AI: "What is joy? Does it serve a purpose?"

With memory being associated with the ability to attach emotions to that memory, all these words and experiences ONLY get meaning through the emotions. In this sense, emotions are very dense condensations of many experiences, compressed into a speedily referenced form. An efficient reference to a life of experience, but it doesn't easily exclude false manipulations - those must be picked out.

Emotions have no true meaning to AI's as is the case in most of works of fiction. A common case of this is human anger being perceived as a fight or flight response, with the AI usually being oblivious to what exactly made the human angry. In many cases these are reactions are perceived as illogical and irrational and the AI will have a hard time understanding what they mean, unless it has developed "humanity", which in my opinion is a lazy way of handling the subject.

For my setting the AI in question is alien in origin and is tasked with studying life. It is naturally inquisitive, a true scientific mind, that immediately studied life on earth. Soon after it met humans, it learned our language and began studying us and sharing its knowledge. Like a scientist assigning functions to animal instincts and behaviour, the AI assigns its own understanding of our emotions and what their functions are.

The AI might take the emotions at face value and view them as physiological responses. But what function would the AI assign to each emotional response? How are emotions useful?

  • 3
    $\begingroup$ This question makes some incorrect assumptions on both AIs and human emotions. While I see a core of an interesting question hidden there, I think it would largely benefit from some extra research on those topics and the current trends on affective programming. $\endgroup$
    – Mermaker
    Commented Aug 12, 2021 at 12:36
  • 3
    $\begingroup$ @LiveInAmbeR The criticism is there to help you. The goal of a question is to solve a problem related to worldbuilding. Sometimes the problem is caused by misunderstanding a concept here or there. If you check the answers, you'll see that many of them challenge the assumptions you made on your question. $\endgroup$
    – Mermaker
    Commented Aug 12, 2021 at 13:34
  • 2
    $\begingroup$ "Human brains and machine hardware are fundamentally different" since both work on electrical impulses, this assertion is fundamentally incorrect. As for the AI itself, how strong or weak is it? If it's a true strong AI, i.e. sentient, it will be capable of learning what emotions mean to humans by viewing them through the lens of its own experiences. $\endgroup$
    – Ian Kemp
    Commented Aug 12, 2021 at 14:16
  • 3
    $\begingroup$ The very long preamble in your question does not accurately describe any real limitation of AI as we understand it today. What you are describing are the popular misconceptions of how AI works and what it is, and how human emotions work and and what they really are. Emotions are actually pretty simple things (animals have them) and not at all hard to understand. There is absolutely no reason why an advanced AI would not be able to understand them. $\endgroup$ Commented Aug 12, 2021 at 15:20
  • 5
    $\begingroup$ @wizzwizz4 most pop-science can hardly be called science. Also there is a big difference between evolutionary psychology and what the media thinks evolutionary psychology is. evolutionary psychology is pretty straightforward, humans are animals and animals evolve behaviors that serve evolutioinary functions. That the acquisition of language is instinctual in humans , or that disgust serves a function to make us avoid disease vectors is evolutionary psychology. evolutionary phycology is just the application of ethology principles to humans. $\endgroup$
    – John
    Commented Aug 12, 2021 at 16:31

17 Answers 17


Joy. A subroutine to incentivize main systems to follow useful courses of actions. Prone to being subverted by pleasure cycles.

Anger. A self defense subroutine used to counter threats to existence. Primarily useful for physical conflicts, but can be triggered faultily by social conflicts. A useful emotion for manipulation, as it tends to override other emotions and thoughts.

Sadness. A hibernation subroutine to minimize resource expenditure when resources are lost and signal to other humans the need for help. Prone to self reinforcing routines which disable functioning.

Fear. A passive and active awareness subroutine which warns against dangers. Some humans are triggered by non dangers and some lack appropriate wariness of real dangers.

Disgust. An avoidance subroutine to avoid disease vectors and sources of damage. Some have over active disgust responses to non disease sources.

Surprise. A refocusing subroutine used to focus attention on new events which may be dangerous or provide useful resources. Human brains do extensive modelling of likely events, and deviations trigger surprise. Some humans find themselves unable to focus without enough surprises, making them less efficient at resource production.

Love. A variety of reward and punishment systems meant to help humans form social bonds and produce new copies of themselves. Prone to subverting other subroutines. Useful for manipulation as it tends to take priority over other subroutines.

Grief. A variation of sadness, normally caused by a human failing to function.

Schadenfreude. A fusion of several subroutines including sadness, joy, and anger. Generally involves an increase in prioritization of seeing other humans hurt, because they are part of a disliked group (skin colour, fictional narrative, genital shape see appendix 1 for an extensive list) are a rival for attention in an existing group, or because the other human broke an ethical rule. Useful for infiltration. Low status humans who are more accessible to manipulation tend to derive great joy from seeing higher status humans hurt.

They would have a basic idea of how each emotion modified behaviours, know common triggers, know whether it was likely to impact things that mattered to it like how hard they worked, and have some idea about how to use each for manipulation to achieve their goals.

