# What purpose would be served by developing AI's that experience human-like emotions?

In a recent Wall Street Journal article, Yann LeCunn makes the following statement:

The next step in achieving human-level ai is creating intelligent—but not autonomous—machines. The AI system in your car will get you safely home, but won’t choose another destination once you’ve gone inside. From there, we’ll add basic drives, along with emotions and moral values. If we create machines that learn as well as our brains do, it’s easy to imagine them inheriting human-like qualities—and flaws.

Personally, I have generally taken the position that talking about emotions for artificial intelligences is silly, because there would be no reason to create AI's that experience emotions. Obviously Yann disagrees. So the question is: what end would be served by doing this? Does an AI need emotions to serve as a useful tool?

• Possible duplicate of Why would someone want to simulate emotional desicion-making within an AI? The accepted answer ("Meh, it'll just be easier for human programmers to maintain the AI programs.") may not be exciting or wonderful, but at least it seems semi-sensible and reasonable. – Left SE On 10_6_19 Aug 19 '16 at 22:57
• There's also an alternate possibility: we won't necessarily want to develop AIs that experience human-like emotions (because then they'd be unpredictable, we'd have to give them rights, blah blah blah), but that more advanced AI techniques (such as deep learning) will inadvertently develop AIs that experience human-like emotions. It's not that we want them, it's that they might come anyway. Do Artificial Reinforcement-Learning Agents Matter Morally? discusses this possibility. – Left SE On 10_6_19 Aug 19 '16 at 23:01

The answer to this question, unlike many on this board, I think is definitive. No. We don't need AI's to have emotion to be useful, as we can see by the numerous amount of AI's we already have that are useful.

But to further address the question, we can't really give AI's emotions. I think the closest we can get would be 'Can we make this AI act in a way a human would if that human was insert emotion?'. I guess in a sense, that is having emotion, but that's a whole other discussion.

And to what end? The only immediate reason coming to mind would be to create more lifelike companions or interactions, for the purposes of video games or other entertainment. A fair goal,
but far from necessary. Even considering an AI-imbued greeting robot in the lobby of some building, we'd probably only ever want it to act cordial.

Yann says that super-advanced AI would lead to more human-like qualities and flaws. I think it's more like it would 'give our AI's more human-like qualities or in other words flaws'. People have a tendency to act irrationally when sad or angry, and for the most part we only want rational AI.
To err is human, as they say.

The purpose of AI's and learning algorithms is to create systems that act or 'think' like humans, but better. Systems that can adapt or evolve, while messing up as little as possible. Emotive AI has uses, but it's certainly not a prerequisite for a useful system.

I think the fundamental question is: Why even attempt to build an AI? If that objective is clear, it will provide clarity to whether or not having emotional quotient in AI make sense. Some attempts like "Paro" that were developed for therapeutic reasons requires they exhibit some human like emotions. Again, note that "displaying" emotions and "feeling" emotions are two completely different things.

You can program a thing like paro to modulate the voice tones or facial twitches to express sympathy, affection, companionship, or whatever - but while doing so, a paro does NOT empathize with its owner - it is simply faking it by performing the physical manifestations of an emotion. It never "feels" anything remotely closer to what that emotion evokes in human brain.

So this distinction is really important. For you to feel something, there needs to be an independent autonomous subject that has the capacity to feel. Feeling cannot be imposed by an external human agent.

So going back to the question of what purpose it solves - answer really is - It depends. And the most I think we will achieve ever with silicone based AIs will remain the domain of just physical representations of emotions.

• I tend to agree. And maybe my thinking is just too limited, but most of the applications of AI that I think about, would not benefit from the AI having emotions. Well, at least I think so, sitting here now. Maybe I'll look back one day and realize I was wrong. – mindcrime Aug 23 '16 at 0:23
• Requiring that AI must/should have emotions is a misplaced expectation. AI basically is a set of tools, and its application therefore must also be decided based on the needs. Some use cases may require emotional aspect, some may not. Who knows! – Kingz Sep 16 '16 at 6:19

I think emotions are not necessary for an AI agent to be useful. But I also think they could make the agent MUCH more pleasant to work with. If the bot you're talking with can read your emotions and respond constructively, the experience of interacting with it will be tremendously more pleasant, perhaps spectacularly so.

