by Brian Tomasik
First written: 30 March 2014; last update: 11 Oct. 2015

Summary

The difference between non-player characters (NPCs)a in video games and animals in real life is a matter of degree rather than kind. NPCs and animals are both fundamentally agents that emerge from a complicated collection of simple physical operations, and the main distinction between NPCs and animals is one of cognitive and affective complexity. Thus, if we care a lot about animals, we may care a tiny bit about game NPCs, at least the more elaborate versions. I think even present-day NPCs collectively have some ethical significance, though they don't rank near the top of ethical issues in our current world. However, as the sophistication and number of NPCs grow, our ethical obligations toward video-game characters may become an urgent moral topic.

Other versions of this essay:

  • Audio podcast (.mp3)

Introduction

Video games raise a number of ethical issues for society. Probably the most fiercely debated is whether gory video games affect proclivities for real-world violence. Some studies find no causal relationship between video games and violence, while other studies, including experimental manipulations, do suggest some causal pathway toward real-world aggression (see Wikipedia's article for more). My subjective experience jives with the "desensitization" hypothesis, though some will probably argue in the reverse direction, that video games allow for releasing aggression in a harmless way.

Another ethical issue sometimes raised is the addictive potential of video games, which can decrease school or job performance, impair relationships, and so on. Some also fear that video games are more mind-numbing than books and will inhibit the intellectual growth of youth. Others point out that video games are a form of art that can be quite complex and challenging, and MMORPGs like World of Warcraft may even teach players planning, teamwork, and leadership.

These topics are worth exploring, but they're not my focus here. In this piece, I ask a less common question: Do the characters in video games matter ethically, not just for instrumental reasons but for their own sakes? I don't know if I've ever heard this question asked before, at least not with respect to present-day video games. Most people would brush it off as ridiculous. But I think the answer is not so obvious on further examination.

Fiction vs. reality

Children learn to separate fiction from reality around the ages of 3-5 years old. Context helps determine the distinction. As an example, in Mister Rogers' Neighborhood, the show's real characters lived in the outside world, while the fictional characters lived in an explicitly separate "Neighborhood of Make-Believe." Children learn that their families, friends, national leaders, and science facts are part of reality, while fiction books, movies, and video games are fictional. It's okay if fictional characters get hurt, because they don't actually exist. When a child sees a character in a movie fall injured, a parent can console the child by explaining that the people on screen are just actors, and the blood is not real. Likewise, it seems obvious that the violence in video games is not real and hence doesn't matter unless it leads to antisocial behavior in the actual world.

But what is "reality" anyway? Our real world, of cars and families and love and future plans, is a complex arrangement of tiny physical particles into higher and higher levels of abstraction: atoms to molecules to proteins to cells to organs to people, which interact in societies. We are extremely elaborate machines built from masses of dumb, elementary physical operations that are put together in ways that form something beautifully dynamic and meaningful. "Reality" consists of all the higher-level abstractions that we use to make sense of these complicated particle movements. These abstractions are what matter to us and determine whether we're happy or depressed, hopeful or afraid. Some neuroscience research by Anna Abraham and D. Yves von Cramon suggests that our sense of what's real is determined by its relevance and degree of personal connection to us.

Now consider an elaborate, intelligent video-game character. It, too, is built in layers of abstraction from tiny physical stuff: electrons and atoms, to transistors, to logic gates, to computer hardware, run using binary code, which is created from higher-level code, which itself is organized in a modular fashion, from individual statements, to functions, to classes, to software packages, all within computers that can interact in computer networks.

We could say, "The computer character isn't 'real' because it's just a collection of dumb physical operations that are each extremely simple and unfeeling." But we too are built from dumb, "unfeeling" physical operations. The transistors that comprise the video-game character's "brain" are just as "real" as the neurons that comprise yours; in both cases, we can even point to the clumps of matter that are involved.

"Ok," one might acknowledge. "The character's brain is as real as mine, but its world is fake. I'm interacting with real physics, while this character is interacting with fantasy physics." But is there really such a big distinction? The trees, rocks, and walls in your world are clusters of fundamental particles behaving in sophisticated aggregate ways. For instance, the atoms in your chair are acting so as to hold themselves together. The video-game character also interfaces with objects (atoms and electrons in transistors, for example) that are made of fundamental physical particles acting in particular ways that define their behavior. The behavior of those particles is different between the cases, but the kind of thing going on is similar.

If it helps, we could imagine the dungeon walls of a Legend of Zelda game as being actual physical walls that the atoms defining the protagonist, Link, bumps up against when he tries to move into them. It's not the case that these walls in the CPU are spatially arranged in the same configuration as the walls in the game, but they're logically arranged in the configuration in the game, according to a complex mathematical mapping from atoms/electrons to virtual walls. (What is the mapping? It's defined by the game's code.) In a similar fashion, the building walls that we experience are themselves complicated mappings of the basic elements of reality. We don't experience decoherence as we split into the many worlds of quantum mechanics, nor do we experience the seven extra dimensions of M-theory beyond 3 spatial and one temporal dimension. As Donald D. Hoffman notes in "The Interface Theory of Perception": "Just as the icons of a PC's interface hide the complexity of the computer, so our perceptions usefully hide the complexity of the world, and guide adaptive behavior." The game character's perceptions likewise interface appropriately with the logical boundaries of its world.

"Real" entities are patterns that emerge from low-level physical particle interactions, and meaningful patterns can be seen in changes of currents and voltages inside microprocessors as well as in configurations of biomolecules inside organisms.

Goal-directed behavior and sentience

If video games can be seen as "real" in a similar way as our own world, what distinguishes video-game characters from real people and animals? I think it comes down to differences in complexity, especially with regard to specific algorithms that we associate with "sentience." As I've argued elsewhere, sentience is not a binary property but can be seen with varying degrees of clarity in a variety of systems. We can interpret video-game characters as having the barest rudiments of consciousness, such as when they reflect on their own state variables ("self-awareness"), report on state variables to make decisions in other parts of their program ("information broadcasting"), and select among possible actions to best achieve a goal ("imagination, planning, and decision making"). Granted, these procedures are vastly simpler than what happens in animals, but a faint outline is there. If human sentience is a boulder, present-day video-game characters might be a grain of sand.

Digital agents using biologically plausible cognitive algorithms seem most likely to warrant ethical consideration. This is especially true if they use reinforcement learning, have a way of representing positive and negative valence for different experiences, and broadcast this information in a manner that unifies different parts of their brains into a conscious collective. Yet, I find it plausible that other attributes of an organism matter at least a little bit as well, such as engaging in apparently goal-directed behavior, having a metric for "betterness vs. worseness" of its condition, and executing complex operations in response to environmental situations. Many NPCs in video games have some of these attributes, at least to a vanishing degree, even if most (thankfully) don't yet have frameworks for reinforcement learning or sophisticated emotion.

Many NPCs exhibit goal-directed behavior, even if that just means implementing a pathfinding algorithm or choosing in which direction to face in order to confront the game's protagonist.

Consider an example from a real video game: Doom 3 BFG edition (source code). In the file AI.cpp, we can see implementations of how AIs can KickObstacles, FaceEnemy, MoveToEnemyHeight, MoveToAttackPosition, etc. Of course, these functions are extremely simple compared with what happens if a real animal is engaged in an action. But the distinction is one of complexity, of degree rather than kind. Real animals' actions to kick obstacles, face enemies, etc. involve many more steps, body parts, and high-level systems, but at bottom they're built from the same kinds of components as we see here.

Especially in RPGs, some NPCs have explicit representations of their "welfare level" in the form of hit points (HP), and the NPCs implement at least crude rule-based actions aiming to preserve their HP. In some turn-based RPGs like Super Mario RPG or Pokémon, an NPC may even choose an action whose sole purpose is to bolster its defenses against damage in subsequent rounds of the battle. The extent of damage may affect action selection. For example, in Revenge of the Titans (source code), drones select a building to target based on a rating formula that incorporates HP damage:

rating = cost * (damage / newTarget.getMaxHitPoints()) * factor * distanceModifier; 

Even NPCs without explicit HP levels have an implicit degree of welfare, such as a binary flag for whether they've been killed. NPCs that require multiple strikes to be slain -- for instance, a boss who needs to be struck with a sword three times to die -- carry HP state information not exposed to the user. They also display scripted aversive reactions in response to damage.

And maybe representations of valuation could be seen more abstractly than in an explicit number like HP. In animal brains, values seem to be encoded by firing patterns of output nodes of certain neural networks. Why couldn't we also say that the patterns of state variables in an NPC encode its valuation? Animal stimulus valuation exists because of the flow-on effects that such valuation operations have on other parts of the brain. So why not regard variables or algorithms that trigger flow-on effects in NPCs as being a kind of at least implicit valuation?

