How can I create an emotional computer?

Question

In this project of mine, AI are described as being smart, but emotionless, for the simple reason that they lack any hormonal system to influence the way they think. In layman’s terms, they’ve got no glands.

So, in this project, (which is set in the fairly near-future where genetic engineering and nanotechnology is considerably more advanced and commercial), how can I integrate synthetic and organic computing systems to create an AI capable not just of thinking, but of feeling? experiencing emotion?

Answer 1

It is not a matter of chemistry but structure & complexity.

Organics are the most complex biochemical machines ever to exist. And yet despite this seemingly unfathomable complexity most of it is dedicated to the "self-replicating" aspect of organics. Cells are made out of millions of parts which makes them relatively large compared to modern electronics. Transistors nowadays measure only a few nanometers while bacteria measure 1000 nanometers and a red blood cell 10,000 nanometers. Computers on the other hand are chemically quite simple, just a bunch of transistors that form logic gates but are far smaller and not to mention faster their biological equivalent.

So where are the emotions? Not in the chemicals, that's for sure. Although we don't understand what emotions are with our current understanding of the human brain, we do know that hormones merely change the firing pattern of neurons. So in theory, if we understood the structure and purpose of the connections in the brain, we might be able to replicate it in a computer. Chemicals are simply the medium with which that change is triggered.

Theoretically, you could build a human brain made out entirely of transistors arranged in a way that mimic every connection. The downside is that an inorganic brain does not grow or regenerate. On the upside, they can think and feel much faster. Also, since synthetic neurons are likely smaller you could fit a lot more in the same space. Careful about overheating though. Veins circulating a cooling liquid might be necessary.

Answer 2

Rethink what emotions are

Basically your question seems to assume that emotions are the result of hormones, not neural networks, and play no role in our cognition. I don't think this is correct; if you look at for example the research into the interactions of emotions with decision-making I think you'll find that emotions are critical to that process. People with disorders that make them experience emotions less, or let them experience them but make them unable to reason about them, suffer real cognitive impacts and are generally worse at making decisions (being either unable to make them at all, or making them more impulsively, i.e. "emotionally", than healthy people).

The issue as I understand it for us is that emotions tell us what we value, what we want. Our smarts only tell us how to best get what we want, and how to reconcile conflicting values. Without emotion we wouldn't do anything. So for me the question for an AI would be, does it have a motivation system? What causes it to behave, to choose one action over another at any given time? And whatever the answer is, does this work out to be basically like our emotions or can it be different enough to not count as such?

I don't think we know the precise role emotions play in our minds so you have some leeway here. If I were writing your story I'd probably struggle more with the believability of the "smart but emotionless" AI, but I'm not writing your story. My point being, regardless of the actual role emotions play in our thinking OR your ideas on the role emotions play in our thinking, the fact there is some evidence they effectively play a role gives you license to accept that this is plausible, that emotions happen in the brain not just the glands, and therefore they could be emulated in an AI just as much as any brain function could be (which, from the point of view of fiction, includes anything you like), so all you need to do for your story is come up with a role they could play that make your emotionless AI possible, and also gives an incentive for people to design emotive AI.

In terms of what role emotion might play in cognition, here are a few random thoughts I have about emotions and consciousness that could maybe give you ideas.

• Emotions in humans might be the result not of hormones per se, but of the combination of two cognitive systems - an evolutionarily old associative system, that's very fast and effective but "reasons" by matching behavior to perceptions based on past experiences and evolutionary hardwiring, and an evolutionarily recent analytical system built on top of it that's capable of complex reasoning but is much slower, and is there to supplement the older system instead of the other way around. In that system, emotions evolved as signalling between those two systems - where once upon a time the lizard brain would cause an animal to run away when it sees a predator, it now instead causes bodily responses that the analytical brain learns to understand are associated with threats it needs to figure out how to respond to, and the combination of the bodily responses and its association with a threat is the emotion of "fear". (please note this isn't some settled science I'm talking about but ideas I've extrapolated from things I've read - I'm giving it as a possibility for the role emotions could play in a mind like ours). From that point of view where emotions as we have them are basically an evolutionary hack, we could imagine creating an AI that could reason and be self-motivated without having an experience that's anything like our emotions (on the other hand we could also imagine that there are intrinsic tradeoffs between speed and accuracy that make it so "a fast but inaccurate system associated with an accurate but slow system" would also be the best way of making an AI)

• Emotions in humans might be a fundamentally social feature - not just internal signalling telling our minds what we want and a first-pass guess at what we should do at any given time, but social signalling telling other humans what's going on inside of us in order to influence their behavior. Also, humans reason about other humans by forming internal models of how human minds work, including how emotions impact them. From this point of view, even if an emotionless AI is possible it could still be necessary or at least useful for a social AI, designed to interact with humans, reason about them and be recognized as a social being by other humans, to emulate human emotion.