Imagine contacting a human call center representative today with a complaint about your bill or a product. You anticipate conflict. You may have even decided NOT to call because you know this experience is going to be painful, either combative or frustrating, as someone misunderstands what you say or responds hostilely or stupidly.

Now imagine calling the kindest smartest most focused customer support person you've ever met -- Commander Data -- whose only reason for existing is to make this phone call as pleasant and productive for you as possible. A big improvement over most call reps, yes? Imagine then if call rep Data could also anticipate your mood and respond appropriately to your complaints to defuse your emotional state... you'd want to marry this guy. You'd call up call rep Data any time you were feeling blue or bored or you wanted to share some happy news. This guy would become your best friend overnight -- literally love at first call.

I'm convinced this scenario is valid. I've noticed in myself a surprising amount of attraction for characters like Data or Sonny from "I Robot". The voice is very soothing and puts me instantly at ease. If the bot were also very smart, patient, knowledgable, and understanding... I really think such a bot, embued with a healthy dose of emotional intelligence, could be enormously pleasant to interact with. Much more rewarding than any person I know. And I think that's true of not just me.

So yes, I think there's great value in tuning a robot's personality using emotions and emotional awareness.

Emotion in an AI is useful, but not necessary depending on your objective (in most cases, it's not).

In particular, emotion recognition/analysis is very well advanced, and it's used in a wide range of applications very successfully, from robot teacher for autistic children (see developmental robotics) to gambling (poker) to personal agents and politics sentiment/lies analysis.

Emotional cognition, the experience of emotions for a robot, is much less developed, but there are very interesting researchs (see Affect Heuristic, Lovotics's Probabilistic Love Assembly, and others...). Indeed, I can't see why we couldn't model emotions such as love as they are just signals that can already be cut in humans brains (see Brian D. Earp paper). It's difficult, but not impossible, and actually there are several robots reproducing partial emotional cognition.

I am of the opinion that the claim "robots can just simulate but not feel" is just a matter of semantics, not of objective capacity: for example, does a submarine swim like fish swim? However, planes fly, but not at all like birds do. In the end, does the technical mean really matters when in the end we get the same behavior? Can we really say that a robot like Chappie, if it ever gets made, does not feel anything just like a simple thermostat?

However, what would be the use of emotional cognition for an AI? This question is still in great debates, but I will dare offer my own insights:

1. Emotions in humans (and animals!) are known to affect memories. They are now well known in neuroscience as additional modalities, or meta-data if you prefer, of long term memories: they allow to modulate how the memory is stored, how it is associated/related with other memories, and how it will be retrieved.

2. As such, we can hypothesize that the main role of emotions is to add additional meta-information to memories to help in heuristic inference/retrieval. Indeed, our memories are huge, there are a lot of information we store over our lifetime, so emotions can maybe be used as "labels" to help retrieve faster the relevant memories.

3. Similar "labels" can be more easily associated together (memories of scary events together, memories of happy events together, etc.). As such, they can help survival by quickly reacting and applying known strategies (fleeing!) from scary strategies, or to take the most out of benefitting situations (happy events, eat the most you can, will help survive later on!). And actually, neuroscience studies discovered that there are specific pathways for fear-inducing sensory stimuli, so that they reach actuators faster (make you flee) than by passing through the usual whole somato-sensory circuit as every other stimuli. This kind of associative reasoning could also lead to solutions and conclusions that could not be reached otherwise.