Consciousness and power

A recurrent criticism of this piece is the observation that NPCs don't have sophisticated self-monitoring functionality of the type that would give them truly "conscious" experience rather than going through life "in the dark." I acknowledge that this complaint is relevant, and maybe I would agree that upon further examination, NPCs don't matter very much because their self-reflection is often not well refined. But we should think about how robust our view is that self-monitoring specifically is the crucial feature of emotional experience.

In "Feelings Integrate the Central Representation of Appraisal-driven Response Organization in Emotion" Klaus R. Scherer suggests five different components of emotion (this table is copied from Table 9.1, p. 138 of the text):


CNS = central nervous system; NES = neuro-endocrine system;
ANS = autonomic nervous system; SNS = somatic nervous system

Emotion function Organismic subsystem and major substrata Emotion component
Evaluation of objects and events Information processing (CNS) Cognitive component
System regulation Support (CNS; NES; ANS) Neurophysiological component
Preparation and direction of action Executive (CNS) Motivational component
Communication of reaction and behavioral intention Action (SNS) Motor expression component
Monitoring of internal state and organism-environment interaction Monitor (CNS) Subjective feeling component

We can see that emotion is a complex process with many parts at play. Why would we then privilege just the subjective-feeling component as the only one that we think matters? One reason might be that it's potentially the most central feature of the system. In Scherer's view, the monitoring process helps coordinate and organize the other systems. But then privileging it seems akin to suggesting that among a team of employees, only the leader who manages the others and watches their work has significance, and the workers themselves are irrelevant.b In any event, depending on how we define monitoring and coordination, these processes may happen at many levels, just like a corporate management pyramid has many layers.

Daniel Dennett's Consciousness Explained is one of the classic texts that challenges views separating "unconscious" from "conscious" processing. He says (p. 275): "I have insisted that there is no motivated way to draw a line dividing the events that are definitely 'in' consciousness from the events that stay forever 'outside' or 'beneath' consciousness."

I think the reason people place such emphasis on the monitoring component is that it's the part of the system that the centers of their brains that perform abstract analysis and linguistic cognition can see. When we introspect, we introspect using our self-monitoring apparatus, so naturally that's what we say matters to us. What our introspection can't see, we don't explicitly care about. But why couldn't the subsystems matter if we knew from neuroscience that they were there, even if we couldn't identify them by introspection? Intuitively we may not like this proposal, because we tend not to care about what we can't feel directly; "out of sight, out of mind." But if we take a more abstracted view of the situation, it becomes less clear why monitoring is so essential. Why couldn't we also care about a collection of workers that didn't have a manager/reporter?

I'm skeptical of higher-order theories of consciousness. I think it's plausible that people adopt these views because when they think about their emotions, of course in that moment they're having higher-order perceptions/thoughts about lower-order mental states. But that doesn't mean consciousness is always a higher-order phenomenon, even when you're not looking. To be sure, I think there is an important element of subjective experience that comes from monitoring, and metacognition probably contributes something to the texture of consciousness, but I'm doubtful about assertions that monitoring is all there is to consciousness. After all, what makes those computations special compared with others? One could say that so-called first-order perceptions are actually second-order perceptions about first-order objects in the world.

The focus on monitoring and reportability rather than the "unconscious" subcomponents seems to be an issue of power. The parts of us that think about moral issues are disposed to care about the monitoring functionality, in a similar way as the president of the United States gets all the attention by the media, even though it's his underlings who do most of the work. If one asks oneself whether one is conscious of low-level processes, one thinks "no, I'm not". Yet as blindsight, subliminal stimuli, and various other "unconscious" phenomena demonstrate, effects we can't uncover via explicit introspection still have effects and implicitly matter to us.

One way in which so-called "conscious" (globally available) information is distinct from "unconscious" information is that we can hold on to the conscious information for an indefinite time, as Stanislas Dehaene explains in Consciousness and the Brain. Our minds have what Daniel Dennett calls "echo chambers" that allow contents to be thought about and re-thought about and kept around for inspection and metacognition. Unconscious information comes and goes without leaving as much of a lasting footprint. This helps explain why we can only talk about the conscious data. But does only the conscious data matter? It too will eventually fade (unless it enters long-term memory, and even then we'll eventually die). Shouldn't relevance be assessed by the degree of impact something has on the system, not the binary fact of whether it can be reported on?

Consider also a split-brain patient. When presented with an object visible to only her right hemisphere, she says she can't see it; this is because linguistic areas are concentrated in the left hemisphere. But she can still reach for the object with the hand that her right hemisphere controls. It seems implausible to suggest that only the left half of the brain is conscious merely because only it can make linguistic reports.

Once we can see roughly how affective subsystems are operating in computational brain models, do we feel inclined to give more power to those voiceless subcomponents or do we want to continue to favor the conscious monitoring part that our instincts incline us to care about? I don't have a settled opinion on this, but I think we shouldn't jump to the conclusion that only the monitor matters. This question can be particularly relevant in cases where the non-conscious components diverge from the conscious summary or are not reported at all. And it's also relevant for agents like NPCs that may have only rudimentary monitoring going on relative to the amount of so-called "unconscious" computation.

In any case, NPCs do have extremely crude monitors of their internal status, such as display of hit points or verbalized exclamations they make (e.g., screaming). And we can see how the other components of emotion in the above table also show themselves in barebones ways within NPCs.

Dennett's stance levels

NPCs differ in their degree of sophistication. For example, consider the game "Squirrel Eat Squirrel" in Ch. 8 of Al Sweigart's "Making Games with Python & Pygame." The program for this game is extremely simple -- just 396 lines of code, not counting what's imported from Pygame. Squirrels of various sizes walk onto the screen from various directions. If they're smaller than the protagonist squirrel, the protagonist eats them; else, they hurt the protagonist. They may occasionally change direction at random, but this is all they do. They don't plan routes, seek to avoid damage, or even try to face the protagonist. They just move randomly, and if they hit the protagonist, either they're eaten or they cause injury.

The NPC squirrels in this game could be adequately described by what Daniel Dennett calls the "physical stance": behaviors comparable to particles moving around or other crude physical processes. (These are the "dumb particle updates" from which all higher-level stances appear as emergent phenomena.) For instance, the squirrel game could be seen to resemble planets in space. If a big planet like Earth runs into a small asteroid, the asteroid is absorbed and makes Earth somewhat bigger. If Earth runs into a planet bigger than itself, it gets blasted apart and damaged. We could think of plenty of other physical analogies, like water droplets collecting together.

In Super Mario Bros., Goombas and Koopa Troopas can be seen purely with the physical stance, because they just walk in one direction until they hit an obstacle, at which point they go the opposite way. A bouncing ball is comparable (indeed, arguably somewhat more complicated due to acceleration, friction, etc.).

Non-animal-like environmental features of games could be seen as comparably sophisticated, such as the dangerous water drips or falling rocks in Crystal Caves. Each level of Crystal Caves also features an air station, which explodes if the game's protagonist accidentally shoots it. The behavior of this air station is not much simpler than that of the most basic game enemies (after all, the air station "dies" if shot, although it doesn't move around or cause damage to the protagonist), but it can be modeled entirely with the physical stance.

Other NPCs are a little more complex. For instance, the Hammer Bros. in Super Mario Bros. turn to face Mario, and if Mario waits long enough, they begin to chase after Mario. This behavior could put them closer to the level of the "intentional stance" (they have beliefs about where Mario is located, and they have the goal of hammering him, so they move in order to position themselves to better achieve that goal). Many advanced game AIs are best described by the intentional stance, even if they don't use sophisticated academic AI algorithms.

Of course, that something should be modeled by the intentional stance doesn't obviously imply that we should care about it ethically. Many computing operations outside the realm of video games also seem to require the intentional stance. Load balancers, query optimizers, and many other optimization processes are best modeled by considering what the system knows, what it wants to achieve, and predicting that it will act so as to best achieve that goal (within the space of possibilities that it can consider). Search engines are probably best modeled by the intentional stance; given beliefs about what the user was looking for, they strive toward the goal of providing the best results (or, more pragmatically, results that will lead to satisfied-seeming user behavior). Outside the computing realm, we could see corporations as intentional agents, with beliefs about how to best achieve their goals of maximizing shareholder value.

If we're at least somewhat moved by the intentional stance, we might decide to care to a tiny degree about query optimizers and corporations as well as video-game AIs. I just wanted to make sure we realized what we're signing up for if we include intentional behavior on our list of ethically relevant characteristics. The fact that video-game characters look more like animals than a search engine does might bias our sentiments. (Or we might decide that superficial resemblance to animals is not totally morally irrelevant either.)