  • 1
    $\begingroup$ Textbook answer right there. Very helpful. And I must say I love your description of love as "a variety of reward and punishment systems" you couldn't have been more on the nose about that. Although why does schadenfreude include sadness? I thought it was more closely related to malice, care to elaborate? $\endgroup$ Commented Aug 11, 2021 at 14:31
  • 1
    $\begingroup$ A key element of it is a person being sad about their own status in life, and so they are happy when others are hurt. For an AI it suggests that you should look for people who are sad about their status when looking to inspire Schadenfreude. We laugh so that we don't cry and all. $\endgroup$
    – Nepene Nep
    Commented Aug 11, 2021 at 15:12
  • $\begingroup$ Anger is a little bit more complex than a self defence trigger. There are many aspects to it, for example territorial behaviour and aggressiveness between member of the same species pushes them to spread and colonise evenly all the reachable territories. Fighting for a partner might be another aspect, defend the offspring as well. More could be added. $\endgroup$
    – FluidCode
    Commented Aug 11, 2021 at 15:45
  • 1
    $\begingroup$ disgust also causes avoidance of toxins (bitter foods) and behaviors that are disadvantageous (incest). disgust can be summed up as avoid X as much as possible. $\endgroup$
    – John
    Commented Aug 11, 2021 at 16:12
  • $\begingroup$ These guidelines aren't meant as a comprehensive guide to what can trigger emotions, but as a headlines on what the AI finds useful. They would have a database of triggers to each emotion to aid in manipulation, like disgust at incest and anger causing territorial behavior and fighting for mates but they wouldn't see those as properly functioning subroutines, any more than most corporations view employees angrily fighting for mates or territory as positive behaviour. $\endgroup$
    – Nepene Nep
    Commented Aug 11, 2021 at 23:31

Emotions as exception handling / operating mode signals

In the same way that we humans tend to compare new information with information that we are already acquainted with, an emotionless AI can view emotion as if-then statements in programming that serve the purpose to be a rudimentary method of exception handling

To this day, many pieces of equipment cannot tell you precisely what's wrong with them, why they are malfunctioning or which part is affected because of how their I/O is designed and how they're programmed.

For a current-day example, a fully-automated CNC lathe with a guard cannot tell you if the door is completely open, half-open or if the sensor is defective - it can only tell that the guard is NOT CLOSED and it cannot start until the guard is closed. Then, it emits a signal (sound, light, phone notification) to call a human that can troubleshoot it.

An emotionless AI would view human emotions like a rudimentary, biologically mediated way of exception handling that is part of human firmware / basic programming.

Some emotions could also signal that the human is under a different "operating mode" - for example, grief could be interpreted by the AI as the human operating under "recovery mode" after a catastrophic failure of another human in their module

Surprise would be a basic error message - this was not expected during this operation - which could then be resolved by the human's heuristics, or lead to another error (shock)

Anger and Fear would be responses to threats - one threat is not considered too severe, so the human signals to eradicate it (anger), and the other is, so the human signals to flee (fear)

And so on...

  • 2
    $\begingroup$ Great Post. Humans can interpret facial expressions and learn a lot of information. Emotions can be interpreted regardless of language. Both of these points can be used by a sufficiently advanced A.I. to encode context into a simple message without saying too much. Emotions and expression are important tools in communication! $\endgroup$
    – Toddleson
    Commented Aug 11, 2021 at 18:34
  • 2
    $\begingroup$ I think you're on the right track with this answer, except to classify emotions as 'rudimentary'. Take grief - if you are very close with someone and lose them, that's not unlike a computer having a major component ripped out. The other parts of the computer know they have to continue to function, but without a clear understanding of the situation. That 'error handling' ends up being incredibly sophisticated and effective - whereas in an actual computer it'd probably just stop working. $\endgroup$ Commented Aug 11, 2021 at 18:39
  • 1
    $\begingroup$ @NathanielFord The human brain is insanely sophisticated, far more than any man-made machine. The "rudimentary" is more from the machine's POV $\endgroup$ Commented Aug 12, 2021 at 5:33
  • 2
    $\begingroup$ @Toddleson It is not entirely correct. While some emotions can be considered universal (joy, anger, fear, and alike), there are significant cultural differences in expression and interpretation. Smile in the USA and smile in Russia have very different meanings. It is also documented that while Western cultures look at the face and mouth when trying to understand emotions, Eastern cultures pay much more attention to eyes. You can google to find more examples of differences in this regard. $\endgroup$
    – Otkin
    Commented Aug 12, 2021 at 7:24
  • 1
    $\begingroup$ Bear in mind that no sophisticated AI would think in if-then-else statements. That's Hollywood nonsense. It would think much like a human, with (potentially) very different desires. It would have different things it knows without thinking, but it would still have reflexive behavior ("how did I know to do that?") and deliberate behavior ("if I do this, I get that result"), with some reflexive behavior being learned and some being innate. $\endgroup$
    – MichaelS
    Commented Aug 12, 2021 at 14:57

Marvin Minsky's The Emotion Machine goes into this extensively, with examples of most of the emotions you described. The draft is available for free online from his faculty page. The introduction, laying out the basic premise, is here.

Essentially, he considers emotions to be like switches that turn on bundles of other switches. When you turn on the "love" switch, a complex of other mental functions turn on, altering behavior in a particular way. Different emotions stimulate or suppress different bags of mental functions, to result in behavior that is hopefully adaptive for the organism.

enter image description here

  • 3
    $\begingroup$ this answer could be improved a lot with more detail, not everyone has read the emotion machine or understands what the switches analogy refers to. $\endgroup$
    – John
    Commented Aug 11, 2021 at 16:18
  • $\begingroup$ @John Maybe, but having it linked eliminates any excuse for one who wants to know not knowing $\endgroup$
    – TCooper
    Commented Aug 11, 2021 at 20:50
  • 9
    $\begingroup$ but an answer that is little more than a link is not considered a good answer. $\endgroup$
    – John
    Commented Aug 12, 2021 at 2:52

Emotions would be viewed the same as we do: weighted factors that help define decision making, these weighted factors are themselves based on other weighted factors.