4. By feeling empathy, this could ease robots/humans interaction (eg, drones helping victims of catastrophic events).

5. A virtual model of an AI with emotions could be useful for neuroscience and medical research in emotional disorders as computational models to understand and/or infer the underlying parameters (this is often done for example with Alzheimer and other neurodegenerative diseases, but I'm not sure if it was ever done for emotional disorders as they are quite new in the DSM).

So yes, "cold" AI is already useful, but emotional AI could surely be applied to new areas that could not be explored by using cold AI alone. It will also surely help in understanding our own brain, as emotions are an integral part.

I think that depends on the application of the AI. Obviously if I develop an AI that's purpose is plainly to do specific task under the supervision of humans, there is no need for emotions. But if the AI's purpose is to do task autonomously, then emotions or empathy can be useful. For example, think about an AI that is working in the medical domain. Here it may be advantageous for an AI to have some kind of empathy, just to make the patients more comfortable. Or as another example, think about a robot that serves as a nanny. Again it is obvious that emotions and empathy would be advantageous and desirable. Even for an assisting AI program (catchword smart home) emotions and empathy can be desirable to make people more comfortable. It would be much nicer to be welcomed by an empathic home assistant than by one with no empathic responses at all, wouldn't it?

On the other hand, if the AI is just working on an assembly line, there is obviously no need for emotions and empathy (on the contrary in that case it may be unprofitable).

### Strong AIs

For a strong AI, the short answer is to call for help, when they might not even know what the supposed help could be.

It depends on what the AI would do. If it is supposed to solve a single easy task perfectly and professionally, sure emotions would not be very useful. But if it is supposed to learn random new things, there would be a point that it encounters something it cannot handle.

In Lee Sedol vs AlphaGo match 4, some pro who has said computer doesn't have emotions previously, commented that maybe AlphaGo has emotions too, and stronger than human. In this case, we know that AlphaGo's crazy behavior isn't caused by some deliberately added things called "emotions", but a flaw in the algorithm. But it behaves exactly like it has panicked.

If this happens a lot for an AI. There might be advantages if it could know this itself and think twice if it happens. If AlphaGo could detect the problem and change its strategy, it might play better, or worse. It's not unlikely to play worse if it didn't do any computations for other approaches at all. In case it would play worse, we might say it suffers from having "emotions", and this might be the reason some people think having emotions could be a flaw of human beings. But that wouldn't be the true cause of the problem. The true cause is it just doesn't know any approaches to guarantee winning, and the change in strategy is only a try to fix the problem. Commentators thinks there are better ways (which also don't guarantee winning but had more chance), but its algorithm isn't capable to find out in this situation. Even for human, the solution to anything related to emotion is unlikely to remove emotions, but some training to make sure you understand the situation enough to act calmly.

Then someone has to argue about whether this is a kind of emotion or not. We usually don't say small insects have human-like emotions, because we don't understand them and are unwilling to help them. But it's easy to know some of them could panic in desperate situations, just like AlphaGo did. I'd say these reactions are based on the same logic, and they are at least the reason why human-like emotions could be potentially useful. They are just not expressed in human-understandable ways, as they didn't intend to call a human for help.

If they tries to understand their own behavior, or call someone else for help, it might be good to be exactly human-like. Some pets can sense human emotions and express human-understandable emotion to some degree. The purpose is to interact with humans. They evolved to have this ability because they needed it at some point. It's likely a full strong AI would need it too. Also note that, the opposite of having full emotions might be becoming crazy.

It is probably a quick way to lose any trust if someone just implement emotions imitating humans with little understanding right away in the first generations, though.

### Weak AIs

But is there any purposes for them to have emotions before someone wanted a strong AI? I'd say no, there isn't any inherent reasons that they must have emotions. But inevitably someone will want to implement imitated emotions anyway. Whether "we" need them to have emotions is just nonsense.

The fact is even some programs without any intelligence contained some "emotional" elements in their user interfaces. They may look unprofessional, but not every task needs professionality so they could be perfectly acceptable. They are just like the emotions in musics and arts. Someone will design their weak AI in this way too. But they are not really the AIs' emotions, but their creators'. If you feel better or worse because of their emotions, you won't treat individul AIs so differently, but this model or brand as a whole.