Does the intentional stance cover too much?

Peter Carruthers notes that on Dennett's view, "it is almost trivial that virtually all creatures [...] possess beliefs and desires, since the intentional stance is an undeniably useful one to adopt in respect of their behavior."

Consider these examples:

  • A washer is tied to a string. If you hold the string, the washer will always orient to face down towards the Earth. If you pull up the washer, it will fall back down if it can. You can put the washer on a table and "thwart" its efforts to hit the floor.
  • An electron is circling around a positive charge. If you push the electron away, it tries to go back toward the positive charge. (Thanks to a friend for inspiring this example.) In The House Bunny, a character says: "I want to be your girlfriend more than an electron wants to attach to a proton," which suggests that attributing a goal to the electron makes sense.

The boundaries between the physical, design, and intentional stances aren't perfectly clear. Maybe the difference is one of intelligence on the part of the intentional agent? For instance, electrons seeking protons aren't very smart at avoiding obstacles. But then again, neither are Hammer Bros. Still, ability to achieve one's goals isn't obviously crucial. Imagine that you have an overwhelming desire to discover the theory of everything that describes physics, but you can't even understand first-grade math due to a learning disability. Or suppose that your loved one has died, and you have an overwhelming desire to see her again, even though this isn't possible for you (due to not having reconstructive nanotech or whatever). Do your desires not matter very much in these cases because you're not smart enough to accomplish them?

Intentionality is not binary

In "Consciousness, explanatory inversion, and cognitive science", John Searle assumes as step 1 of an argument that "There is a distinction between intrinsic and as-if intentionality" (p. 586). "As-if" intentionality is the metaphorical kind that we use to ascribe goal-directed or mental states to a system, such as when we say that a lawn is thirsty for water. But, claims Searle, this is different from the intrinsic intentionality that animals like humans have when they're thirsty.

I maintain there is no fundamental divide. Goal-directed behavior is not different in kind from mechanical behavior; rather, goals are conceptual frames that we impose on mechanical systems. It's just that for some systems the intentional stance makes more sense than for others -- e.g., because the system is intelligent enough to still achieve its goals in the face of obstacles. In contrast, an apple whose "goal" of falling down is thwarted by the stem on which it hangs is not intelligent enough to extricate itself from the stem.

Searle compares the brain in our heads, which has intrinsic intentionality, with the "brain in our guts", which has as-if intentionality (p. 586-87):

Now does anyone think there is no principled difference between the gut brain and the brain brain? I have heard it said that both sorts of cases are the same; that it is all a matter of taking an "intentional stance" toward a system. But just try in real life to suppose that the "perception" and the "decision making" of the gut brain are no different from the real brain.

I don't claim there's no difference between the gut brain and the brain brain -- it's just that the difference is a matter of degree rather than kind. Distinctions between "mechanical" versus "goal-directed" systems are continuous gradations along dimensions of complexity, intelligence, robustness, and metacognition.

Searle claims that my view opens up a reductio (p. 587):

If you deny the distinction it turns out that everything in the universe has intentionality. Everything in the universe follows laws of nature, and for that reason everything behaves with a certain degree of regularity, and for that reason everything behaves as if it were following a rule, trying to carry out a certain project, acting in accordance with certain desires, and so on. For example, suppose I drop a stone. The stone tries to reach the center of the earth, because it wants to reach the center of the earth, and in so doing it follows the rule S = (1/2) g t2.

But Searle's modus tollens is my modus ponens. Searle has outlined one of my main points in this essay.

Searle encourages cognitive science to adopt an "inversion analogous to the one evolutionary biology imposes on preDarwinian animistic modes of explanation" (p. 585). He says simple cognitive processes like the vestibulo-ocular reflex don't deserve to be thought of as goal-directed. Rather than eyes moving with the goal of stabilizing vision, eyes just move mechanically, and evolution has selected mechanisms such that they achieve evolutionary ends (p. 591). But once again, I maintain there isn't obviously a sharp distinction between selection and goal-directed cognition. Suppose we're choosing where to go to college. This is a paradigm case of a goal-directed process: We have criteria and long-term plans, and we evaluate how well each college would score along those dimensions. But inside our brains, it's a mechanical-seeming selection process, in which inclinations toward different colleges compete, and the strongest wins. Likewise, an AI agent might engage in goal-directed behavior via an evolutionary subroutine to refine certain action parameters. In general, we could think of the "goal" as the fitness/objective function and "selection" as the cognitive processes that pick out a candidate that best achieves the goal. Both biological evolution and human-conscious goal-seeking involve both components.

Mysterious complexity

Often when a system is simple enough that people can understand how it works, people stop caring about it. When we can see something operating mechanically, we trigger the "inanimate object" viewpoint in our brains rather than the "other mind" viewpoint. Vivisectionists formerly cut open animals without anaesthesia because of the (technically correct though ethically misguided) belief that animals were "just machines."

The "AI effect" is similar: As soon as we understand how to solve a cognitive problem, that problem stops being "AI." As Larry Tesler said: "Intelligence is whatever machines haven't done yet."

Inversely, when a system is large, complicated, and arcane, it can feel to us like magic. Suddenly we can dismiss our prejudices against understandable processes and instead be awed by the mystery of the unknown. The fact that our brains are so big and messy is perhaps a main reason we attribute mysterious consciousness to our fellow humans. Because we haven't understood the mechanical steps involved, we can retain the "magic consciousness" feeling.

Many video games are bulky and massively complex as well. Here are some estimates from two articles:

Game Lines of code
Quake II 136,000
Dyad 193,000
Quake III 229,000
Doom 3 601,000
Rome 2 3,000,000

The article on Rome 2 adds that each of the game's soldiers have 6000-7000 polygons and 45 bones or moving parts. The game's lead designer, James Russell, explained: "Games are arguably the most sophisticated and complex forms of software out there these days."

Of course, code in video games has to run everything about the game world, and the amount of code devoted specifically to NPC behavior is a tiny fraction of the total. Still, even just the NPC systems may become messy enough that our intuitions that "this can't be sentient" give way to more magical ideas about emergence of consciousness.

We should bear in mind that codebase size is not equivalent to sophistication, intelligence, or moral importance. Some extremely powerful algorithms are really simple, while some trivial operations can be made complicated by adding a lot of frivolous or inelegant appendages.

Appearance vs. implementation

The main reason we don't care intrinsically about people being stabbed to death in movies or plays is that these media use deception to fool us. What looks like a sword in a man's chest in a play is actually a sword sticking under his armpit. When actors pretend to cry out in agony, their emotions are not actually distraught to a significant extent. In other words, what looks bad at first glance is not something we actually consider bad when we know all the facts.

Video games are somewhat different, because in their case, the NPCs really do "get hurt." NPCs really do lose health points, recoil, and die. It's just that the agents are sufficiently simple that we don't consider any single act of hurting them to be extremely serious, because they're missing so much of the texture of what it means for a biological animal to get hurt.

There is some degree of "acting" going on in video games when characters play out a scripted sequence of movements that would indicate conscious agony in animals but that is mainly a facade in the video game because the characters don't have the underlying machinery to compute why those actions are being performed. For instance, pain might cause a person to squint his eyes and scream because of specific wiring connecting pain to those movements. In a game, a character can be programmed to squint and scream without anything going on "under the hood." The usual correlation between external behavior and internal algorithms is broken, just as in the case of movie acting.

The same reasoning suggests that recordings of video games (such as screenshots and gameplay walkthroughs on YouTube) have a less significant moral status than playing the game itself, because the behaviors are all "acting," without any of the underlying agent-behavioral computations going on. That said, insofar as the image rendering of gameplay in a YouTube video still traces the outlines of what happened in the game, if only in an adumbral way, is there at least an infinitesimal amount of moral significance in the video alone? A more forceful example would help clarify: Is there any degree at all to which it would be wrong to repeatedly play a video of a person enduring a painful death, even if no one was watching it, if only because the pattern of pixels being generated has some extremely vague resemblance to the pattern of activity in the person's body during the experience? I don't know how I feel about this, but either way, I think mere videos of something bad don't rank high on the list of moral priorities.

I once saw a comment that said something to this effect: "I was going to 'like' this post, but it has 69 'like's already, and I don't want to ruin that!" This seems to be a relatively clear case where mere appearances bear essentially no relation to a moral evaluation. Whereas NPCs do execute extremely simple versions of the types of processes that comprise emotions in animals, the numbers "69" do not implement any obvious versions of sensory pleasure, not even in an extremely simple form. (Of course, one still might invent joke interpretations of those digits according to which they represent pleasure, but this is probably no more true for "69" than for "70".)

