AI and the nature of play

In 2019 OpenAI Five not only won, but reportedly “steamrolled” the DOTA world champions playing the game. According to the OpenAI team the AI had played over 10,000 years of games against itself to achieve this feat. This sets up a good backdrop for us to gain insight to why is it we play games and how our relationship with games changes with the introduction of AI.

Why do we play?

Lions and other animals exhibit playful behavior that is often interpreted as training the skills they need later in the adulthood to survive. It would take a person with a peculiar world view to understand playing hide and seek or tag as preparing children for the challenges of adulthood. This hypothesis get even more difficult when we move on to playing with cards computers etc.

There is some evidence that especially roleplay and physical games do play an important role in the development of healthy individuals. Lately there has been more and more research on the effect of videogames and their cognitive benefits. How does an AI that wins every game it plays fit into this picture?

Games can be seen as a form of entertainment, just like watching a movie, but if we observe children at play, we immediately notice that even though games are played for fun, they are not taken lightly and there is a large investment of emotional and cognitive capital in every game where winners and losers are present.

There is also an observable trend even among adults to avoid playing with someone who always wins, preferring to play against those who offer challenge of roughly equal skill level. The people whose only motivation to play games is to get better at them, and thus welcome big hurdles, are few and far between.

In fact, observations in mice connect playful behavior to an understanding of fairness and morality by Dr. Jaan Panksepp. He studied wrestling mice and noticed that if the bigger mouse did not let the smaller mouse win at least around 30% of the time, the weaker mouse stopped playing with its bigger counterpart.

Before the event of machines, someone playing chess against the best in the world could firmly believe that there is actually no one capable of beating him in the game. At this point of time, you only need to log to any chess site for free to have a machine dismantle your strategies that took a lifetime to acquire.

This kind of capacity to steamroll human players raises interesting questions as to how our relationship with playing changes.

Drilling in

So far we have mentioned few traditional games such as hide and seek and tag, but AI, as it currently stands, has not really made it’s greatest contribution to gaming in the form of an advanced cheetah-bot that runs after children.

Most of the actual applications are in videogames, and even then, it’s not necessarily in the form of opponents to beat. Many different game systems that work behind the scenes could easily be categorised as having artificial intelligence. Systems that curate the experience such as those in Overwatch or Fortnite that make sure people are matched with other people of similar skill level.

Whether or not call these systems AI can be legitimately hard to discern, but their impact on the way we experience play is not.

In 2017 the Lego Foundation published a white paper(A white paper is a report or guide that informs readers concisely about a complex issue and presents the issuing body’s philosophy on the matter.) on the role of play in children’s development. Despite being published by the Lego Foundation its writers seem to have been level headed regarding their findings since they begin by stating that research on the field has not found firm causal links between play and enhanced “child outcomes”. So your child outcomes are left to fend for themselves without any necessary help from the Lego company and its products.

What makes the paper interesting is that it separates different forms of play as they relate to developmental outcomes. We might start by weighing how relevant is AI for those different forms and what can it tell us about our relationship with play.

The paper outlines 5 different forms of play: physical, playing with objects, symbolic play, pretend play and games with rules. Mind you this does not mean these elements can’t be found in the same game or that this list somehow defines precisely what games are, but that these are the separate aspects of play relevant to the development of children. Note that this sets up the definition playfulness and playing in somewhat teleological terms, meaning that we are on some level talking about inherent goal oriented activities for us as humans.

Next we will take a deeper look into the different forms of play and sketch a few ways in which they might interact with AI.

5 Forms of play

Physical play seems to have the least to do with AI, as cheetah bots are a thing of the future. The paper notes positive effects through exercise and rough’n’tumble play helping social skills later. On these social benefits the researchers note that it is not clear that they are tied the physical. Would making children play fighting games or other direct form of one on one competition like chess produce same social benefits?

These results read in a goal oriented manner could imply that play helps us train conflict resolution or competitive situations in a manner that is not harmful to our actual social standing. How does AI change this situation?

It is undoubtedly the case from my own experience that a match against another human being in a game is a different experience from playing against the computer. With an AI opponent you simply have no social baggage or the need to negotiate. There is no shortage of videos of angry players getting killed in a disrespectful manner by another human beings, where as the AI would just have shot you and not teabagged your corpse. Even if an AI teabagged you, it would not probably cause much anger since you would not read disrespect or belittling as a motivation behind the AI’s actions.

If social cues are put to the forefront it is worth asking what kind of differences in setting, ie. competing against a group of friends, against strangers or battling against AI, will bring in when we are talking about games’ social effect on the individual.

A related discussion surfaced when Fortnite started adding bots to matches to allegedly give players a bigger chance of a positive experience and keep them playing. Most people weren’t happy since they now can’t trust that their elimination of the whole enemy team was “real”. There is an obvious social and skill layer to the game that is being undermined by bots. Imagine telling to a friend about your star moment not being sure if it was just orchestrated by a group of bots.

Playing with objects is pretty much what it sounds. Manipulating and solving concrete problems. In one study playing freely with a tool was contrasted with teaching the proper use of a specific tool before engaging in a problem solving activity.

There was some indication that when playing freely with objects or tools before solving a problem the children persisted longer at the task and came up with better solutions to the problems at hand. Some spatial, mathematical and reading ability score correlations were present when children engaged in free construct play.

The white paper is pretty underwhelming when it comes to firm results on any of the topics, but it now gives us two new definitions of what playing might mean: manipulating objects and a free exploration of the capacity of tools.

Almost all games require almost by default some sort of 3D or 2D object manipulation in space, but apart again from other systemic and social settings curated by AI, object manipulation is not necessarily in itself that affected.