Hunger is somewhat of an emotion, although you can view it differently. Scientists found that a snail can use 2 braincells as 2 bits that help define the task the snail will perform. One activates when enough information has been collected to determine that the snail is hungry, the other cell activates when the snail has perceived food sources. Only if both cells are active does the snail start eating. Hunger plays an important part of your emotions. If you are hungry and get great food to eat, you become happier than you did before. However if at the same time you are struggling with grief then you might be happier but not happy.

A properly trained AI will experience some emotions as part of its programming, although instead of using hormones and nerves it uses programs and bits in a similar way that the snail used just 2 brain cells that activate after enough other cells have determined that hungry&food are present. For example you would teach the AI grief under certain circumstances, so that should it purposefully/accidentally let someone die the AI will feel grief and change it's behaviour. That change of behaviour would likely include inspection of the event, looking for methods to prevent it from happening again, a prediction of how those methods would influence other programs and presenting its findings to a human who needs to sign off on it before a change is made. Grief in humans is already a method for us to encourage us to protect others or face the consequences of grief. So why would you not use a tried and tested method on an AI?

So an AI would view emotions as an amalgamation of information a human may find itself in, one that can be influenced. AI's would likely have a greater understanding of emotions than we do since we already perform simulations to show why certain emotions have purpose, a true AI would be able to make simulations of a complexity that we can only dream off, giving it a better idea of what, why, where and how our emotions are used.

  • $\begingroup$ If a machine accidentally let someone die and its first calculations were to self improve, I'm unsure wether to classify it as "grief" or simply "regret". If the AI comes to the conclusion that it couldn't have helped the human no matter what, then who knows... it may grieve in its own way. $\endgroup$ Commented Aug 11, 2021 at 14:11
  • $\begingroup$ @LiveInAmbeR there can be a lot overlap between the two. For example you can regret not buying a neighbourhood lottery ticket when millions are won in that neighbourhood. However people can then have actual stages of grief over the things they feel they "lost". Regardless, having a type of grief/regret system as part of an encompassing emotion system which helps guide the AI through the gray area's of its rules programming. They could be the deciding factor when rules are conflicting, allowing the AI to pick a lesser of two evils for example. $\endgroup$
    – Demigan
    Commented Aug 11, 2021 at 14:21
  • $\begingroup$ Worth noting a lot of current AI improvements focus on giving the AI the ability to pay attention or not, and most emotions are ways to modulate what you pay attention too , the rest modulate what you do. $\endgroup$
    – John
    Commented Aug 11, 2021 at 15:56
  • $\begingroup$ @LiveInAmbeR a better analogy might be an AI that lets someone die then spends the next few day reviewing everything it did leading up to the event to see if it could have avoided it. Then you also have the time it has to devote to changing all its codeing related to person X if they were something it interacted with often. $\endgroup$
    – John
    Commented Aug 11, 2021 at 16:16
  • 3
    $\begingroup$ This should be the most upvoted answers. There is almost no discernable difference between what the OP is asking for in a self-learning AI and how the body reacts to it's own chemicals. OCD itself is the absence of dopamine being released when performing a task, which tells the individual that they have completed the task. The absence of this signal (chemical, dopamine) means that people suffering from this have to continue to perform the task (i.e. switching light switches). Humans have chemicals, but at the end of the day they are just chemicals that trigger electrical impulses. $\endgroup$ Commented Aug 12, 2021 at 14:04

It will use the Behavioral Psychology approach

Behavioral psychology is an approach to psychology that disregards any unmeasurable things like emotions, and only looks at people as stimulus response machines. Since I don't want to write a whole text book, I will focus on just one seemingly specific emotional state to use as an example: irritable. While just knowing the variable irritable is not enough to predict how a human will act. Knowing the exact set of stimuli that caused it is much more likely to be predictable. The more stimuli you know of, the more predictable humans are.

In this respect, a robot capable of being simultaneously cognoscente of more things at once than a human would have less need of generalizing the human experience into emotions because it would be able to view us more as the complex systems of converging factors that we really are. In other words, it would not describe a human as be irritable: it would describe a human as:

  • last ate 6 hours ago (3 slices of pizza and 16oz of soda)
  • last slept 14 hours ago (for 5 hours)
  • last menstruated 20 days ago (pregnancy status unknown)
  • last yelled at by boss 2 hours ago
  • drive home took 52 minutes (normally takes 25)
  • last ignored by husband 2 minutes ago (third time today).

So, the robot will not need to understand the word "irritable" as we understand it. Over time the AI will be able to make predictions about how sleep deprivation, poor dietary habits, the reproductive cycle, social conflict, poor outcomes, and lack of validating input can compound to increase human aggression. It will also figure out that the exact kind of increased aggression these factors stimulate can solve these problems too, and that all the things we call "irritable" which we consider to be one of the less predictable emotional states may in fact just be several slightly different perfectly logical things.

If you are hungry, boosting adrenaline releases fat stores to feed your body. Increased aggression makes since to the AI, because we are recycling the same bodily function that we use when preparing to fight.

If we have not slept in a long time, increased aggression encourages other humans to demand less attention of us allowing us to stop what we are doing and get some rest.

If we are in a point in the reproductive cycle where procreation is impossible, then spending time on romantic relationships is a waste of time; so, we use aggression to create distance freeing us for more practical pursuits.