An argument against mere images having moral standing

Suppose we film an actor who feigned a painful death in a movie scene but was really enjoying the acting experience. The acting is very convincing, and the video itself is indistinguishable from a video of a real, painful death. In fact, suppose there exists another video of an actual, painful death that is pixel-by-pixel, frame-by-frame identical to the pretend one. Then if we replay the pretend video, are we "tracing the outlines" of the awful real death or the enjoyable fake one?

To some extent, this puzzle resembles the underdetermination of the meaning of a computation in general. We always have to decide what interpretation to apply to a set of symbol movements, though typically, unless the computational system is extremely simple, there's only one interpretation that's not hopelessly contorted in its symbol-to-meaning mappings.

Maybe an important feature of a morally relevant agent is counterfactual robustness. Hurting an NPC matters because if you hadn't hurt it, it wouldn't have recoiled, and if you had hurt it in a different place, it would have recoiled there rather than here. Another possible difference between interactive video games vs. fixed videos is that in an NPC, there are computations going on "under the hood" of the agent, in its "mind," telling it how to react. In contrast, video frames are collections of memorized pixel values. If the violent video were not statically recorded but dynamically computed based on some algorithm, at that point I might indeed start to become concerned.

An actor can enjoy the portrayal of being hurt because his brain includes extra machinery that allows him to know how to mimic injury without actually being injured; this machinery is not present in an NPC. Is there any extent to which portions of the actor's body do actually suffer a tiny bit by faking death, and they're just washed out in his brain's overall assessment?

I should add that sometimes actors do take on the moods of the characters they're playing to a nontrivial degree. Indeed, this is the premise behind the method acting school of thought. Two examples:

  • Angelina Jolie "preferred to stay in character in between scenes during many of her early films, and as a result had gained a reputation for being difficult to deal with."
  • When shooting The Aviator, Leonardo DiCaprio reawakened his childhood obsessive-compulsive disorder (OCD). According to Professor Jeffrey Schwartz: "There were times when it was getting not so easy for him to control it. By playing Hughes and giving into his own compulsions, Leo induced a more severe form of OCD in himself. There is strong experimental evidence this kind of switch can happen to actors who concentrate so hard on playing OCD sufferers."

Why visual display might matter

A few people do seem to feel moral revulsion about video-game violence, and presumably visual display is the reason for this: The images appear disgusting and may trigger our mirror neurons. Without visual display, very few people at present would care about NPCs. But the visual display seems "superficial," in the same way that a person's appearance doesn't matter compared with his inner feelings. So is visual display totally irrelevant for serious altruism?

One way we might think it matters is that it provides more unity to characters that would otherwise consist of lots of distributed operations by different parts of the CPU. The display localizes the effects of those operations to a set of pixels on a screen that evolves slowly in time, analogous to biological creatures being a set of atoms that evolves in time according to the brain's algorithms. We might think of the NPC as "being" those pixels, with its "brain" computations happening non-locally and then being sent to the pixel body. Of course, this is just one possible conceptualization.

The screen could also be seen as playing some roles of consciousness. For instance:

  • The global-workspace theory of consciousness proposes that some lower-level processes become "broadcast" to the rest of the brain in a serial fashion, and this corresponds to consciousness of that information. We could see the screen as playing a broadcasting function, and in fact, the global-workspace theory has been described using a theater metaphor, in which consciousness consists in what happens on stage, while lots of other computing may go on in the background as well. However, unlike in a global-workspace model, the screen doesn't have effects back on the brain itself. I think these effects back on the brain are pretty crucial, and indeed, Daniel Dennett considers consciousness to be the effects of "famous" or "powerful" brain coalitions on the rest of the system -- not just a broadcast of the information that doesn't do anything. We would not say "I'm conscious of X!" if the global broadcast about X didn't transmit to our verbal brain centers.
  • The higher-order theory postulates that consciousness consists in thoughts about one's mental states. A screen can be seen as a "thought about" the operations that the CPU is performing, where instead of being written in words, the screen's message is written in pictures.

I think both of these interpretations are rather weak and strained, but perhaps they can at least convince us that what the screen does may not be completely irrelevant to a moral evaluation.

How important are NPCs?

I think video-game NPCs matter a tiny amount, even the present-day versions that lack reinforcement learning or other "real AI" abilities. They matter more if they're more complex, intelligent, goal-directed, adaptive, responsive to rewards and injuries, and animal-like. However, I don't think current game NPCs compete in moral importance with animals, even insects, which are orders of magnitude more sophisticated in brain complexity and intelligence. Maybe game NPCs would be comparable to plants, which, like the NPCs, mostly engage in scripted responses to environmental conditions. Some NPCs are arguably more intelligent than plants (e.g., game AIs may use A-star search), although sometimes plants, fungi, etc. are at least as competent (e.g., slime molds solving shortest-path problems).

Fortunately, the number of video-game NPCs in the world is bounded by some small constant times the number of people in the world, since apart from game testing, NPCs are only computed when people play video games, and a given game contains at most a couple highly intelligent NPCs at once. The NPCs also run at about the same pace as human players. So it's unlikely game NPCs will be a highly morally significant issue just on the grounds of their present numerosity. (A more likely candidate for an algorithm that's significantly ethically problematic based primarily on a numerosity argument would be some silent, bulk process run at blazing speeds for instrumental purposes.)

At the same time, the levels of violence in video games are basically unparalleled, even compared against the spider-eat-insect and fish-eat-zooplankton natural world.

NPCs typically come into existence for a few seconds and then are injured to the point of death. It's good that the dumb masses of NPC enemies like Goombas may not be terribly ethically problematic anyway, and the more intelligent "boss-level" enemies that display sophisticated response behavior and require multiple injuries to be killed are rarer. Due to the inordinate amounts of carnage in video games, game NPCs may be many times more significant than, say, a small plant per second. Perhaps a few highly elaborate NPCs approach the significance of the dumbest insects.

On the other hand, it's important not to overstate the degree of suffering of NPCs either. Their decrements to health upon injury and efforts to stay alive aren't significantly different from those of more mundane computational processes found in other software or other complex systems. We may be tempted to give too much weight to video-game violence because of its goriness relative to what the algorithms behind the scenes actually warrant.

Video games of the future

If video games are not among the world's top problems at the moment, why talk about them? One reason is just that they provide a helpful case study with which to refine our intuitions regarding what characteristics make an agent ethically relevant. But a larger reason is because video games may become highly significant in the far future. As more sophisticated cognitive and affective components are added to NPCs in the coming decades, the moral importance of each NPC will grow. Likewise, the sheer number of video games might multiply by orders of magnitude, especially if humans colonize space and harness the stars for computational power.

Eliezer Yudkowsky's fun theory suggests that a valuable galactic future should involve ever-increasing challenges. Others like Nick Bostrom and Carl Shulman suggest a spreading of eudaemonic agents, who likewise value many rich, complex experiences.

Where would such fun, eudaemonic experiences take place? Presumably many would be within virtual-reality video-game worlds, filled with many NPCs. (Actually, in such a case, the distinction between player and non-player characters may break down.) These NPCs could include immense numbers of insects, wild animals, and imaginary creatures. In order to make these NPCs lifelike, they might consciously suffer in sophisticated ways when injured. We can hope that future game designers heed this as an ethical concern, but that possibility is far from guaranteed, and it may depend on whether we encourage compassion for wild animals, insects, etc. in the shorter term.

What about enemy aggression?

According to the top-voted comment on the main reddit discussion of my argument:

[NPCs'] primary goal is to oppose the player and foil his attempts to progress through the game. They take steps to stay alive because it makes them more effective as an obstacle. If the player has a goal, and the NPC's have a goal of preventing the player from achieving that goal, then there is a collision of rights in an environment with an intrinsic competitive mechanism for resolving the situation.

The obvious oversight in this argument is that the player is the person who set up the bloody arrangement to begin with. Suppose you build a colosseum and hire a rival to fight you to the death. During the battle, it seems you have the right to kill your opponent because he's trying to kill you. But from a broader perspective, we can see that it was you who created the whole fight in the first place and thereby ensured that someone would die.

Do player characters matter?

In this piece I've focused on NPCs because they're more autonomous, numerous, and abused in video games than player characters (PCs). But do PCs also have moral significance? It's natural to assume that PCs are subsumed within the moral weight we place on the player herself; PCs might be significant in the same degree as a spoon, shovel, or other extension of the player's body. Any planning, goal direction, or behavior of the PC would already be morally counted in the brain of the player.