The second, free exploration of tools is pretty interesting since it somewhat overlaps what we conventionally think of as playing or being playful. It is also not a long reach to take it up a notch in abstraction and think of tools in terms of game mechanics or even in research as methods, settings, math etc. We are not the only animal who uses tools but we sure have stretched the definition the most. We are species capable of playing with ideas.

Interesting related research paper was covered by a youtube channel talking about AI and computer science research. The AI was programmed to be curious when playing games which in mathematical terms meant it maximized the number of unpredictable outcomes. In course of the study the AI learned to play the games better than AI just made to maximize game scores.

Though, it is easy for an AI to try 5000 million times to pass a level in a million different ways but humans usually require some direction and motivation to do so. When was the last time you suddenly freely explored what different things your toilet seat could be used for? Most of us don’t spend our curiosity by the hour on just any question. Setting and goals matter.

It is easy to map tool use on video games as well since they are in most cases almost defined and marketed by the mechanics they offer to the player. Fps, rpg, strategy, driving, sports etc.

On a concrete object manipulation level though, you could make the case that every game has the same mehanical input(the controller) so that is the object manipulation that counts.

Luckily we have some research that shows that it’s not only the use of controller that gets better when we manipulate virtual 3D objects and environments. Argument akin to that would be that solving math problems on paper only made our penmanship better.

You would be somewhat hard pressed to make the argument that manipulating abstractions is not an important part of being a human. I hope I don’t need references for this.

The goal oriented reading then says we play games to practise handling different abstractions. Even the most seemingly mindless shooter can have a deep strategy layer for those looking for it. Take for example this short video of a Doom shotgun mod and its use that probably has more honest thinking put into it than many college level essays.

It’s not necessarily AI in particular that will have an obvious effects on the use of tools in games rather than the game design itself. Artificial intelligence understood as a human like agent is probably most impactful when it comes to social interaction within a game.

Funnily enough when Fortnite added bots to their servers, many players started acting like bots to trick other players into letting them get close, thus giving them the advantage over more naive players. Thus changes in the social gaming environment were leveraged as tools to achieve goals inside the game.

Symbolic play can refer to the inclination to play with words, pictures, music, dancing etc. and ties heavily to self expression. There is something to say about symbolic, but I think some of it goes under the next form of play which according to the paper is the most studied, called pretend play.

In pretend play we imagine being someone we are currently not. Research did find some correlation with self control, executive function and pretend play in children. But it said nothing what would happen when the environment is the kind of ready made world many fantasy games have not conjured up by the children themselves. The paper says nothing on the kinds of enviroments where humiliation and trolling other players is the lifeblood of the game.

One notable thing in games is that pretending with other people is a different experience from pretending with AI characters. Saying people can act notably different anonymously online than in real life is at this point a bit of an understatement. I’m not sure we still have decided whether purposefully toxic online behavior is pretending or not, but we do know intuitively it doesn’t really exist if all the other players are bots.

In a way it is empathy and reading the minds of others that makes it possible and powers teabagging and griefing other players in game. You can make up your own mind on whether to see this as destructive as Kurt Vonnegut put it:

We are what we pretend to be, so we must be careful about what we pretend to be.

or a healthy way of letting out steam without dire consequences. Many cultures have had and have rites of passages and liminal states that let people veer off from moral customs in order then to integrate them more tightly into the moral fabric of the community.

A more universal and advanced form of this question was posed by the series Westworld. Specifically whether people killing or raping human like robots in the theme park had any moral consequences to them as people.

Currently I’m not convinced that Vonneguts point extends to outside of our social circles into gaming, but that would be a whole another essay in itself.

Games with rules is the last category of games that has a lot going for it. Rules in childrens’ play context are often inherently social so not surprisingly playing them seems to benefit childrens capacity for taking other peoples point of view, turn taking etc. Games with rules also work as scaffolding for learning, something that the reserachers note is an untapped field of study for enhanced learning outcomes.

Most of the information is heavily caveated by the researchers saying that there is not much research to back up any radical claims to any direction. Some positive outcomes on number processing when playing board games with that particular requirement are noted.

AI’s impact on social interaction is notable when coming up with community rules. Let’s take a dueling game For honor, which emphasises one on one duels between players. If you start you soon realize that the community has a set of guidelines and moral rules that affect your experience. Use a trick to kick someone down a well and not engaging them in a “fair fight” and they might soon let you know what they think of your mom.

Greeting other players before a duel is considered good manners, which some might take advantage of. Depending on your personal inclinations and your opponent your experience is wildly different.

Take for example as duel which you have begun in good faith by greeting your opponent which they take the advantage of by stabbing you. The rest of the engagement you let your opponent stand up when you knock them down and continue to act in good faith, but still win. You make your opponent pay a heavy moral cost, since you won without trying to take advantage of every trick in your arsenal.

This complex interaction would simply not be possible with AI, since it heavily relies on you reading you opponents mind.

Game mechanics themselves are also rules to be mastered but they’re not as heavily affected by AI understood as a human like agent.

Conclusion

Cheetah bots aside, AI might currently most heavily change our experience of social interactions and those behavioral aspects which rely on our understanding of other peoples minds during games.

Taking the example of the OpenAI FIVE we started with, the human players had hard time understanding some of the decisions made by the AI during their matches and always ended up losing the games by a margin. The AI could predict by the half way point if it would win or lose by over 90% accuracy even when the human counterparts thought they coud still win.

This on one hand opens up the way of improving our playing by analysing the games but also begs the question what is the point of even playing games against AI that can transcend human performance.

To sidestep this question we are required to look at gaming/playing as a worthy goal in itself that cultivates us as humans. This article has sketched some of these more qualitative aspects of gaming/playing and how the application AI might affect them.

In the end no matter how things might go forward, I firmly believe cheetah-bots are the future.