Alternatively someone could plant some personallities like in a role-playing game there. Again, there isn't a reason they must have that, but inevitably someone will do it, because they obviously had some market when a role-playing game does.

In either cases, the emotions don't really originate from the AI itself. And it would be easy to implement, because a human won't expect them to be exactly like a human, but tries to understand what they intended to mean. It could be much easier to accept these emotions realizing this.

### Aspects of emotions

Sorry about posting some original research here. I made a list of emotions in 2012 and from which I see 4 aspects of emotions. If they are all implemented, I'd say they are exactly the same emotions as of humans. They don't seem real if only some of them are implemented, but that doesn't mean they are completely wrong.

• The reason, or the original logical problem that the AI cannot solve. AlphaGo already had the reason, but nothing else. If I have to make an accurate definition, I'd say it's the state that multiple equally important heuristics disagreeing with each other.
• The context, or which part of the current approach is considered not working well and should probably be replaced. This distinguishes sadness-related, worry-related and passionate-related.
• The current state, or whether it feels leading, or whether its belief or the fact is supposed to turn bad first (or was bad all along) if things go wrong. This distinguishes sadness-related, love-related and proud-related.
• The plan or request. I suppose some domesticated pets already had this. And I suppose these had some fixed patterns which is not too difficult to have. Even arts can contain them easily. Unlike the reasons, these are not likely inherent in any algorithms, and multiple of them can appear together.
• Who supposedly had the responsibility if nothing is changed by the emotion. This distinguishes curiosity, rage and sadness.
• What is the supposed plan if nothing is changed by the emotion. This distinguishes disappointment, sadness and surprise.
• The source. Without context, even a human cannot reliably tell someone is crying for being moved or thankful, or smiling for some kind of embarrassment. In most other cases there aren't even words describing them. It doesn't make that much difference if an AI doesn't distinguish or show this specially. It's likely they would learn these automatically (and inaccurately as a human) at the point they could learn to understand human languages.
• The measurements, such as how urgent or important the problem is, or even how likely the emotions are true. I'd say it cannot be implemented in the AI. Humans don't need to respect them even if they are exactly like humans. But humans will learn how to understand an AI if that really matters, even if they are not like humans at all. In fact, I feel that some of the extremely weak emotions (such as thinking something is too stupid and boring that you don't know how to comment) exist almost exclusively in emoticons, where someone intend to show you exactly this emotion, and hardly noticeable in real life or any complex scenerios. I supposed this could also be the case in the beginning for AIs. In the worst case, they are firstly conventionally known as "emotions" since emoticons works in these cases, so it's easier to group them together, but very few people seriously think they are, just like the example I gave.

So when strong AIs become possible, none of these would be unreachable, though there might be a lot of work to make the connections. So I'd say if there would be the need for strong AIs, they absolutely would have emotions.

By emotions he doesn't mean to add all sorts of emotions into an AI. He only meant the ones that will be helpful for taking vital decisions. Consider this incident for a second:
Suppose an AI self drive car is driving through the highway. The person sitting inside is the CEO of a company and he is running very behind on schedule. If he didn't get on time there will be loss of millions of dollars. The AI in the car has been told to drive as fast as possible and reach the destination. And now a rabbit (or some other animal) comes into the way. Now if the car puts emergency brakes then the passengers will get seriously hurt and plus there will be loss of millions as CEO won't be able to get to the meeting.

Now what will the AI do?
Since for an AI, their decisions are only based on their utility function. Hitting the rabbit and keep going will logically show a better option. But, should the AI take that decision.

There are many questions like these where an AI might stuck into a situation where moral based decisions will play a vital role.
The above scenario is just for an example point of view.

## Theory of mind

If we want a strong general AI to function well in an environment that consists of humans, then it would be very useful for it to have a good theory of mind that matches how humans actually behave. That theory of mind needs to include human-like emotions, or it will not match the reality of this environment.