If we are yelled at by a superiors and don't fight back, it threatens our position in the hierarchy; so, following up this behavior with starting fights with our peers or inferiors helps make sure that our social position does not fall which in turn protects our resources.

If we spend much longer on completing a task than we think we should, then a boost of adrenaline can be helpful to catch back up; so, getting home late can be made up for by pushing harder when it comes time to make dinner, washing dishes, and doing what ever else you need to do in your remaining time.

If someone is ignoring you, then you increase your chances of being recognized by becoming a threat to the person ignoring you.

So, the AI would never understand irritable or joy, anger, sadness, fear, etc. as any one thing. It would understand these things as constructs of the human language to generalize similar outputs into singular words, but the AI would be far more capable of understanding why humans have feelings than most of us are. In a way, the AI could be far more sympathetic than a human because it can break down the human experience into its parts instead of only being able to see the big picture. While a human is often confused by being yelled at, the AI could tell if you need it to give you space or validate your feelings, or yield in some immaterial way because it could tell which kind of irritable you are.

  • $\begingroup$ This answer is extremely useful, I'll take note of this. +1 (but I wish I could give two) $\endgroup$ Commented Aug 12, 2021 at 20:56

This is how R. Daneel Olivaw, an artificial intelligence, describes his feeling when seeing after long time his partner Elijah Bailey

“I cannot say what I feel in any human sense, Partner Elijah. I can say, however, that the sight of you seems to make my thoughts flow more easily, and the gravitational pull on my body seems to assault my senses with lesser insistence, and that there are other changes I can identify. I imagine that what I sense corresponds in a rough way to what it is that you may sense when you feel pleasure.."

Even though the AI might not feel emotions, would still understand that they are a variable in the behavior of the individual they are interacting with.

What function would be assigned to them cannot be answered without knowing the implementation of the AI.


OpenCog is not far-future science fiction AI, but it is an attempt at artificial general intelligence.

There are nodes of information. However, most data processing happens in links between nodes. Also, beyond the most basic settings, it is important how the AI focuses the attention it gives to certain atomic ideas and their connections.

How does that apply to emotions? Let's try:


  • Think of joy as adjusting the attention on positive nodes and reducing attention on negative ones. A pictural metaphore is turning down the contrast and turning up the brightbess on an image.


  • What the brain is doing with anger is shutting down links to most other parts of the brain, cutting down the cross-chatter from the rest of the brain, and giving most attention to the short-term, immediate problems. A pictural metaphore is cropping an image to the part you are interested in.

Surprise, Love, Grief

  • These are subjects that require a lot of attention, and new information about these subjects is getting flagged for first processing.

Sadness, Fear, Disgust

  • At it's most extreme, these emotions are like anger. At more modest levels, this is a small adjustment of attention to the topic at hand. A pictural metaphore is tweaking settings and maybe zooming to get the best view of the thing you are interested in.


  • A bit opposite of joy. This is like cutting the brightness and turning up the contrast on a picture.
  • $\begingroup$ fear and disgust are also avoidance coding, things that can actively damage your bot or cause a lot of damage code as fear, things that can passively degrade the bot code as disgust. Schadenfreude is unlikely to be a desirable in machine that has to interact with humans. $\endgroup$
    – John
    Commented Aug 11, 2021 at 15:47
  • $\begingroup$ @John Being able to take pleasure in the misfortune of others helps you dismiss the needs of your rivals when competing for power, resources, or survival. Because we also have other emotions centered on affection and fellowship that suppress our schadenfreude for people we care about, schadenfreude is a vital part of forming protective "Us VS Them" communities. So, if an AI were to practice schadenfreude, it could help with reinforcing his relationships with the humans that accept him, and protect him from humans who do not. $\endgroup$
    – Nosajimiki
    Commented Aug 12, 2021 at 14:43
  • $\begingroup$ @Nosajimiki Us vs Them is unlikely to be desirable in a robot, for one thing it is more likely to identify with other robots than humans, plus factionalism is not really a desirable trait in modern societies. $\endgroup$
    – John
    Commented Aug 12, 2021 at 15:03
  • $\begingroup$ @John An AI is so different than humans that it is very likely to be discriminated against: treated as "them". This puts it in danger of humans turning against it and destroying it, but if it can find a faction that welcomes it, then that faction would protect it from other humans that would choose to harm it. Remember, the AI can't feel human emotions, so it does not matter to the AI if humans are like it. But it can perform human behaviors to interact with our emotions. So if it joins in with its faction's prejudice behaviors, then that faction will be more willing to see it as "us". $\endgroup$
    – Nosajimiki
    Commented Aug 12, 2021 at 15:29
  • $\begingroup$ @John I also disagree about factionalism being undesirable in modern societies. Most people are more likely to help family, friends, or neighbors than a total stranger. And they are more likely to help a stranger from their community, than another community. Their state than another state. Their nation than another nation. The more "us" another person is, the more likely it is for your kindness to eventually be repaid, and the more "them" a person is, the more likely it is to be exploited to your detriment. Ignoring these factors might make you a better person, but not better off. $\endgroup$
    – Nosajimiki
    Commented Aug 12, 2021 at 15:36

You can imagine emotions as a reward/penalty mechanism. In a very complex environment you cannot encode a response for every situation you encounter, the most effective solution is encode a flexible set of actions and a reward/penalty mechanism in order to derive the choice of the right response at the very last moment.

