Twitch,Fish,Pokémon and Plumbers: Game live streaming by nonhuman actors

Existing Research

Ideas of agency and hence interactivity were instrumental in the initial formulations of game studies (Ruffino, 2020:15), and agency is thus now a well-theorised concept within the field (Jennings, 2019). Without intending to contribute to discourse surrounding definitions of agency, we broadly take Murray’s (2016:159) description of agency as ‘the satisfying power to take meaningful action and see the results of our decisions and choices’. Such a framing seems readily applicable to games in which players are given various choices to make and then experience the effects of these choices on the game world. However, while offering a thorough overview of the topic, Girina and Jung (2019):11 note that the idea of agency can become a ‘surrogate’ for play, with an emphasis on the ‘myth of choice and by the abundance of paths available to players’ replacing an understanding of games’ ‘non-negotiable algorithmic nature’. Which is to say: can game players be said to possess agency if they play only within confines laid out by a game’s designers? Game structures influence the decisions that players can make and hence compromise their agency (Consalvo et al, 2019; cf. Girina and Jung, 2019), marking out games as complex spaces for assessing what agency truly entails. Metagaming practices that undermine these rules of playing, such as speedrunning or glitching, complicate this even further (Boluk and Lemieux, 2017).

Yet, what of the agency of nonhumans? In considering the agency of technological nonhumans, as others have noted (Švelch, 2020), there is little critical or game studies scholarship on game AI or game systems, but an initial appraisal of their ‘agency’ is possible. Perhaps, the first critical consideration of nonhuman ludic agency was that of Giddings (2005) who examined the cybernetic logics of games as simulated systems; any interaction with a game system inevitably involves interacting with a nonhuman agent (Ruffino, 2020:16). Fizek (2018) in turn encourages us to think seriously about such concepts in the context of ‘movement-simulating bots, self-acting nonhuman agents, or game worlds changing without the direct input from the human player’ (Ibid:203). Her analysis presents us with frameworks and examples for thinking about how to move beyond humans as the ‘sole meaningful agents’ in games (Ibid:206), and we will draw upon this work several times throughout our paper as we extend such an understanding into the realm of live streaming. We additionally note Ouellette & Conway’s paper in which they question whether ‘the phenomenon of play [is] even an ontological possibility’ for a computer (Ouellette and Conway, 2020:11), something we will address in considering the difference between an AI actively playing a game, and an algorithmic system incidentally enabling play to occur. Yannakakis & Togelius’s work (Yannakakis and Togelius, 2018) identifies limits in what games AIs are able to play, particularly when it comes to developing and implementing strategy in strategy games, which they link with intention (2018:135), a topic we will return to in this paper.

In the literature on biological nonhumans, meanwhile, there are major bodies of scholarship on nonhuman animal play and considerations of animal agency, although these two do not appear to be especially well unified. Although of course humans are ourselves animals, we will simply use the term ‘animals’ from this point to refer to biological nonhumans. The study of animal play is vast but evolutionary approaches are unsurprisingly common (Bekoff and Byers, 1998; Burghardt, 2005; etc), highlighting the many beneficial effects of play for a wide range of species. In terms of animal agency, Špinka (2019) proposes agency as central to animals’ lives since they must consistently navigate complex or unpredictable environments as well as the agency of other living things, while Jamieson (Jamieson, 2018:111) notes, perhaps most crucially, the overwhelming evidence that ‘most of the properties that were once thought to distinguish humans from other animals are shared with other animals’. In combining the two it does not therefore seem unreasonable to suggest animals enact agency when engaging in play, although this – as we will return to later – would require an animal to be aware that play is taking place.

We now move to live streaming, the practice of playing games live for an audience over the Internet, facilitated by platforms such as Twitch. These platforms have developed their own cultural practices, value systems, and arrangements of collective identity (Jackson, 2020), and as part of the ongoing growth of live streaming we have begun to see research addressing questions of agency. For example, Taylor’s (2018:80) work considers live streaming as an ‘assemblage’ in which ‘a variety of actors (human and nonhuman)’ actively contribute to making ‘play, performance, and work possible’ on live streaming platforms. This is echoed by Egliston’s (2020:243) observation that ‘Twitch is a useful tool for visualising the distributed agencies at play in videogaming’. The first author of this current paper meanwhile has studied the economic and professional agency of financially successful live streamers, identifying how live streaming offers compelling careers to individuals from a tremendous range of backgrounds often lacking in social or economic capital (Johnson and Woodcock, 2017), and in doing so enables self-actualisation and control over one’s life in an increasingly challenging job market for young people (Johnson et al., 2019). Partin (2020) has studied struggles between the users and the Twitch platform, highlighting how the platform’s ‘agency’ does much to profoundly shape and structure the experiences and potential interactions of its users. Finally, Orme (2021) alludes to yet another form of agency we will discuss – spectatorial agency – by showing how those who might feel excluded from gaming culture nevertheless exercise their desire to experience gaming in some form via Twitch spectating.

