Over two decades have passed since IBM’s Deep Blue bested reigning world champion Garry Kasparov in chess, 3½ games to 2½. The reaction oft repeated since was that this heralded the onset of true machine intelligence: “It is a depressing day for humankind in general,” read the Guardian.1 The world eventually realized, however, that while the achievement was a remarkable feat of engineering, the computer wasn’t exactly playing chess—at least not the way that humans do. The computer showed no sign of intuition, no savvy, no real spark of creative thought.2 Blue may have won, but in retrospect its greatest accomplishment was perhaps not the solutions it provided on the chess board, but the questions it raised about games, creativity, and the nature of human play.
Fast forward 20 years: Google’s AlphaGo defeats Lee Sedol in Go, 4‐1. Go’s set of possible board states is massive compared to that of chess, meaning this was an altogether different situation. The brute‐force approach of Deep Blue simply wasn’t an option in the development of AlphaGo. Indeed, the programming behind this stunning display of AI research required a deft maneuvering of current-position evaluations and possible next moves, enabled by a complex combination of classical tree search and modern neural networks.3 Furthermore, some of AlphaGo’s moves set the world abuzz with talk of genuine insight and creativity. In the second match, AlphaGo’s 37th move surprised and stunned Sedol, not to mention the tens of millions of viewers watching live. The move, registered by many as a mistake at first, turned out to be a strategic play that turned the tide of the game in AlphaGo’s favor and revealed that the program had begun to approach the game of Go in an entirely new way, as of yet unstudied by human players. The move not only hadn’t been pre‐programmed, it had never even been imagined.
AI has made great strides in many other venues of gameplay since then. Programmers have used structured Bayesian networks to construct systems that can teach themselves optimal winning strategies for a variety of Atari games, in turn indicating the ability to learn causality and intuition.4 DeepStack and Liberatus, two poker‐playing AI, have shown how AI might succeed in games of imperfect information.5 Most recently, artificially intelligent systems playing the popular strategy game Dota 2 have successfully collaborated in order to defeat professional human players.6 One by one, the hurdles facing AI are falling, and the arena of gameplay is where some of the greatest battles have been fought.
While each new achievement crosses a different challenge off the list, one aspect of this research remains constant: A language of goals, competition, and problem‐solving dominate the development of game-playing AI. Concepts of winning and losing permeate the research, indicating that the strategies employed ultimately seek to abstract and compress the set of possible outcomes to the point that a winning answer—a correct answer—can reliably be chosen. In other words, current models of game‐playing AI aim to reduce the complexity of the playing scenario to the point that its solution space is computationally tractable.
Play, however, is about more than just winning and losing—in fact, it is about more than just games. While the domain of gameplay might serve as a productive armature for the discussions of optimized search techniques and complexity reduction strategies dominating artificial intelligence discourse, the topic of play has much more to offer, especially concerning creative design. In fact, a review of the literature on play indicates that it stands at odds with notions of optimization and goal orientation. Players exhibit delightfully un‐optimal behavior. They become curious, they get bored, they obsess, they give up—and while their actions may be impossible to predict, their behavior is still entirely attributable to relatively simple, computationally implementable causes.
In “Resisting Reduction,” Joi Ito calls for models of participant design that accept the irreducibility of the real world.7 Play enables designers to do just that. If we are to equip our computational agents with the degree of creative freedom utilized in design, we must craft a model of play that accepts—in fact, capitalizes on—the irreducibility of the real world. Doing so necessitates a closer look at play and an analysis of what makes it so fruitful for design.
Play, like intelligence or creativity, is what Marvin Minksy might have dubbed a suitcase word, an idea so broad and vague that any number of definitions can be stuffed inside. To tackle the term, it helps to adopt a structure similar to that used by computer scientists Stuart Russell and Peter Norvig in their exhaustive review of AI. In Artificial Intelligence: A Modern Approach, the authors use two lenses to separate AI into four categories: Thinking Humanly, Thinking Rationally, Acting Humanly, and Acting Rationally.8 With play, we adopt two slightly different lenses. The first is to consider whether the author is discussing the subject (the player) or the activity (the play). The second is to ask whether the author is empirically relating play’s qualities (observation —the corollary of weak AI) or surmising what is happening behind the scenes (inference—strong AI).
With these two questions in mind, we generate four categories of play research: Some authors study observable qualities, describing the behavior of the typical player or the notable attributes of the play activity. Meanwhile, others attempt to peek behind the curtain, inferring the unobservable intentions of the player or what they believe to be the true value of play (FIGURE 1).
Figure 1. Four categories of play research.
With these lenses providing focused categories, we can ask:
BEHAVIORS: What do players do?