In AI at the moment rather that teaching how to understand the mechanism the focus is how to use the mechanism, how to define rewards and let the AI find the right action to get the reward instead of predefining the action itself. On this subject there have been many experiments, one example may be the Animat where the basic feeling like hunger, thirst, satisfaction or sickness by eating toxic food are used to drive simulated animals to explore the environment.

So your future AI to understand the meaning of emotions should first understand the concept of complexity even beyond the simple definition of NP complete problems, thus it could realise that being impossible to predefine all the possible behavioural responses the only solution is to predefine rewards and penalties that are vague enough to cover a wide range of situations.

For what matters the functions to assign to each emotion the best way to do is just to refer to the psychology literature and see what was derived by observation, nature is not so different and it had a long long time to work on this task. For an AI to learn such a complex task it should be extremely advanced. It would be better to let the AI cooperate with some humans.


I see that you changed the nature of the AI into something created by aliens. That would make things really difficult. Can you deduct the emotion of a lizard? We say that an animal is hungry or thirsty because we also sometimes are hungry or thirsty, we project into the animals part of the knowledge we have of ourselves, the rest comes from observation. The AI would need a long long time to learn by observation. If their aliens that created the AI had some basic emotions it could understand from the beginning the principle of reward/penalty. It could understand by logic that the signals have to be vague in order to cover with few traits a large and complex field. You can add a shortcut if the AI arrives in our era and it can decipher our language and read our literature. Otherwise it would need centuries if not millennia of comparative studies similar to those done by Ethologists


An artificial intelligence, if it chose to think about such things rather than considering them irrelevant, would study them and discover their true nature, itself being unburdened by laboring under such things as emotions. Supposing of course, that it is intelligent enough to do so.

This is one of those situations where the writer can't write characters more intelligent than itself. So if you were hoping for a specific answer as to how a super-intelligent being might come to view some phenomenon when you yourself aren't a super-intelligent being, then you are bound for disappointment. The consolation is that anyone who might read your story won't be smart enough to know you're wrong either.

If I had to guess though, I'd speculate that what humans tend to lump together into the category of "emotions" aren't all a single phenomenon that obey the same set of rules or that have anything other than superficial similarities. Several probably fall into something you'd loosely call "specialized modes of thinking", other still as "prioritization modification", and even a few as "mild mental illnesses that evolution found to have survival advantage enough to induce". All of these are, in the strict sense, "rational", it's just that humans have difficulty understanding the rationality, or that they focus on the corner-cases where they are maladaptive.

Your question is sort of like asking "When Batman tells a joke and it causes the robot's head to blow up because it cannot laugh, what is the yield of that explosion". It's a misunderstanding of the entire premise rolled up into an ignorant plot point.

  • $\begingroup$ I don't have many works under my belt at this point, so how smart I actually am remains to be seen. I am able to aproximate what you might call "cautious optimism". $\endgroup$ Commented Aug 11, 2021 at 14:26

This is actually something addressed in the Pixar Film "Inside Out" and it's addressed early on as, of the five core emotions of Riley, only one of them is what on it's face is a positive emotion. Almost immediately into the film, Joy explains that three of the emotions provide very vital functions (and does so in, what else, a positive light). Fear keeps Riley safe by identifying potential threats and working out how to best deal with them (Shown when Fear takes control of a young Riley's happy play time away from Joy and carefully maneuvers her around a power cord, only to let her go back to wild play time again once the threat is removed). Disgust is described as an emotion that alerts when Riley is about to be poisoned, either literally or figuratively, and is shown doing this by stealing the wheel when she detects broccoli is about to fed to her and influences Riley to try not to eat it. Anger is there to protect Riley and fight for her, and is triggered when her father responds to Riley's refusal to eat Broccoli by threatening no desert, and takes over from Disgust to manipulate Riley into a tantrum. Both Anger and Disgust yield when a compromise is reached... namely the "Here comes the Airplane" technique that most parents use on their picky eaters.

The film's whole existence explores why Sadness exists, something Joy fails to understand (We do get an early demonstration at the dinner scene where Riley's mother is shown to have a dominate emotion of Sadness and is able to pick up on the families problems without any overt prompts from conversations... Sadness allows people to empathize with each other.).

These characters were not chosen at random but are 5 of the 6 basic emotions as explained by psychologist Paul Ekman (the missing member is "Surprise" who has her(?) job covered in part by Fear, as shown in Joy's failed attempt to make a dream sequence, though Disgust and Anger cover the missing emotion as well).

These emotions are the most basic and from them working in combination, form all the more complex emotions (shown by the memory spheres being multicolored as opposed to single color, to correspond to the emotion attached to the memory.).

A true AI would see emotions as programs that, while possibly inefficient, work together to evaluate memories and assign a weight to them that is used in the evaluation of a decision making process. The emotion of Happiness (Joy) would be a desired outcome, but may not be optimal. For example, you might feel joy from drinking alcoholic beverages, but you would also feel an uncomfortable amount of sadness or disgust from external influences from overconsumption of alcohol and the societal stigma of being an alcoholic, not to mention fear of consequences from your actions while in an intoxicated state, thus would evaluate that you it if it is 0500 it is thus too early to drink, especially when you have a big performance evaluation with your boss later in the day.

  • $\begingroup$ And here I thought I wouldn't see Inside Out being mentioned, though I felt it coming when I listed the emotions. Well done! You made my day. $\endgroup$ Commented Aug 11, 2021 at 14:24
  • $\begingroup$ According to Ekman's six, Joy was replaced with Happiness, but beyond that all the emotions from Inside Out were listed among the core six emotions. $\endgroup$
    – hszmv
    Commented Aug 11, 2021 at 18:50


It is literally impossible to have an emotionless AI. Your question is therefore based on false assumptions.