Considering agency in live streaming, we particularly note research on the famous Twitch Plays Pokémon (TPP) phenomenon. This is an important background for the analyses we present here. In 2014, (many tens of thousands of) Twitch viewers input text commands that were then received and implemented by the version of Pokémon Red played on this stream. This was achieved through a custom IRC (Internet Relay Chat) bot that could make sense of terms like ‘up’ or ‘right’ typed into the chat window (Ramirez et al, 2014). The result was a gameplay session where the Pokémon Red in-game player character—swamped by hundreds of inputs a minute complicated by the time delay between text and stream, reducing the extent to which viewers could send ‘useful’ inputs – moved back and forward, behaved erratically, constantly performed useless actions, and (very rarely) progressed in the game. The uniqueness of this broadcast and the curious behaviour of the player character led to the emergence of a tremendous volume of attention in the gaming press (e.g. Farokhmanesh, 2014; Makuch and Haywald, 2014) and wider publications including major newspapers (e.g. O’Mara, 2014; Vincent, 2014), leaving an indelible mark on gaming and on live streaming culture.

In attempting to theorise the phenomenon numerous proposals were soon put forward by scholars. For example, Margel (2014) defines TPP as a ‘crowdsourced’ game, Mallory (2014) termed it a form of ‘community-based play’, while Lessel et al (2017) instead opted for the phrase ‘user-generated gaming’ in trying to pin down the nature of the stream. We do not disagree with any of these, yet they all focus on the actions of the thousands of users who were influencing the game’s play, in the process omitting the concomitant observation that play was not influenced by any streamer. The game was ultimately played by a computer system designed to take account of thousands of individuals’ instructions and process them. TPP is hence a nonhuman stream, but also highlights an important point: that a mass of human agents are not themselves a human agent. TPP marks the highly visible emergence of the nonhuman gaming stream, but here we move beyond TPP as a stream controlled by a nonhuman agent consisting of thousands of humans, towards nonhuman agents which lack even that degree of human involvement. As Egliston (2020:242) states, Twitch is an extremely promising space for ‘studying how [gameplay] emerges from assemblages of human and nonhuman’, which we will demonstrate in this paper through four case study streams.

Lastly, we should also note that there are many ‘expectations’ or ‘assumptions’ about what a game live stream should entail or what a game live stream requires, especially in terms of labour, which nonhuman streams complicate significantly. Some these are specified by Twitch through documents such as their ‘community guidelines’ or ‘terms of service’ (TOS) which make requirements that streamers not broadcast illegal content, for example, and maintain a reasonably pleasant community (cf. Ruberg, 2021). Equally, while there is no formal requirement to interact with one’s viewers, this is one of the strongest implicit norms in live streaming, and indeed one of the things that most significantly marks it out from other media formats (Sjöblom and Hamari, 2017). There are also significant labour demands for human game streamers (Johnson and Woodcock, 2017; Johnson et al., 2019) who are expected to maintain such relationships with viewers, be entertaining or witty, develop stream aesthetics and the detail and distinctiveness of their channels, stream for long hours – especially if pursuing financial ‘success’ on the platform – and so forth. These labour demands are interesting to consider when we think about what kind of labour – if any – is being performed on these channels, and what nonhuman streaming labour means more broadly for questions of leisure, automation, and nonhuman forms of work.

Case Studies

The following case studies – each with no human streamer – were selected for their variety in terms of both the actor controlling or influencing gameplay on the stream and the diverse range of associated audience behaviours and responses, the understanding of which is (as we will show) important to fully theorising these streams. These selections emerged out of two ongoing longer term projects studying game live streaming via (in part) observational methods. In the case of the first author examples of nonhuman, game streams became apparent over several hundred hours of online observational work on Twitch between 2015 and 2021, in a project primarily examining questions of labour and culture in game streaming. During this time the first author has conducted hundreds of hours of observational work across over a hundred different Twitch channels and, quite deliberately, across a diverse range of broadcaster demographics and channel sizes. This has included observation of channels hosting the latter two case studies in this paper. Field notes that were (and continue to be) taken across this project focus on streamer behaviours, the aesthetics and visuals of channels, viewer interactions with each other and with streamers, and the emergence of phenomena such as in-jokes and memes in different Twitch channels. In the case of the second author nonhuman game streams were identified during an ongoing study of performance on Twitch consisting of several hundred hours spent streaming and spectating others’ streams. This author conducted participant observation within a range of Twitch channels, in particular noting user behaviours and how stream environments are constructed to facilitate forms of spectator interaction, as well as how spectators take up these contextual offers to interact with and within the stream environment. These observations underpin the first two case studies described in this section, which are in turn connected with similar approaches to archived footage of and secondary references to the third and fourth case studies to create a holistic impression of the role of nonhuman agency on Twitch. Field notes were collected alongside screenshots taken from relevant channels, coded thematically.