Players display engaged, improvisational, exploratory behavior. The player actively tests and pushes against the particular set of rules or restrictions that make up her play space, occasionally adjusting the rules to change the play.9 The player sometimes moves spontaneously, improvising a step without knowing what will happen.10 Upon entering the play space, whether it’s a totally new space or just newly altered, the player commences exploring all of its corners and possibilities.11 To quote psychologist Mihaly Csizkszentmihalyi and sociologist Stith Bennett: “Play is going. It is what happens after all the decisions are made—when ‘let's go’ is the last thing one remembers.”12
ATTRIBUTES: What does play look like?
Play is governed by rules, yet it often appears quite irrational. These rules act as either restrictions or allowances, defining the type of activity that will be allowed during the play. Players can heed the rules or decide to break or change them, but each action is committed in direct response to the rules.13 The rules form the playground, or setting, for the play. Dutch historian and early play theorist Johan Huizinga refers to this as the magic circle of play.14
Within this space, play is entirely immersive. The playground is the player’s reality; the typical rules of everyday life have been substituted for the rules of play, making it unreal in the sense that the rules offer an alternative to the rules of reality.15 The resulting strict separation from reality frees the player from external constraint, making possible the exploratory, improvisational behavior mentioned previously.
INTENTIONS: Why do players play?
The player can play for a variety of reasons: to escape reality,16 to cope with ambiguity,17 or even to satisfy some external objective. Much of the pleasure in play lies not in its outcome, but in its engagement: in setting up rules and seeking new possibilities, in engaging alternate realities, and in suspending rational constraints.18
Most importantly, players play because they want to. Psychologist Jerome Bruner describes this quality as the player’s inner‐directedness.19 Immersed in play and freed from external judgment, the player is able to explore any and all solutions at his or her discretion, knowing that in the alternate reality of play, the normal consequences do not apply. Curiosity takes hold as the player investigates options simply to see what will happen.20
VALUE: What is the purpose of play?
Philosopher Hans‐Georg Gadamer notes that play exhibits a tendency for self‐renewal, indicating that play is entirely self‐fulfilling.21 This brings us to play’s most interesting characteristic: Play is conducted for play’s sake, and nothing more. This idea echoes throughout the research on play,22 but philosopher Hilde Hein summarizes it best: “to value [play] for its possible consequences is a denial of its essence.”23 As a consequence, play cannot be connected to external expectations. While it may prove useful to a host of cultural, social, and developmental functions, these results rely on play for their explanation, and not the reverse.
These characteristics are summarized in Figure 2, sorted according to the lenses applied.
Figure 2. Using the definition of play to address all four categories of play research.
At the risk of adding another definition to an overly crowded landscape:
Play is the curious behavior of an autotelic subject temporarily exploring a system of rules.
This definition concisely responds to each of the four categories described earlier. The behavior of the subject at play is explicitly described as curious exploration, implying the nonlinear, improvisational search for new possibility that dominates the play behavior. This exploration of rules also illuminates the attributes of play, as it is the instantiation of the rules of play as a collective system that implies new ground for exploration, and new possibilities to discover.
But most importantly, the intentions of the player and the value of play are accounted for using a single term. Play is described as autotelic, meaning it is internally‐motivated and self‐fulfilling. As will be discussed later, it is the internal drive of autotelism that enables the player to explore with the level of freedom and curiosity that we see in full‐fledged play.
Curiosity, engagement, immersion, self‐interestedness: These qualities mark not just the player, but the creative individual as well. Designers, artists, musicians, and craftspeople all play, not just as a means to an end, but because it is in play that creative exploration is richest. And what result as intentional behaviors must be practiced for years: Discussing the difficulties of teaching novice jazz musicians to improvise, anthropologist Eitan Wilf describes how students struggle to resist using the techniques (or licks) they have memorized, even if these techniques fit within the typical vocabulary of jazz. To overcome this, teachers often instruct their students to adopt unfamiliar instruments and “rearrangement(s) of their corporeal schemata,” opening up the creative space and affording truly new methods of improvisation.24
As Pip Mothersill notes in “Inviting Feedback,” it is this “messiness and openness to new possibilities,” this “slow, convoluted, surprising friction,” that appeals to designers in particular.25 Design rarely follows a clear or straightforward path. Instead, it is an iterative, meandering search for interesting questions and possible values in an ambiguous and dynamic space of possibilities. Designers make their way through this indeterminate space, often with very little guidance. Play frees the designer to explore without considering consequence. For designers, the playful path is a meandering one; at times unpredictable enough that it surprises even the designer. But these behaviors are not unfortunate byproducts that the experienced learn to ignore. Instead, they are core components in the expert designer’s creative practice. To quote architecture professor Malcolm McCullough, “The master at play improvises.”26
The image of designers‐at‐play emerging here is quite different from the depiction of play borne by most modern game‐playing AI. True, in both cases we see exploration, inquiry, perhaps even flickers of creativity, but while some of the observable behaviors and attributes may appear the same, there exists an important difference: The play that designers exhibit indicates a willingness to change and an embrace of complexity that goes beyond traditional notions of goal‐oriented problem‐solving.