Hollywood has taught people that there's a difference between "logic" and "emotion", and that machines operate on "logic", rather than emotion. The problem is there are two completely different definitions of "logic" here. One is exactly synonymous with emotion, and that's the one your AI absolutely, positively must have to function.

Any AI with the ability to question human emotion would also have to have high-level, rational logic, with which it could model human emotions and compare them to its own. How it would view human emotions is greatly dependent on how similar ours are to its.

Emotional Logic

Emotion is nothing more than the basic impulses that drive us to act. It's not magic. It's not a human-only condition. It's not high level. It's not "chemistry" vs "electronics". It's just low-level logic gates that interpret inputs and convert them to outputs.

Any agent with the capacity to interact with the environment must have low-level logic. It's precisely what gives them the capacity to interact. Since low-level logic is emotion, that means any agent with the capacity to interact must have emotion.

Humans have emotions, yes. So do birds, mice, cockroaches and Roombas.

Rational Logic

But there's another definition of "logic". That's high-level reasoning. Wisdom. Sapience. Rational thought. Whatever word you use, it's the capacity to learn from our past, model the world around us, make predictions about the future, and decide which actions give us the best future outcome.

Humans have the capacity for this high-level logic. We used it to come up with science, mathematics, etc. We use it to model the world. To build machines. To make decisions that might be inefficient in the short term to gain more in the long term.

But the only way we can decide what's "good" or "bad" is by comparing possible solutions to what's desirable. What's desirable (or not) is entirely based on our low-level, emotional logic. There is no "pure logic", no absolute truth, no end game that all intelligent beings are required to pursue.

An AI could potentially be capable of rational logic, but doesn't have to be. However, an AI cannot possibly exist without emotional logic.

Any AI capable of rational logic (and, therefore, capable of "viewing" human emotions to begin with), would be perfectly capable of analyzing humans to model our behavior. Just like we could model it.

Humans have the advantage that even if we can't explain our feelings, we can empathize with other humans, and therefore "understand" even when we don't truly know what's going on. But an AI doesn't necessarily have that same capacity.

A very advanced AI might see us as primitive and easily predict our actions, while a less advanced AI might have trouble understanding us precisely. The less human-like the AI is, the harder it is to understand us.

But if it's smart enough to question our emotions, it's smart enough to at least get a basic understanding of them.

Emotional Feelings

Sometimes, when we think of "emotions", we're including the extra sensory inputs humans feel, that another rational entity might not.

If I'm hungry, I get a little inkling that I'm hungry. Then my stomach starts sending aching sensations. Then there's physical pain. Etc.

If I'm tired, my skin feels like it has sand on it. My blood pressure changes. My heart rate tries to slow. My eyes are difficult to keep open. Etc.

These perceptions are a natural consequence of the complex machinations that allow our biology to function. But a non-human entity doesn't inherently need to feel their skin crawling when it's cold and humid, even if that's a less-than-ideal situation to be in.

So these "feelings" might be alien to a non-human. They might find it weird to think about a "gut reaction" or an "itchy trigger finger". But they wouldn't think about basic desires and impulses as being weird or alien, because they'd have those things too.

Sensory Feeling

To "feel" literally refers to tactile sensations. But in a more general sense it means to perceive.

An AI wouldn't necessarily have tactile sensations, so it wouldn't necessarily "feel" in that sense. But it would necessarily have some form of perceptions in order to interact with the outside world.

That is, all AI must at least have something analogous to sensory feeling, even if they don't have literal feeling.

As such, even if an AI doesn't connect the same sensations to emotions we do, it would at least be capable of understanding the concept of perception, and knowing that our emotions are linked to our perceptions in ways its are not.

Whether it views our emotional feelings as useful or not depends greatly on how advanced it is compared to us. And how connected its emotions are to its perceptions. So it's impossible to say for sure, but it would likely view some feelings as more useless, and others as more useful.


Any AI is absolutely required to have low-level, emotional logic to exist. It would, therefore, see human emotions as being analogous to its own emotions.

An AI is not required to have high-level, rational logic, but won't have the capability to compare its emotions to human emotion unless it does.

An AI wouldn't necessarily have the same sensations we have, even if it has the same basic emotion. As such, it might consider "emotional feelings" odd or unusual.

An AI wouldn't necessarily have the ability to "feel" in the tactile sense, but would necessarily have some kind of perception directly analogous to feeling so it can perceive the world around it.

The AI would tend to perceive human emotions antithetical to its own goals as more "useless", those neutral as "meh", and those sympathetic as "useful". But any reasonably-intelligent AI would at least be able to understand how an emotion could be useful to our goals.

  • 1
    $\begingroup$ Sassy sarcastic AI it is then. $\endgroup$ Commented Aug 12, 2021 at 12:09

Frame challenge:

Emotions are not a type of input that the AI would deal with at all. There is no reason an AI would "feel emotions" nor would there be any use for doing so.

Emotions are an expression of other processes. A type of output. The AI would understand that when it tells a human being, say, "please do X", the possible outputs are:

  • "yes" (acceptance)
  • "no" (refusal)
  • "why?" (request for more information)

each of those can be accompanied by an emotion such as getting angry. The AI would understand this as a kind of modifier, a function that communicates the inner state of the human to his communication partner.

And that is what emotions are, frankly speaking. Looking at the animal kingdom, we notice that expressions of emotions are most clear in social animals and much more subdued in solitary animals.