Possibly there are ways in which the PC could count a little extra beyond the human player, such as through visual reactions to injury, autonomous responses based on HP level, or automatically orienting to face an enemy. Regardless, the experiences of the human player probably matter more.

Comparison with biocentrism

Biocentrism proposes that all living organisms -- animals, plants, fungi, etc. -- deserve ethical consideration because they have (at least implicit) goals and purposes. Paul Taylor suggests that each organism matters because it's a "teleological centre of life." His "The Ethics of Respect for Nature" explains:

Every organism, species population, and community of life has a good of its own which moral agents can intentionally further or damage by their actions. To say that an entity has a good of its own is simply to say that, without reference to any other entity, it can be benefited or harmed.

This captures a similar idea as my point that game NPCs have at least implicit goals that can be thwarted when we injure them. While this might seem like a separate criterion from sentience -- and Taylor clarifies that "the concept of a being's good is not coextensive with sentience or the capacity for feeling pain" -- I think the distinction between sentient and non-sentient teleology is one of degree. Even fairly trivial systems exhibit a vanishing degree of self-reflection, information broadcasting, and other properties often thought to be part of sentience.

Taylor's biocentric theory assumes that life is good for organisms:

What is good for an entity is what "does it good" in the sense of enhancing or preserving its life and well-being. What is bad for an entity is something that is detrimental to its life and well-being.

But if biological organisms have their goals thwarted enough (as is the case for most plants and animals that die shortly after birth in the wild, as well as for survivors enduring frequent hardship), maybe they would be better off not having existed at all. Whether we disvalue conscious suffering or so-called "unconscious" goal frustration, it seems a real possibility that the disvalue of nature could exceed the value. Hence, even a biocentric theory isn't obviously aligned with ecological preservation.

J. Baird Callicott pointed this out:

Biocentrism can lead its proponents to a revulsion toward nature—giving an ironic twist to Taylor's title, Respect for Nature—because nature seems as indifferent to the welfare of individual living beings as it is fecund. Schweitzer, for example, comments that

the great struggle for survival by which nature is maintained is a strange contradiction within itself. Creatures live at the expense of other creatures. Nature permits the most horrible cruelties.… Nature looks beautiful and marvelous when you view it from the outside. But when you read its pages like a book, it is horrible. (1969, p. 120)

So even if one doesn't adopt a traditional sentiocentric viewpoint, one can still feel that wild-animal suffering (or, more precisely, wild-organism goal-frustration) is perhaps an overridingly important moral issue.

Taylor admits that he is "leaving open the question of whether machines -- in particular, those which are not only goal-directed, but also self-regulating -- can properly be said to have a good of their own." He confines his discussion only to "natural ecosystems and their wild inhabitants." In general, environmental ethics leaves out the "machine question," as David Gunkel calls it. Some, like Luciano Floridi, have challenged this focus on the biological. Floridi proposes an "ontocentric" theory of "information ethics" that covers both biological and non-biological systems.

Does NPC suffering aggregate?

Some people have an intuition that small pains like a pinprick don't add up to outweigh big pains like torture. I sympathize with this sentiment, though I'm not sure of my final stance on the matter. There's a counterargument based on "continuity," which says that one prick with a 3 mm pin is less bad than thousands of pricks with a 2 mm pin, and one prick with a 4 mm pin is less bad than thousands with a 3 mm pin, and so on. Eventually it seems like one stab with a sword should be less bad than some insanely big number of pinpricks. I'm personally not sure suffering is continuous like this or whether I would declare some threshold below which suffering doesn't matter.c

If you did think that pains only add up if they exceed some threshold of intensity, then does NPC suffering not matter because it's so tiny? I personally think NPC suffering should add up even if you have a threshold view. The reason is that I don't picture NPC suffering as being like a dust speck in your eye. Rather, I picture it as being really awful for the NPC. If the NPC were smarter, it would trade away almost anything to avoid being shot or slain with a sword. The reason this quasi-suffering has low moral weight is because the NPC is exceedingly simple and has only the barest traces of consciousness. Even though NPC suffering ends up having small total importance, I still put these numbers in the category of "major suffering" (relative to the organism), so I think these numbers should still add in a way that pinpricks may not.

To put it another way, I would do aggregation as follows. Let wi be the moral weight given to organism i based on its complexity. Let the positive number Si be the suffering of that organism relative to its typical levels of suffering. Let T be a cutoff threshold such that pains less intense than the threshold don't matter. I propose to aggregate like this:

Σi wi * (if (Si > T) then return Si, else return 0)

rather than like this:

Σi (if (wi * Si > T) then return wi * Si, else return 0),

because in the latter case, if wi is small, wi * Si will never exceed T, so even an NPC in agony won't count given a non-tiny threshold T.

The second of these equations would make more sense if you hold a view according to which suffering is a tangible thing that can be quantified in some absolute way. If NPCs have lower moral weight, it must be because they suffer less in a universal sense. But then they should be like pinpricks and should fall below the threshold T.

The first of the equations makes more sense if you instead hold my perspective, according to which consciousness is not a thing but is an attribution we make. We decide how much consciousness we want to see in various physical processes. Interpersonal utility comparisons are not absolute but are ultimately arbitrary, and the wi in the equation is interpreted as an interpersonal-comparison weighting based on brain complexity, not as a "true" statement about how much absolute suffering is being experienced. The organism-relative suffering for, say, an NPC being shot is high, because the NPC would do almost anything to avoid that outcome if it had the opportunity. In other words, the von Neumann-Morgenstern (VNM) disutility of being shot is very negative relative to other outcomes for the NPC. (In practice, NPCs rarely obey the VNM axioms, but I'm speaking in a loose sense here, referring mainly to the idea of assigning utilities based on tradeoff choices. In any case, most people don't always obey VNM axioms either.)

I don't think it's right to imagine ourselves experiencing an extremely tiny pain when thinking about what it's like to be an NPC, because the NPC itself would trade away almost anything to avoid what it's experiencing (if it were capable of thinking about such tradeoffs). This suggests that intense suffering would be a more apt analogy than minor suffering when trying to put ourselves in the place of an NPC that's being killed. Pain and pleasure don't have objective magnitudes; rather, we make up how much we want to care about a given physical system. The only reason I give an NPC (much) less weight than I give a human is because I have a prejudice to care more about complex and intelligent things -- not because there's any objective sense in which a human suffers more than an NPC does.

Should we aggregate suffering at all?

Sometimes the objection against aggregation goes further: It's claimed that not only does NPC suffering not aggregate, but no suffering does, not even human suffering. C. S. Lewis expressed the idea in The Problem of Pain (p. 116):

There is no such thing as a sum of suffering, for no one suffers it. When we have reached the maximum that a single person can suffer, we have, no doubt, reached something very horrible, but we have reached all the suffering there ever can be in the universe. The addition of a million fellow-sufferers adds no more pain.

My first response is that our ethical calculations don't have to refer to something "out there" in the world. To say that two people suffering is twice as bad as one person suffering is just to say that we want to care twice as much in the former case. It's perfectly sensible to conceptualize our computational brains making calculations on the basis of aggregated suffering, or using computers to do so. A deontological prohibition against lying also doesn't refer to any object "out there" but is still a perfectly coherent moral principle. I find it intuitively obvious that ethical concern should scale linearly in the number of instances of harm, and this is all I need in order to be persuaded.

If that's not convincing, I'd make another point. It's claimed by the anti-aggregationists that the suffering of an individual is real, while aggregate suffering is a fictional construct. But this isn't the case. The only "real" things that exist are fundamental physical primitives -- quarks, leptons, bosons, etc., or perhaps more fundamentally, strings and branes. Anything else we may describe is a conceptual construct of these basic components. Molecules are helpful conceptualizations of more fundamental particles behaving in coherent ways together. Likewise with cells, organs, organisms, societies, planets, galaxies, and so on. When we talk about the suffering of an individual, we're referring to an abstraction: The collection of neurons of that person firing in certain patterns that we consider to constitute "suffering" in some classification scheme. I don't see why we couldn't apply a similar abstraction to describe the collective suffering of many organisms, viewed as a coherent group. Maybe one difference is that neurons in a brain are more integrated and interdependent than people in a society, but in that case, couldn't we at least aggregate the collective suffering of a close-knit human community when they, for example, experience a common loss? Like a collection of neurons, they are all responding to an input stimulus in ways that trigger aversion, and these responses interact to form a whole that can be seen as larger than its parts. To pump this intuition further, we could imagine hooking together the brains of all the community members, until even C. S. Lewis admits they form a single individual. Then slowly degrade the connections between community members. At what point do they stop being a single individual? Why can't their verbal and physical contact still count as communication within the collective brain?