To tease this distinction out, consider a parallel dialogue adopted by craftsman David Pye. In The Nature and Art of Workmanship, Pye discusses the difference between the workmanship of certainty and the workmanship of risk. The former, he argues, is an attempt to predetermine the result of the work before it ever begins. Pye uses the example of printing, where the process is unified to the point that each produced print is an exact copy of the others.27 The certainty in the process is absolute, and the predetermined result irons out variegation in favor of reproducible content. Anthropologist Tim Ingold has another term for certainty: the mechanical ideal.
…while I would concede that the perfect machine is an ideal that cannot be realised in practice, the mechanical ideal is nevertheless driven by an aspiration of systemic closure. While in operation, the machine is designed to be as exact as possible in the execution of a course determined by settings fixed in advance. It should not feel anything even though perhaps it does.28
Contrasting these depictions, Pye describes the workmanship of risk as:
...workmanship…in which the quality of the result is not predetermined, but depends on the judgment, dexterity, and care which the maker exercises as he works... The workmanship of risk has no exclusive prerogative of quality. What it has exclusively is an immensely various range of qualities, without which at its command the art of design becomes arid and impoverished.29
The workman incorporates risk intentionally, even when it might be avoided, knowing that doing so opens up a wealth of creative avenues. This distinction can be difficult to spot, as it exists regardless of the activity engaged. Thus, whereas the novice jazz player turns to the certainty of known licks to mimic a risky performance, the expert classical musician breathes risky new life into a seemingly certain set of notes. The difference is one of disposition: It is the orientation to the work, regardless of final outcome or possibility.
Play occupies a critical role in our design process, but this does not necessarily make it difficult to model. In fact, it is entirely possible to generate playful behavior in autonomous computational agents. Computational play is a proposed framework for this implementation.30 It adopts as its motivation two core beliefs: First, that creative design tasks will eventually be within reach of autonomous, artificially intelligent systems, and second, that creative design processes involve more than the optimized search strategies and complexity reduction techniques currently dominating artificial intelligence discourse (and, in turn, modern game‐playing AI). The framework is a step toward equipping autonomous computational agents with the freedom they need to engage in play‐driven creative design.
The first component of the framework stipulates that computational play is facilitated by multimodal representation conducted by the player. This can take a variety of forms, but the process must be significant enough that it requires the player to translate its own work between different modalities, reinterpreting the content along the way. An example is the designer’s tendency to sketch. This externalizing process is key to the designer’s ability to shed a cognitive bias and approach a topic from a different perspective.31 Insisting on multiple modes of representation acknowledges that the process of translation often ignites our subconscious, causing the kind of insight that often proves so fruitful to the design process.
Second, computational play is iterative. Combined with the insistence upon multimodal representation, this lends play a cyclical structure. Iteration opens the design process up to shifting contexts and changing requirements, granting the same level of importance to both question and problem. At the same time, it also results in a natural discretization of the creative process that is conveniently applicable to computation.
The third characteristic of computational play is its focus on generative techniques that do not rely upon results for justification. Players—as well as designers, artisans, craftspersons, and many other creative professionals—improvise. With this, they demonstrate a tendency to act first and evaluate second. This unrestricted generation of content yields material that can then be incorporated in the next iteration of the design process.
It is not enough to simulate playful behavior through observable qualities, however. We must also come to terms with the inferred, the personal, the very subjective nature of play: its autotelism, or reliance on internal motivation. As the activity of an autotelic subject, play is pursued for play’s sake, leading to an autonomy and independence from external constraint.32 Designers rely on this detachment from external concern in order to explore illogical ends, non‐optimal tangents, and personal pursuits. An autotelic designer engaging in true play is able to embrace risk in ways that enable her to deal creatively with life’s complexities. Driven by autotelism, play stays the whip, and resists the reduction of the creative process to goal‐oriented problem‐solving.
These characteristics are displayed in Figure 3, using the same structure as before, albeit with play and player merged into a single lens.
player - at - play
Figure 3. Computational play framework. The player and play are merged into one category, and both observable and inferred characteristics are addressed.
Autotelism, then, is what lends play the freeing qualities that cement it as central to creative mastery. Autotelism internalizes all of the decision‐making criteria in a playful process, requiring the player to decide if playing should continue, regardless of any external consideration. But if there is no guiding model, what does the player use to make the decision?
Divorced from any external heuristic, the first parameter in a computational model of autotelism is simply the size of the remaining possibility space. What possibilities are left uncovered? As the number of possibilities increases, so does the likelihood of play continuing. This captures designers’ natural curiosity about uncharted possibilities and reflects the tendency of the playful designer to explore simply for the sake of exploring.