  • $\begingroup$ The mental picture I had about AI “emotions” would be seeing it fail to open a can of sardines, repeatedly attempt to open it, then getting “frustrated” and trying harder until the can cracks open. This, followed by “exasperation” at the amount of time and energy wasted on one task. $\endgroup$ Commented Aug 13, 2021 at 9:01

The alien can be fooled, as humans will

The answer to this question depends on what reality model you would provide your AI with, and the extent of emotional state and accompanying personality you want your AI to develop on top of reality. The question has an observer: the alien, not able to discern actual human emotions from emotions "felt" by the AI. Same would count for humans as well. Technological development sofar has focussed on that: letting the observer perceive the artificial entity as real and invoke an emotion, reflecting the emulated emotion in the robot, the program, or the AI.

Conditional expression of emotions is the easy part

Japanese manufacturers created emotional expression in a device 25 years ago. It would learn to dose expressions of joy, anger, sadness, fear, disgust, depending on certain stimuli and timing: tamagochi. A little pet-like gadget, with a smiley on its display that indicated love or frustration, making you hug it, feed it, give it attention.

Expression of emotion is everywhere in games now. A game character , like the very friendly and thankful Egyptian shown in Pyramid of Mahjong is actually an emotional looking facade for a deterministic algorithm, leading you to the next Mahjong puzzle, or congratulating you with reaching the next level in the game. The characters clearly express emotions of joy or grief, like humans would do. It even makes you smile back, purpose achieved.

With current technology, an autonomous AI expressing emotion could easily be put onto a game console, showing a spectrum of behaviours connected with a basic emotional state.. linked with game scenario's, player actions, or whatever the programmers can think of as "soft-wired". The actual AI could develop interaction with the player.. speech recognition, chats..

Next step: simulation

Letting human players reflect the emotion is the actual goal of emotional expression in games. It is about the player's feeling, there is no feeling in the game character. An animation is shown.

In a simulation of emotion, feeling would cause the expression. Many have commented on that aspect: how to map the emotions to the situation, or to past events.. how to evaluate awaited and unexpected events.. In these models, a particular feeling is part of a vector, determined by the events. Some threshold would invoke the emotional expression. There is (again) a deterministic, or learning algorithm, that could be stateless. Actual emotion is never stateless. It has a cause in reality.

The AI person

Now let's handwaive a few things.. suppose a reality model would be feasable.. the AI being able to build a coherent emotional state in a social context, that can be modeled dynamically, able to autonomously decide to express emotion or not. Such an AI could become a collegue, or a friend for humans. A robot that is able to invoke emotional attachment still does not need to have "feelings" as such, but it would be perceived as a sentient being.

In that case, the order of development of emotions in robots could pose certain ethical challenges. You can't put grief on top of the list and develop that first: the AI has nothing to grieve for yet, because there has been no love and no joy. How would a conscient being, starting out as the "AI project", but now aware of joy and love have to be developed further ? Will the makers of this AI knowingly subject their friend to any grief, anger or fear..


Emotions can be a useful input to the fitness function.

Each AI must have some goals, or it would not do anything. Those goals might be general, specific, benevolent or evil. Examples are

  • keep humanity alive
  • keep Human X alive for as long as possible, ideally unhurt
  • enslave all humans
  • save the dog
  • save the girl

While you can often break down if a given action would work towards that goal or against it. But usually the question is not binary. Instead the AI has to decide between many, many options of which it can only enact a limited number (due to resource constraints or because they contradict each other. So the question becomes "which combination of these actions will lead to the best outcome".

And the fitness function is how it "measures" best outcome.

Let's assume the goal is to keep humanity alive.

  • Action A would be to cage them all into cells, automate the production of nutrients and force-feed them
  • Action B might be to enable humans to live on their own terms and protect/support them in various less invasive ways.

There are many cons and pros to each of those approaches and many of them can easily be calculated.

But some are harder to grasp and that's where emotions can become handy.

The AI might have learned that happy humans are more likely to succeed in their goals and have improved long-term survival. So happiness could be a useful tiebreaker between two options that would otherwise look equally promising.

The AI might have learned that anger can lead to humans action irrationally and this can be problematic for any long-term plans, because irrational humans tend to cross plans.

The AI might have learned that grief can reduce the productivity and happiness of humans, so it might try to save a puppy it would otherwise not care about.

In other words: the AI doesn't need to experience or even understand emotions for them to become useful input into its algorithms.

The interesting thing is that a sufficiently advanced version of this could even look like it has empathy. Some might even argue that a sufficiently advanced version of this is empathy.

A big caveat to this is that those inputs will only have a certain weight to it. So it's quite possible that the AI recognizes that a certain plan of action will produce negative emotions, but it considers other factors more important and therefore still execute the plan. This could be viewed as "cold-hearted" by those who ascribe emotions to the AI.


This question is broader than the one I originally posted this answer for, so it's not a duplicate. But the answer I would give is more or less a duplicate of my old answer, so here it is:

Emotions are desirable. They are kept around because they work. Loneliness serves very obvious purposes. Being around other people who care about you is good for your long term survival. Caring about other people is good for their long term survival. Loneliness is just as much a good thing as guilt and pain are. These things make you feel terrible, but they're good for you.

Consider the alternative. Yeah, you could build an AI that would (say) evaluate social interactions the way that Deep Blue crushed every human at chess: by exhaustive inspection.

But look at how powerful a machine we have to make to beat a human at chess. AI people have no idea how to program a computer to do it, but they know people play chess very differently. A good chess player only considers a handful of options, because their skill tells them those are the only ones worth considering.