For us, an often used shortcut is explicitly thinking "what would I have done in this situation?" "what event could have motivated me to do what they just did?" "how would I feel if this had happened to me?". We'd want an AI to be capable of such reasoning, it is practical and useful, it allows better predictions of future and more effective actions.

Even while it would be better for it the AI to not be actually driven by those exact emotions (perhaps something in that direction would be useful but quite likely not exactly the same), all it changes that instead of thinking "what I would feel" it should be able to hypothesize what a generic human would feel. That requires implementing a subsystem that is capable of accurately modeling human emotions.

Careful! There are actually two parts to your question. Don't conflate meanings in your questions, otherwise you won't really know which part you are answering.

1. Should we let AGI experience emotion per "the qualitative experience"? (In the sense that you feel "your heart is on fire" when you fall in love)

There doesn't seem to be a clear purpose as to why we'd want that. Hypothetically we could just have something that is functionally indistinguishable from emotions, but doesn't have any qualitative experience with respect to the AGI. But we are not in a scientific position where we can even begin to answer any questions about the origins of qualitative experience, so I won't bother going deeper into this question.

1. Should we let AGI have emotions per its functional equivalence from an external observer?

IMHO yes. Though one could imagine a badass AI with no emotions doing anything you'd want it to, we do wish that AI can integrate with human values and emotions, which is the problem of alignment. It would thus seem natural to assume that any well-aligined AGI will have something akin to emotions if it has integrated well with humans.

BUT, without a clear theory of mind, it doesn't even begin to make sense to ask: "should our AGI have emotions?" Perhaps there is something critical about our emotions that makes us productive cognitive agents that any AGI would require as well.

Indeed, emotions are often an overlooked aspect of cognition. People somehow think that emotionless Spock-like characters are the pinnacle of human intelligence. But emotions are actually a crucial aspect in decision making, see this article for an example of the problems with "intelligence without emotions".

The follow-up question would be "what sorts of emotions would the AGI develop?", but again we are not in a position to answer that (yet).

What purpose would be served by developing AI's that experience human-like emotions?

Any complex problem involving human emotions, where the solution to the problem requires an ability to sympathize with the emotional states of human beings, will be most efficiently served by an agent that can sympathize with human emotions.

Politics. Government. Policy and planning. Unless the thing has intimate knowledge of the human experience, it won't be able to provide definitive answers to all problems we encounter in our human experience.

• Politics and government point you made was an excellent example. My perspective however is that AI only needs to be able to understand emotion and not necessarily experience it. – Seth Simba Jan 7 '18 at 6:16

Human emotions are intricately connected to human values and to our ability to cooperate and form societies.

Just to give an easy example: You meet a stranger who needs help, you feel empathy. This compels you to help him at a cost to yourself. Let's assume the next time you meet him, you need something. Let's also assume he doesn't help you, you'll feel anger. This emotion compels you to punish him, at further cost for yourself. He on the other hand, if he doesn't help you, feels shame. This compels him to actually help you, avoiding your anger and making your initial investment worthwhile. You both benefit.

So these three emotions keep up a circle of reciprocal help. Empathy to get started, anger to punish defectors and shame to avoid the anger. This also leads to a concept of justice.

Given that value alignment is one of the big problems in AGI, human-like emotions strike me as good approach towards AIs that actually share our values and integrate themselves seamlessly into our society.

• ok points, but if you will go father you will see that emotions are just shortcuts for logical explanation why they are or are not useful. Rules could be derived from emotions and our biology. And those rules are kept in compressed and non obvious form as emotions and biology. For us it is necessity, but not for ai, although as for us it might work as optimization for AI if cost of such implementation in digital form will be lower then rules based implementation. Problem with rules is only one, we ourselves do not know them fully, but AI task is more narrow then our, might not need much rules – MolbOrg Sep 10 '16 at 15:55