For more on viewing consciousness at many layers of abstraction, see the next section.

Individuation vs. entitativity

The different components of a video-game character may be diffusely distributed. The NPC may have one set of functions controlling behavior, another set computing physics, another set performing graphics rendering, and so on. And perhaps some of these functions refer to shared libraries or DLLs, which are common to many different characters or objects in the game. So is each character really a separate individual, or are they all part of one big program?

Questions like these are not unique to video games. This section of another piece explores similar puzzles in biological contexts. Ultimately, there is just one multiverse, one unified reality. The way we slice it up when interpreting and describing it is up to us. Of course, some divisions seem more natural, like separating the United States from Russia at the Bering Strait. Many biological organisms can be separated based on physical localization, although slime molds, and even the bacteria in our guts, provide challenges to an absolutist approach of that type.

How do we count moral value if lower systems and higher systems can both be considered conscious? For instance, in the case of a China brain, both the individual Chinese citizens and the collective brain that they play out are conscious at the same time. We could further imagine the China brain as being one member of a larger Milky Way brain. And the brains of each Chinese citizen are themselves composed of many subcomponents that we may consider conscious in their own rights. How do we quantify the total amount of consciousness in such a system? The problem is analogous to quantifying the amount of "leaf-ness" in a fractal fern image. Do we just count the lowest layers, and let the big parts matter more because they have more smaller parts? But this ignores the way in which the big parts themselves form holistic units; a bunch of smaller parts arranged in a more random fashion would be less important. Do we just count the big brain as one more brain like the smaller brains? But presumably the China brain is more significant (or less significant??) than any individual Chinese citizen. I don't have a good proposal for how to proceed here. I do have an intuition that the lower-level components account for much of the total value, but this view might change.

These questions seem like something out of postmodernism or Hindu metaphysics,d but they have ethical significance. The reason our prevailing ethical paradigm for individuation focuses on humans and human-like animals as its central units of value is because it's humans who have the most power and make claims about their consciousness and self-unity -- not because humans are somehow ontologically special as a level of organization compared with lower and higher levels. We do also see some higher levels of organization assert their power and self-unity, including nation-states and corporations, though none of these has the machinery to organically generate the stance of claiming that "there's something it's like" to be itself, the way humans do.

Susan Blackmore describes the sense of self as an "illusion" that our brains construct. Elsewhere she clarifies that by "illusion" she doesn't mean to say that consciousness doesn't exist but only that it's not what we initially thought it was. Blackmore continues:

"Everything is one" claim mystics; "realising non-duality" is said to be the aim of Zen; "dropping the illusion of a separate self" is the outcome for many meditators. These claims, unlike paranormal ones, do not conflict with science, for the universe is indeed one, and the separate self is indeed an illusion.

See also ego death and egolessness.

Dennett's Cartesian theater can help us see the constructed nature of self-other distinctions. The Cartesian theater is the idea of "a crucial finish line or boundary somewhere in the brain" where formerly unconscious contents become conscious. The "lights turn on," so to speak. Neuroscientists may reject this explicitly. Yet many of them still talk as though the brain is conscious, while the external world is not. But in this case, the brain itself is "a crucial finish line or boundary" where formerly unconscious contents (e.g., photons or sound waves from the external environment) become conscious. How is this any different? Where in the environment-organism-environment feedback loop does consciousness "start"? It doesn't start anywhere; there's just a big loop of stuff happening. We call some parts more conscious than others because of constructions we're developing in our heads, not because of a discontinuity in the underlying physics of the world.

Case study: SimCity

In SimCity (2013), the player acts as the mayor of a city, working to build it and address needs as they arise. The game includes individual people, who move around using sidewalks, cars, and buses. The 2013 SimCity is an agent-driven simulation, so the dynamics of how the city evolves are complex and not computable from simple spreadsheet formulas, unlike what was the case in previous versions. That said, any given individual is relatively simple -- perhaps not much more complicated than water molecules moving through the capillaries of a tree.

However, the city as a whole can be seen as a tiny bit sentient. The city has a welfare level based on the happiness of its citizens -- and when the mayor does something good, smiley-face icons pop up in response. The city also has evaluation metrics on other dimensions, like property value and tax revenue. As agents interact with each other, problems arise, like not enough bus routes, buildings catching fire, too little electric power, congested roads, and so on. The game pops up notifications of these situations so that the player can choose to respond. Sometimes people in the game ask the mayor permission for something, such as holding a fireworks show, and the mayor can accept or deny the request. Citizens may hold protests if they become upset.

The city behaves kind of like a young baby -- growing, developing, carrying out its own functions, but demanding attention from its mother when it needs help and crying if it suffers. Alternatively, we could see the city like a domesticated plant that grows but needs assistance from its gardener, or else it runs into problems (too little water, fungal infection, etc.) and displays its ill health by various signs.

This example is also a case where the distinction between player and non-player characters becomes fuzzy. The game as a whole can be seen like a sentient NPC, but we could also interpret the joint player-game system as sentient. After all, the game can be viewed as "unconsciously" computing metrics, and then when something becomes important, the signal is "broadcast" to the player so that she can give attention to it, select actions in response, learn how to manage better, remember the results of past actions, and so on -- a process reminiscent of the LIDA cognitive architecture. On this interpretation, SimCity would be like an extension of the player's brain and could be valued in the same way we value other brain subcomponents.

Death within a larger program?

I think suffering rather than death per se is the main moral objection to video games, but those NPCs that can be said to have nontrivial goals often do try to avoid death. From a preference-utilitarian perspective, at least, this may make killing them somewhat bad.

One reply to this essay is that "Video game AI characters will all be run by essentially one entity [so] their deaths will be as meaningless as always." But let's compare this with another statement: "Animals are all part of essentially one entity (the world), so their deaths are meaningless." That a subsystem is part of a larger system doesn't make the subsystem unimportant. As the preceding discussion has shown, we really are part of the world in a deep sense, and we can't be extricated from it.

The same commenter's reply added another objection: "Also I have a hunch any characters with any money sunk on them will be respaw[n]ing for someone else." That NPCs can respawn is a point brought up by many of the reddit comments in reaction to my thesis. But compare respawning of NPCs to recycling of nutrients and biotic matter. After a person dies, she becomes food for bacteria and other organisms, and eventually some of her atoms are recycled into other animals, including people. She thereby "respawns" into new bodies. This doesn't change the moral status of her original death. Given that some NPCs try desperately to avoid being killed, is their dying and respawning fundamentally different? Maybe one difference has to do with personal identity: The NPCs come back "reincarnated" with their original identity, while a person who dies will never (except with vanishing probability in a given light cone) return as her exact same "self".

Sameness and empathy

In summarizing my argument on a discussion of these ideas, I said that humans and NPCs are both "physical processes playing out more or less elaborate computational algorithms." I noticed that I was appealing here to an underlying similarity in order to evoke empathy, and this inspired me to write the current section.

How broadly we choose to let our moral sympathies extend is up to us. We can adopt a more parochial view that focuses moral weight on very human-like minds, or at least those with very sophisticated cognitive and emotional abilities. Alternatively, we can adopt a more cosmopolitan approach that gives more widely distributed ethical consideration to more kinds of computational processes. How we draw our moral valuation function is not fully determined by scientific facts or logical consistency alone.

The empathy we extend to other beings seems to depend a lot on how much we see them as being similar to ourselves. In an interview with All In The Mind, James Doty explained:

So the key is how do you expand that in-group to a larger circle if you will? So that's an area of interest Dave DeSteno actually at Northeastern in the US, he's actually doing work that shows that if you can take somebody who you don't see as in your in-group and then look at them and try to sit there and say, is there one thing we have in common? And then you start trying to go down a list, and then you realise that there are probably multiple things. Each time you do that it actually decreases the sense of separateness.

Vittorio Gallese has written on the idea that empathy arises from neural systems that recognize the sameness of others. His "The Roots of Empathy: The Shared Manifold Hypothesis and the Neural Basis of Intersubjectivity" proposes a "multidimensional, 'we-centric' shared space" that humans use to think about their similarities to non-self members of a social group. This sense of sameness "is articulated on many different levels of complexity," and depending on the metric we use, "different species might score differently."

the [sameness] project aims to channel this feature of our psychology to engender altruism:

For every unique feature that makes you different from the 7 billion other people in this world, there is one feature that makes us the same -- we are all people. Humanity is our cause. It's quite simple really.