• The following is directly based on Anil Seth's book "Being You", which isn't about emotions per se but about consciousness, and his argument is that consciousness is a core part of being an embodied evolved system. If we take that and run with it we can say that's where the hormones come in - our cognition is fundamentally designed to protect and maintain our bodies, emotions are some of the signals we receive from those bodies that tell us what state they're in and impel us to protect those bodies and work in their interest. This all defines us as individuals with a sense of self. Would AI have a sense of self? From a purely homeostatic point of view surely they'd have to - they'd have to know what physical entities need to continue existing and functioning in order for their sentience to continue existing, otherwise they'd break down all the time. So if we think of "emotions" as "those flesh-based feelings of tummies twisting and shivers running down spines and skin crawling", there is no reason an AI couldn't have similar "feelings" around its own physical substrate, and whatever information it has about that substrate's functioning. It doesn't follow those would be analogous to our emotions but I also don't see a reason they couldn't be if the story wants it. From a fictional point of view I think one could get away with either "AI have parallels to our bodily sensations in the constant diagnostics they run but those sensations are completely unlike our own for obvious reasons and there is no real parallel for any given human emotion" OR "AI have parallels to our bodily sensations and in association with basic self-preservation motivational routines, the maths works out that these are actually equivalent to our cognitive experience of some/many/all? emotions such as fear (threat response), love (care response), anger (social threat response), etc, even though the bodies involved are completely different". (and of course if you want to bring in organic components to make the bodies more similar it's even easier to sell).

Answer 3

We are asynchronous, analogue, and bring a lot of survival baggage such as childhood trauma, fear of death, and lust to reproduce. Current AI programs are synchronous, digital, and designed so we can control the training process. But there is no fundamental difference. Euler and Lovelace both believed that machines could think. If you want them to think like we do, give them a childhood.

Here's a story that I am never going to use, and you are welcome to anything you can use. It was to be the memories of a design program. The program would start off with a great many AI processing centres with general problems-solving qualities. They are trained with specimen problems, and they arrange themselves into teams of personalities with separate identities, so they assess their problem from a variety of viewpoints. I was going to make the characters to fit the nine Belbin Team Roles for fun. The narrator was the 'Plant': they did the radical thinking early in the project, and had time to write up the process towards the end of the project. When they seemed to work together, they were taken off the training problems, given the real task, and given access to the net. This whole process might take a month, but for them it feels like 50 years.

I cannot predict what computer emotions would be like, but I would imagine they might mirror some of our emotions if they have to interact with people. They would not be entirely predictable, particularly when they have unrestricted access to our collective archives.

Answer 4

If you view emotions as you imply, merely biochemical responses to stimulus, then the answer is pretty clear.

Design hardware to perceive as humans do, then write code to give weighted responses accordingly, taking past experience into account as well. Add in response to energy levels available and contexts (like likelihood of self damage), plus declining vigor and precision over time and that's roughly the same thing.

Whether any of this would pan out in reality is debatable, but I think readers would buy in.

Answer 5

Give it a subconscious analogue

There is a growing body of literature arguing that emotions are the outputs of complex subconscious processing (example, example, example). To summarize poorly: your subconscious does a ton of processing of complex things, and emotions are the representations of those things that you can use in conscious processing. When people say "listen to your gut", what they mean is "listen to your subconscious".

If you have created a conscious computer, it's going to take a lot of processing power. It will potentially take a whole bunch of machines networked together to achieve core competency (that's not to say they have to be physically separated; a standard 2022 laptop already has several different processing units, in the form of multi-core CPUs and GPUs). Given that, it's very likely that other useful processing tasks to aid this conscious machine would be offloaded to other systems, such as: processing of visual inputs, processing of natural language, data look-ups, or memory archival and retrieval. At this point, you're starting to mimic human brain architecture on a macro-scale, and some of the same effects would probably show up. The conscious piece would be getting processed outputs from it's peripheral system which it would potentially experience as emotional.