In reality, playfully creative processes aren’t nearly so linear or predictable. This relationship (between the number of possibilities and the likelihood of continuing) paints only one part of the picture and is still just an objective analysis. Since computational play necessitates the role of a player, any model of autotelism must also take into consideration the subjectivity of the agent. In other words, we must also consider whether or not the player wants to keep playing, and acknowledge that this parameter might exist entirely independent of the number of possibilities.
Combined, these two parameters form a 2D graph of the playful designer’s decision space, in which the X‐axis represents the objective size of the possibility space, while the Y‐axis represents the subjective interest level of the agent (Figures 4, 5, 6). The brightness of any point on the graph represents the likelihood that the subject will continue playing, from black (quit) to white (continue). A medium‐grey value indicates that the decision may go either way.
The predictability of this space is fairly straightforward: If the player has little interest in the task and few available possibilities, the play will likely be terminated, as indicated by the dark lower left‐hand corner of Figure 4. On the other hand, if the interest is high and the possibilities are many, the play is likely to continue, as indicated by the bright upper right‐hand corner. The opposite corners, where one value is high but the other is low, balance out to an unpredictable mid‐grey, indicating that the decision could go either way. This is the same in the center of the graph, where both variables have neutral values and no reliable prediction can be made.
What the graph also captures, however, is a measure of intentionality on the part of the player. Although not all are entirely predictable, decisions that land in any of the four corners are easily attributable, as the extreme values (from one or both axes) yield straightforward cues to behavioral choices. Only in the center, where neither axis holds significant sway over the decision, does attribution dissipate, and chance dominate (Figure 5).
In both the upper left‐hand and the lower right‐hand corners, the mismatch between the size of the possibility space and the interest level of the subject results in a kind of volatile frustration. The behavior here is unpredictable, demonstrated by the medium‐brightness that results from the opposition between the two variables, but entirely attributable, due to the presence of one extremely high variable. Adding a blue‐green color gradient along this diagonal axis helps visualize this much richer field of possible decision‐making behavior (Figure 6).
In the upper left‐hand corner, frustration results from a dedicated player whose possible options are slim. In this situation, for the sake of continuing the play, the player might justify a number of behaviors that defy more rational behavior: She might choose to start the process over, or ignore part of the problem in order to focus on a particular issue. The player might continue with a process that seems particularly fruitless, knowing that it may yield new insights in the future. She might improvise a totally new step, even without believing that it holds any value. And finally, the player may break the rules, modifying her playground in order to open up new possibilities. In the lower right corner, on the other hand, when the interest is low but the possibility space is high, this frustration could have the opposite effect. The player may give up, quitting the drawing process despite the available avenues. While such an avenue might seem undesirable to the outside observer, it is just this detachment that gives play value.
As designers, our curiosity propels us to experiment with an idea, even if there’s no need to do so. Designers also frequently suspend external considerations in favor of personal interests. Such actions – of prioritizing exploration over optimization, subjective interests over objective pursuits, risk over certainty – imbue design with creative potential and demand an autotelic and playful approach. If we expect our machines to play, and one day design, with the same degree of creative freedom that we enjoy, then we must enable the subjective, unpredictable, attributable behaviors that emerge from autotelism and push creative design.
Computational play can indeed be implemented in autonomous machines. In fact, the basic framework is not difficult to model – for some preliminary results from a proof‐of‐concept implementation in an autonomous drawing machine, see . Granted, actually implementing autotelism is a tricky notion. Defining machine subjectivity is a challenging task, and one that many believe should be tackled only after other aspects of intelligence have been mastered. But we must not wait to perfect the subjectivity of autonomous computational agents before we explore its role in the creative process.
Play doesn’t always make sense, and the resulting behaviors may seem odd at first: Why keep searching when a good answer has already been found? Why pursue a wrong answer? Why break the rules? Instead of demanding a useful answer from such questions, I find it illuminating to ask a different set: What motivates us to proceed when we get stuck? What do we do when we confront something we don’t understand? How can we overcome the desire to rely upon answers that already exist? And finally, what opportunities will we uncover when we let loose our creative curiosity and pursue novelty for novelty’s sake? Such questions are grounded in an optimistic, and playful, view of design: One that believes there is value in the process aside from the results, and that there are, and always will be, interesting things to discover. The world of design is not a fixed and definite place, waiting to be fully understood, properly optimized, and successfully predicted. It is a dynamic and diverse landscape, full of interesting surprises that will challenge our understandings for the better. Certainty is the whip that drives play to a premature conclusion, and it is a poor foundation on which to model our understanding of play’s role in the creative process. True mastery of play behavior, whether in a human or a machine, shuns that drive in favor of the boundless creative potential afforded by curiosity, exploration, and risk.