On the other hand, efforts to identify traits that are "uniquely human" represent attempts to highlight the differences between us and them. For instance:

Or people may identify traits to distinguish only some non-humans from the rest of nature:

  • Only mammals have a neocortex.
  • Only animals have neurons.
  • Only biological creatures have carbon-based organic molecules, DNA, protein manufacturing, etc.

We can think of the expanding circle as reflecting a shift from focusing on differences to focusing more and more on similarities with others. If you focus only on differences, then the only thing not different from you right now is ... you right now. Most organisms at least extend their scope of empathy to include their future selves, perhaps with diminished concern as those selves become more distant. Many species further have kin altruism, and some have reciprocal altruism with non-relatives. Some animals have "pure altruism" without expectation of reward, perhaps triggered by seeing their sameness with others of their species or, in some cases, a similar member of another species.

Going further, people often defend the moral importance of "lower" animals by appealing to their similarities with humans in cognitive ability, intelligence, emotion, neural components, behavior, phylogeny, and so on. As we go down lower, the similarities become weaker, since -- for instance -- insects do not in fact have the degree of cognitive sophistication that mammals and birds do. But we can still point to many areas of commonality. When discussing game NPCs, the similarities become fewer, though we can still appeal to the traits discussed in this piece, such as planning and state valuation. And as I observed at the beginning of this section, both humans and NPCs are physical processes playing out goal-directed computational algorithms.

We could potentially go further still and consider all of physics within the scope of our concern. We could appeal to the fact that we all share the same multiverse, operate according to the same physical laws (at least in a given universe of the multiverse), and so on.e I can see some merit to this view, though my heart tells me not to go for it because I have too many obligations to more animal-like agents that I care about more. In any case, even if we care about the whole multiverse, it's not clear what implications this has for agents that don't reveal either explicit or even implicit values or goals. So maybe the implications would not be very significant, although perhaps one could invent reasonable notions of what's better or worse for inanimate parts of physics. (Increased entropy? Reaching minimum-energy states?)

Anyway, I hope this section helps clarify the debate. I think a lot of the argument over whether, and how much, NPCs matter reflects differences in emphasis that people place on different traits. If you focus on very human-like traits, NPCs will be quite different from us. If you think about more shared traits, NPCs will appear more similar.

One might object that "our degree of moral concern for an agent" is not the same as "the agent's degree of sentience." But I think that when our degree of moral concern is based largely on mind-related traits of an agent, the two become basically the same questions, at least if "sentience" is to have morally laden meaning. If we define "sentience" by objective criteria then an agent's degree of sentience becomes a factual question. But most of us think of sentience as "the thing we're trying to get at when we talk about morally meaningful subjective experience," and in this case, tying it with our moral sympathies is the right approach. In any event, if you don't like this terminology, you're welcome to avoid it. The words we use don't change the substance of what we're talking about.

Empathy vs. aesthetics

We tend to care about the suffering of others because of feelings of empathy -- imagining ourselves suffering in their place. As the beings that we consider become more abstract, it becomes harder to imagine ourselves as them, at least without resorting to unwarranted anthropomorphism regarding what their inner lives are like. In the abstract cases, we need to rely more on general principles and concepts, like "goal-directed behavior," "welfare monitoring," and "information broadcasting." These are characteristics that we care about because we can see them in ourselves and because they seem at a theoretical level like the "right kinds of things" for what "sentience" should refer to.

When we take this stance of looking at the cognitive algorithms themselves and what they're doing, we stray somewhat from the realm of pure empathy and partly invoke aesthetic intuitions. For instance, it seems more consistent to weigh different systems to varying extents based on the degrees to which their various features seem "sentience-like." This elegance intuition may encourage us to see trace amounts of sentience even in places (like NPCs, potentially) where it would have seemed absurd to see it relative to our gut empathy intuitions.

To what extent should our moral intuitions be guided by gut empathy vs. abstract aesthetics? I think both parts matter. Unguided aesthetics might lead to a theory that values something like entropy maximization or computational complexity or some other abstract goal without much connection to what we think of as the feelings of others. I find it sad that environmentalism often invokes a person's sense of beauty to justify conservation without much thought to the suffering of the animals that have to live in those habitats. On the other hand, gut empathy can be capricious and arbitrary and may miss vast sources of suffering that we weren't trained to care about by our culture. Trying to balance both impulses is something I struggle with.

How low should we go in assessing sentience?

The logic of nature is very similar at many levels. Similar sorts of mathematical principles can describe physics, chemistry, biology, and sociology. Similar game theory applies both to unicellular bacteria and multinational corporations. Fundamental ideas of feedback control, adaptive response, and learning apply to both extremely simple and extraordinarily complex biological systems. All natural and social sciences have an underlying unity to them, because they describe the same reality, where similar patterns emerge both at small and large scales. Humans tend to place themselves at the center of ethical concern, but we are one tiny part of an enormous universe, and over time, we become increasingly aware of how much else may matter beyond ourselves.

How much ethical consideration should we extend to very abstract systems, possibly including parts of elementary physics that share some commonalities with properties of sentience that we consider relevant? I have mixed intuitions here. On the one hand, I feel like:

This is just intellectual distraction. What really matters are animals screaming out in pain, and that kind of clear, intense suffering is what we should focus on preventing. Forget the tiny computer algorithms and physical particles.

On the other hand, I think:

What about creatures like insects or zooplankton? Are they in the 'animals in severe pain' category or the 'intellectual distraction' category? I can't tell. And the natural world presents a full continuum of life's complexity, from prokaryotes up to humans. Where along that spectrum do you cut off? If you're not sure, might it be that the cutoff extends pretty far down? And given the enormous numbers at stake, perhaps the lower forms of sentience would dominate. After all, wouldn't altruism be vastly more important in a universe where morally relevant sentience was present everywhere rather than a universe where it was rare?

I don't really know where to end up on this spectrum. For now I follow a path that includes some focus on both normal suffering by large animals and weird views that contemplate more ubiquitous sentience.

What is it like to be an NPC?

Let's start with a more familiar question: "What is it like to be a dog?" A dog's brain has many similarities to our own, though it has some differences as well. Dogs have worse visual and better olfactory perception. Dogs can't speak rich language or do algebra. The ways in which information is perceived and processed by the dog's brain may differ. Fundamentally, a dog is its own kind of entity, operating according to its own modes of functioning. It has some similarities to us, but it's not literally the same.

I think we can still talk about "what it's like" in a poetic sense, with the goal of evoking a certain mental image or attitude in our listeners. When we say "there's a unique feeling of what it's like to be a dog," what we're trying to do is suggest: "Picture yourself in the skin of a dog, but then remember that what the dog sees and feels doesn't look quite as you imagine it." The phrase also aims to highlight the ethical point that "we can morally regard the dog in a similar way as we'd regard our own experiences, except with some modifications to account for how the dog is different from us."

"What is it like to be an NPC?" has a similar meaning, except that here the degree of difference is vastly greater than between humans and dogs. NPCs do not have almost all of the higher-level cognition or emotion processing that we have. They can't compose detailed thoughts about themselves, form complex associations, or respond in sophisticated ways. But they do have rudiments of input-output functions, behavioral reflexes, and possibly planning and adaptive behavior. What it's like to be this agent is very different from what it's like to be a human. An NPC's rudiments of consciousness feel very different from ours, so much that even using the word "consciousness" or "feel" can be misleading insofar as it evokes florid imagery that doesn't really apply.

What is it like to be an electron? Here the physical process is even more divorced from (and simpler than) what our higher-level algorithms embody. When we try to empathize with something else, we do so by imagination -- using parts of our own mind to simulate that mind.f But trying to simulate an electron with our brain is like trying to simulate a floating dust speck using an airplane. There's just so much extra irreducible machinery in our brains that we can't accurately reproduce "what the world is like" to the electron itself. For electrons we may have to use a more abstract valuation method.

Maybe this is the best way to think about the question of what it's like to be a very alien mind, like an NPC. Maybe it's not something our brains can understand, because an NPC is just too different. Rather than trying to imagine ourselves as an NPC, maybe the best we can do is realize that an NPC executes basic algorithms that begin to resemble morally relevant mental impulses and let that be enough.

Appendix: Evolution of my views on this topic

I never expected I would begin to care at all about video-game characters. If someone proposed the idea to me in the past, I would have written it off as foolish oversentimentality and anthropomorphism. It seemed just obvious that the reflexive rules and simple algorithms used by NPCs didn't feel like anything. In 2009, a friend raised the possibility that relatively uncomplicated agent-like Python scripts might have some moral significance, and I remember thinking to myself, "No, that's crazy. I don't care about silly Python scripts!"