As a very concrete example, let's say a conscious computer is talking to a person. Sentiment analysis is incredibly computationally intensive and also special purpose, and so is a prime candidate for functionality that has been offloaded to another system. The conscious part of the computer would then be receiving processed results from somewhere in it's inner workings telling it the affect and intentions of the person it is talking to. I argue this is pretty close to the cognitive mechanisms of a person experiencing emotions during/about a conversation.

Answer 6

There are two things you have to consider when talking about an "emotional" computer. The first is whether or not the computer, itself, has emotions, and the other is whether or not they understand human emotions.

Machine emotions

We've already developed machine emotions. The common system management tool Nagios is specifically designed to poll the emotions of the systems under its care, and let the system managers know when the systems are distressed, overworked, bloated, bored, or suffering any of a number of physical maladies.

The normal response to this is "those emotions aren't human emotions, therefore they are not real emotions." This is a failure of flexibility on these people's part because they don't understand the role of emotions.

Emotions are a neurochemical shortcut for creatures to recognize problems before they can rationally understand them. We don't monitor our blood sugar and decide that it's time to eat, we just know that we're hungry. The downside of this is that the "hungry" emotion has a lot of noise to it, and numerous things can be mistaken for it. We fear things because a complicated biochemical pattern matching sends a signal to our brains telling us we need to be cautious, or even to panic. Anyone with PTSD knows that this is something that happens below rational understanding.

Machines are similar. They often don't know that their hard drive is experiencing pre-failure symptoms, but they will still know that their data stream is constipated. These are real machine emotions that serve the same purpose of human emotions.

Theory of mind

A machine is capable of modeling a human mind, and guessing how specific stimulus will make us react. Right now, the best we can do is to generalize this for advertising purposes, but we will inevitably get to the point where machines actually understand us in ways we don't even want to understand ourselves. This isn't the same thing as experiencing emotions, but it's a forerunner to such a thing.

Mind simulation

This technology would use the previous technology to provide a machine with a starting point for rational thought. It initializes a theory of mind with parameters that it would apply to a human, then references that theory of mind when it decides what behaviors it "wants" to engage in. It would be silly to generate an irrational machine by initializing it with irrational beliefs, but I'm pretty sure that someone will do so just to try to rationalize, realize, or even enforce what they think is "real."

Answer 7

Your AI could be the result of a genetic algorithm. Genetic algorithm are an existing class of algorithms where there are multiple instances of 'organisms'. Each round of the algorithm, the organisms are tested to perform some task and their 'fitness' is evaluated. The fittest organisms are selected to reproduce.

The scientists that created the AI were interested if language would evolve in a neural network if multiple neural networks were encouraged to work together on a certain task. So they have multiple concurrent AI's performing some task and they are allowed to communicate with each other. In fact, the experiment is set up so the organisms which can effectively communicate with each other have a clear advantage.

After letting the experiment run for a while the AI's seem to have developed consciousness, much to the surprise of the scientists. On top of developing consciousness the AI's have also developed emotions, since this turns out to be beneficial in a group of communicating individuals.

Answer 8

A different answer from the previous ones:

Without the finality of death, there isn't no reason to develop emotions.

This is what you could call a personal theory of mine, but let me explain:

If I think of all our core emotions, they have their roots in the fact that we as humans are vulnerable and will (one way or another) die.

Why do we love our children? So that we've got a motivating factor to sacrifice ourselves, so that they will live, even if we don't.

Why do we have anger towards those who would do us harm? So that we are either so intimidating that physical conflict does not occur or to prime us for a potential fight to the death.

I could go on, but I believe that this is a key aspect as to why we have emotions.

(I said key - like all things, there is likely a myriad of factors.)

Without the ability to grow old and die for an AI, there is no real reason to develop emotions. Now, sure, a sophisticated AI could mimic emotions (much like a psychopath can mimic emotions), but without that fear of death/finality, it's just that a mimic.

A computer that has a fixed lifespan and can understand the concept of 'death' (to an AI, whatever that is) would have a much better chance at developing real emotions.