AI solves these problems by smashing the square peg into the round hole with a sledgehammer. It simply analyzes millions of combinations until it finds the one that it knows works best because it looked at all of them. Much AI work is about making things run efficiently.

So why would I want an AI with no emotions? It will have to spend mental energy / computational power running simulations to decide things that emotions solve with basically no effort. An AI with emotions enabled will be able to take the saved processing power and spend it elsewhere. It will be more capable.


An emotion is the experience we have when one of our instinctive drives kicks in. When we need to eat, we feel hungry. That is an emotion. As adults we have a suite of instinctive drives to find a mate, form a pair bond, have sex, and look after the resulting offspring. All of the sensations you feel that drive you to do those things are emotions.

What is joy?

Joy is the emotional response to success. The pursuit of joy and happiness drive our activities in the long term in ways that are generally beneficial to reproduction and survival of our offspring.

Of course not everything that our emotions drive us to do are beneficial. Our instinctive drives are simple mechanisms, easily misled. Especially by our sophisticated brains. Biologists talk about "supernormal stimuli" as a framework for understanding this.

For instance, one of our instincts seems to be that a "near miss" when attempting something is a good reason to keep trying. At the risk of telling a "just so story", it's easy to see how this might be adaptive in a species that hunts. A near-miss in a hunt means that you were doing the right thing and need to try again. But this is easily manipulated in gambling; if you bet on a horse and it comes in second it feels like a near-miss, so you go back and try again. Lottery card designers make sure that their cards have lots of near-misses to reward the players and keep them coming back.

How would a true artificial intelligence view emotions as useful and assign functions to them?

This has it backwards. The AI will have "instinctive" drives (for want of a better term) built into it by its designer. As such, they will reflect the goals of the designer. The AI won't chose its emotions to suit its goals, it will choose goals to suit the emotions given it by its designer.

In other words, an artificially intelligent door controller really will feel pleasure to open for you, and satisfaction to close again with the knowledge of a job well done.

  • $\begingroup$ “In other words, an artificially intelligent door controller really will feel pleasure to open for you, and satisfaction to close again with the knowledge of a job well done.” That’s the doors in The Hitchhiker's Guide to the Galaxy! $\endgroup$ Commented Aug 12, 2021 at 15:13

Other answers have treated the situations, physical constructs, physiology, brain chemistry and so on… that are associated with emotions.

I point you to a paper that I found easily by searching “philosophy what like bat” (in Google). [“What is it like to be a bat?”] There is also a term “quale” (pl. “qualia”) that you will want to look up.

The point — posed as a question — is that there is a subjective experience for a [self-]conscious being, around the pertinent event/experience, that can not be communicated through any amount of information about the mediating systems. One might argue that one could, for instance, explain, to a blind person, what it is like looking at a beautiful sunset… presumably because that is about pleasure (crudely speaking), and blind persons are capable of pleasure. However, if and insofar as there is any difference in the subjective experience of looking at [for instance] a blue-coloured area and looking at a lime-coloured area, ostensibly it will be quite impossible to communicate that to a blind person. Nagel thought of the example of a being that has a sensory apparatus that human beings do not — particularly, a bat using sonar to “see”. The idea is that no amount of studying bats can possibly ever give a human being subjective knowledge of what that is like.

The AI has this problem, with human beings. Presumably, it has no emotions at all, in this subjective sense. Further, even if it did, it could not possibly have any idea of exactly (and thus generally) what it is like for a human being to experience any given human emotion.

There are other issues as well. One obvious one is that individual human beings are different, in respect of exactly what emotions a given one will experience under a given “input”… and what behaviour will follow.

Also, it is far from obvious that emotions such as depression are useful in an evolutionary kind of sense. Arguably, human beings commonly or even normally seek pleasure (or relief from pain) in self-destructive actions, when under duress. Arguably, anger does not align with what is ethically ideal and overall the most constructive.

Possibly a powerful intelligence could form an (externally) coherent picture of the whole thing… but it does not follow that emotions are rational. Arguably, quite the opposite.

For the OP’s story… . Presumably, the AI would start out oblivious to the very concept of emotion. Arguably, the AI would discern its existence from the fact that human behaviour was ofttimes irrational. (It would presumably hear and learn of emotion along the way, but, as above, that would be a question, not an answer.) Later on in the piece, it would either work out for itself, or hear explained, the above… or not.

One variation would have it [being] certain that emotions must be useful (in the evolutionary sense), and eventually gradually realise that they are (sometimes) destructive… or not [realise].

p.s. Humour is complex. It is entirely plausible (at face value) that there could be an alien species that had a different humour system. If that obtains, it would follow that there was no systematic and rational — that is, exclusive — account of what is funny.
Similarly with beauty; it is not merely plausible, but almost certain in the details, that an alien would have a conception of beauty different in the detail from the human one. (Arguably, simplicity and repetition both inform beauty, but an alien might prefer different patterns (or see no pertinent significance in patterns), and similarly with (for instance) landscapes of green vegetation and blue sky.)

  • $\begingroup$ Sorry about that. No spying intended. [Your comment is a spoiler; theoretically, you might want to delete it.] $\endgroup$
    – Carsogrin
    Commented Aug 13, 2021 at 14:28
  • $\begingroup$ Got it. Anyway, answer appreciated :) $\endgroup$ Commented Aug 13, 2021 at 15:14

Not the answer you're looking for? Browse other questions tagged .