My views at that time were distorted by a confused understanding of consciousness. Even after I explicitly rejected dualism and understood that it was up to us to decide what algorithms we wanted to call "sentient," I retained what we might call the ethical version of Dennett's Cartesian theater fallacy, in which I felt like the algorithms that gave rise to morally significant conscious emotion were relatively precise, localized, and binary in the sense of either being there or not. After all, that's how consciousness feels to us from the inside. But as I learned more neuroscience, I began to see how the brain's algorithms were not binary but existed along a continuum of complexity, and they could be seen to varying degrees in many computational systems.

"Still," I thought to myself, "the really important parts of subjective experience are things like reward/punishment, reinforcement, and consciousness. The trivial logic updates and planning algorithms of game AIs don't count as part of that." But gradually I began to see that even these kinds of distinctions were not clear-cut. For instance, under the LIDA model of consciousness, inspired by global-workspace theory, even extremely simple programming objects can be seen as marginally conscious insofar as information computed by some routines is bubbled up to be shared with other parts of the program and gets stored in the object's "memory" (state variables). Likewise, reward signals can come in varying shades of detail, the crudest being raw numerical values. And I realized that I would care some about animals even if they didn't learn but merely used genetically pre-programmed rules to determine their behavior. After all, when humans use cached reflexes and strongly entrenched habits to choose actions, we're doing something similar to game AIs executing fixed stimulus-response behavior.

It was becoming clear that a hard separation between sentient animals and simple, "unfeeling" computational agents could not stand. Fundamentally, these creatures lie on a continuum of traits along various dimensions. By early 2014, the implications of these insights for the case of video games began to sink in.

Optional appendix: My personal history with video games

I grew up in a household where video games were discouraged, out of concern for healthy child development. Still, when my family got a computer in the early 1990s, we bought a few dozen games for it, including both educational ones like Math Rescue and Micro! Deluxe and pure entertainment ones like Monster Bash and Cosmo's Cosmic Adventure. I spent probably hundreds of hours playing these.

When I was in elementary school, my best friend often played Nintendo Game Boy, and he encouraged me to try video games. In 4th grade, I got my own Super Nintendo system and played many of the standard games on it (Mario, Zelda, Donkey Kong Country, etc.). For our class's 5th-grade yearbook, we were supposed to pick what we wanted to be when we grew up, and because of my current interest in video games, I said "video-game designer." (Previously my list had included a paleontologist, geologist, and lawyer.)

My enthusiasm for video games continued throughout middle school, and while I was a good student, I spent most of my free time on Super Nintendo and Nintendo 64 games. I also sometimes played older puzzle or action games on my computer. I subscribed to Nintendo Power magazine and looked forward to its arrival every month. Video games were a huge part of my life, so much that when, for example, I accidentally deleted my ~160 hours' worth of accumulated experience points in Pokémon Yellow, I cried really hard and was sad for the rest of the night.

In 8th grade, I became inspired by Ralph Nader to work on making a difference for others. This reoriented my life's direction, such that video games now appeared an unjustifiable waste of time. I stopped playing them at the end of 8th grade, except for occasional computer games, and I sold my Nintendo games and systems for cumulatively ~$80 at the local game shop. I've rarely played video games since then except in special circumstances, though I do remember the old games quite well, having played them through so many times.

Video games were plausibly suboptimal for my development compared against more intellectual pursuits, but I don't think they were completely useless either. Perhaps the most important skill they taught me was persistence in the face of failure, since difficult games can be quite frustrating. It sometimes took many hours to beat a hard level, and occasionally I got permanently stuck until I was older, had more training, and could finally complete the game that had before been impossible. Hard video games are frankly a lot of work -- requiring more concentration and effort than most other tasks in life!

One of the reasons I enjoy software development so much is that programming feels like playing video games. Each task to complete is like a short level in the game, and you "win" the level when the new change you made works. Sometimes it's hard to get a library to install properly or to fix an obscure error, and this is like playing against a hard boss. Occasionally it takes many failed attempts to fix a bug, just like it sometimes takes many tries to get past a particularly tricky part of a game.

In Nov. 2008, I did a job interview with Microsoft, and originally I was suggested for the Xbox team. I had basically zero interest in video games at that point, so I asked for a different option -- hopefully Live Search (later Bing) because I wanted to use machine learning. I ended up interviewing with and joining Bing's ranking team instead.

I now have a renewed interest in video games because of their connection to deeper issues of cognitive science, sentience, and ethics -- though being interested is not the same as being supportive. Video games may be an intrinsic moral issue, and at the very least they represent a test bed for our views, helping us pick apart which components of agents we consider how morally important.

Feedback

Dylan Matthews published a popular piece in which I spoke about the moral status of video-game NPCs, among other things. That interview attracted a lot of attention and sparked the following replies and discussions. The statistics on Vox's article show that as of 13 Aug. 2014, the article was shared on LinkedIn 3 times, Twitter 171 times, and Facebook 1347 times.

Footnotes

  1. According to Wikipedia's article on NPCs: "Nearly always the connotation is that an NPC is allied with, or at least neutral toward, the player, while hostile characters are referred to as enemies, mobs or creeps." In this piece, my discussion focuses on enemy characters, because in many games enemies are by far the most numerous. In addition, gamers generally inflict more damage to enemies than to neutral NPCs, so enemy characters present the most important moral questions. I'm not sure if there's a better term that I should be using. I felt that "enemies" or "mobs" or "monsters" was too specific, e.g., because it rules out non-enemy NPCs, which may also be morally relevant.  (back)
  2. Consciousness is often compared with a CEO, and the analogy with corporate management seems apt. In Consciousness and the Brain, Stanislas Dehaene demonstrates that "In many respects, our mind's subliminal operations exceed its conscious achievements" (p. 86). The so-called unconscious mind does most of the work and can solve many problems on its own. However, it lacks the coordination among diffuse brain regions that consciousness allows. In a similar way, corporate employees do most of the work and can solve problems without their bosses, but cross-domain communication and goal-setting may be helped with the involvement of management. Dehaene makes this analogy explicitly on p. 92. On p. 88, he quotes Henry Moore, who said that consciousness "resolves conflicts, organizes memories, and prevents [a person] from trying to walk in two directions at the same time." A corporate executive (ideally) does many of the same tasks. She also serves as the "public face" of the corporation as seen by the outside world, in a similar way as conscious declarations are what outsiders know about our brains.

    Note that Daniel Dennett would caution against taking a CEO analogy literally as implying a full-blown homunculus: "anyone who goes hunting for the frontal display screen where the Boss keeps track of the projects he is controlling is on a wild goose chase" (Consciousness Explained, pp. 274-75). But the idea of a less personified region for coordination and conflict resolution remains valid.

      (back)

  3. One reason a threshold might be plausible is if brains show qualitatively different dynamics at some intensity of objective damage (say, between a 10 mm pin and an 11 mm pin). For instance, consider the following toy example:

    if(num_pain_neurons_firing < 1000) { algorithm_1(); }
    else { algorithm_2(); }

    This would represent a bifurcation in the dynamical system that is the brain. If we think algorithm 2 is vastly worse than algorithm 1, then there would be a discontinity at the 1000th pain neuron firing.

    While this is a stylized illustration, brains do exhibit bifurcations. For instance, in ch. 4 of Consciousness and the Brain, Stanislas Dehaene discusses how there seems to be a relatively binary threshold between when a stimulus remains unconscious and when it becomes globally broadcast. The difference results because of positive feedback that creates an avalanche of activity when the input exceeds some threshold of intensity. Perhaps there could be similar binary cutoffs in suffering dynamics at some level(s) of physical or psychological damage.

    Regardless, it remains an option for us to create a binary threshold in our valuations at some intensity of pain even if the brain's dynamics appear relatively continuous.

      (back)

  4. Or Spinoza, who said that "when we say that the human mind perceives this or that, we are saying nothing but that God [...] has this or that idea." Views like these should be seen as conceptual poetry on our part, not a literal phenomenon that produces weird, unexplained effects in the way the pseudoscientific Global Consciousness Project claims.  (back)
  5. In the movie Her (2013), Samantha the operating system reflects on "how bothered I was about all the ways that you [her physical boyfriend] and I are different. But then I started to think about all the ways that we're the same. Like, we're all made of matter. [...] We're all 13 billion years old."  (back)
  6. Some believe that we think about our own subjective experiences in a similar way. For instance, David Papineau suggests that when we imagine pain, our brains generate a faint trace of pain to help us conceptualize the experience. NPC "emotions" may be so divorced from ours that we can't even approximately simulate their experiences in our heads without anthropomorphizing. But this doesn't imply that NPCs don't have their own rudimentary forms of experience -- alien to our understanding but still morally relevant.  (back)