventiveness of writers, and analogies with the mental development of children.

Flowers for Algernon is a famous example, and worth reading.

Answer 3

I agree with John Dallmans answer: We don't know. But a lot of people with very different backgrounds have thought about this and related topics so I wanted to throw around some ideas and some existing knowledge and research. Maybe you can use something for your story.

First, humans mostly agree that most adult humans are self aware. Philosophers have pondered on how one would know this. The philosophical zombie and the Chinese room are interesting starting points on this.

Some animals are somewhat self aware. A classic experiment is to just hand a mirror to an animal. Dogs will just see another dog. Chimps realize they see themselves. Elephants are in between, the clever ones get it, the dumb ones don't. This already indicates that self awareness is not a black-and-white thing but something more gradual.

Human infants slowly develop more and more understanding of the world around them eventually leading to self awareness. Piaget stages are one way to mark a number of milestones.

Answer 4

The short answer isnt "we wouldnt know" but "each AI develops differently and there is no 100% perfect way to check what they learned*"

*aside from maybe using another AI with the closest to human values that checks each AI's response with a trillion scenario's

Long anser:

An AI starts as a program that we feed information, this information teaches the AI something, and we move on. The sum of all this information and how the AI writes its own programming to deal with the information is what makes it an AI. That means that if you have perfectly identical programming and give the AI the exact same information but in a different order they will be different AI's. In fact, due to our storage capabilities being far less reliable than we think the copied AI's with the exact same information would develop differently.

This makes AI programming opaque. You dont know what the AI has actually learned. A good example is an AI that works absolutely perfectly. Then a natural disaster happens and the AI is tasked with supervising it, bringing in goods, choosing where to deliver aid first, make life and death decisions. And your AI does this with flying colors, no problem right?

But because the programming is opaque and everything learned adds to the totality of the programming you dont see where else this will affect the AI. A few months later you find all the homes for the elderly bulldozed and replaced with things for younger people, as the AI learned to make life and death decisions based on quality of life. The biggest problem with AI's isnt that they'll go rogue, but that they will simply learn to serve us in a way that destroys us. And the above example is not guaranteed, the AI may realize that some things it learns are controversial or oppose one another, but make the correct decision anyway. We will never completely know what an AI will do in a given situation, even if its a situation the AI has already been in since the AI will learn and adapt constantly.

Answer 5

Human mental growth and AI develop along different lines. If you often hear people doing such comparisons it is because there is a tendency to over-over-oversimplify. Note that a lot of tools dubbed AI in use today are very good at doing a very specialised task, but are extremely dumb at everything else. Advanced from a point of view, primitive from all the other points of view.

You will need completely different stages, and it does not stop there, there is not a single line of evolution, you can have different types of AI evolving in a different way depending on the group of scientists or school that is working on it.

I can make an example of possible stages, but that is only one possibility:

1. Understanding categories and doing correct classification.

2. • Understanding relationships between categories. Understanding different representations of the same object (image, name as a text, name as a sound).

   • Understanding classification when the representation changes. e.g. Recognising a face seen as a profile after learning the front. e.g. 2 Understanding that sentences in different languages describe the same thing.

   • Understanding classification when the object changes with time.

3. Understanding cause effect relationship.

4. • Understanding contexts. e.g. why the picture of a person standing mid-air is strange. e.g. 2 why a spider web cannot catch a 5 kg stone.
   • Making inferences.

5. Planning.

6. Reading some text with the instruction to perform a task and doing it successfully.

7. And so on...

The steps I wrote do not lead to an AI as advanced as an adult human, but going on with further points would be useless. Not even with the crystal ball we could predict what could happen further because we haven't even completed step 1.

Answer 6

A basic question concerning your subject would be, why should machines develop awareness / self-awareness. All the positive attributes associated with the matter of awareness emerge from the anthropocentric understanding that evolution must go this way. If you actually consult state of the art awareness (as it has been institutionalized by Jon Kabat-Zinn) you will recognize that the central element humans want machines to perform is abandoned here: to discern and categorize. In fact being completely aware from that point of you would not easily be associated with awareness. That is because a crucial element of a general human awareness becomes eroded: the ego. As the ego expresses itself in volition among other things a completely aware person would hypotheticall also lack that. Again, it es the anthropocentric view that makes the matter complicated and actually science uses methods to get around this limiting factor. But as for us, I would point out that machines only need infrastructure and energy. They are not in need as humans are and not craving for anything. And they don't need a notion of self and thus don't need to accumulate anything material or immaterial. So, after all they did not undergo the bloody trail and error of biological evolution. This actually makes me think that they would not have the urge to kill anything. Eventually the would not care about live so much and have their own existence. In a way just like they have at the moment - may be because there is no need for awareness. A good read for you is definitely "Golem XIV" by Stanislaw Lem, who by the way is an author who constantly struggles with anthropomorphism in his oevre.