The specific case studies were chosen to demonstrate some of the diversity in these channels, to illustrate some of the many other entanglements of these channels (especially in terms of audience behaviours), but also to point towards commonalities and broader themes. These channels and their viewers were observed by the authors in situ, and these observations are coupled here with commentaries from gaming news sites, commentary sites and so forth. As such, we begin with two case studies of Pokémon trading streams. These constantly running channels use two different kinds of trading across different games to provide players with Pokémon that would otherwise be exceptionally time-consuming to obtain. Our third case study is the Super Mario Maker (2015) level ‘Lucky Draw’, which achieved significant visibility in 2019 as a user-designed stage that required a one in seven million chance to be completed, but no active user input. Our last case study is the channel Fish Plays Pokémon, in which a live fish observed by a motion tracker ‘played’ a game of Pokémon Red with what might, generously, be called mixed success. In each case, we describe the stream as observed, its nonhuman streamer, how viewers respond to or interact with the stream (if at all), and any wider cultural or technological elements that these nonhuman game streams touch upon. These streams each present a distinct and interesting form of nonhuman agency, and cumulatively communicate nonhuman live streaming’s potential as both cultural practice and an object of study, with importance both within the study of game live streaming, and beyond. We now describe our case studies in detail, looking particularly to highlight the different kinds of nonhumans at play, and how the humans in the Twitch chat window responded to these diverse Twitch streams.

Nonhuman Streaming

To start thinking about the roles of the above nonhuman agencies in gaming, and subsequently nonhuman game live streaming, it is useful to begin with thinking of play as a technological-human assemblage. Taylor and Elam (2018:250) use this lens to propose a model of gaming expertise as ‘autoplay’, privileging speed and accuracy over narrative elements of a game. Taking this one step further, it is apparent that player and game agencies more generally operate in a relational capacity. As we noted in the literature review, there is an existing history in game studies, traceable through Giddings (2005) via Fizek (2018:207), of ‘recognis[ing] technological agency and shy[ing] away from the anthropocentric assumption that agency resides solely in the human’. Broadening this argument, Straeubig’s (2020) typology of AI roles in video games presents technological agency as it ranges from operating in response to player input, to providing the parameters within which the player may operate. Not only do we engage with these ideas but we also propose to extend these existing technological notions of nonhuman gaming agency to incorporate biological gaming nonhumans, and to consider how streaming – with its spectacle, affective and other dimensions – complicates the picture.

Our first intervention is to decouple agency from affective intentionality. Here, we understand intentionality through Daniel Dennett’s notion of an ‘intentional stance’ (Dennet, 1987:18), in particular as it emphasises the relationship between behaviour intended to produce desired outcomes and, in terms of agency, an intellectual and emotional capacity to understand the results of enacted agency. While the relationship between intentionality and human agency necessarily has an affective component, this does not necessarily translate well to nonhumans. This is not to say that nonhumans don’t ‘see the results of [their] decisions and choices’ (Murray, 2016:159); they do. However, nonhuman – in particular, technological – perception differs vastly from human perception (Ouellette and Conway 2020) and so this actor-consequence connection varies. Considering ‘Lucky Draw’, the game ‘sees’ the result of each algorithmically chosen transformed block and responds accordingly, ultimately either triggering the P switch or not, although there is no deliberate link between these individual choices and the successful completion of the level. The human spectator meanwhile understands these results through the lens of either success or failure.

This is an example of ‘the game engine’s scripts running in the background’ as one element of ‘a decentralised assemblage actor’ (Fizek 2018:208). In FPP, Grayson is unaware that a game is even being played and yet human spectators are still invested in the outcome. The fish exercises agency in movement, but not intentionality in those movements being translated into gameplay inputs. During Pokémon trading streams, the value of the Pokémon being traded exists only to the human traders, while the nonhuman facilitates these trades without consideration to what is being traded (beyond rules around whether the Pokémon being created are ‘possible’ within the rules of the games). In these moments, ‘the human becomes a witness to the system’s alleged agency, and a delegator of play onto the algorithms’ (Ibid: 206), and hence generates a (false) sense of intentionality in what is taking place. Players choose to take stock of Grayson’s progress through Kanto, or to echo Mario’s seemingly endless descents into the offscreen abyss with their own descent into madness, or to watch others’ trades as they wait on their own. Regardless of what is taking place, humans – with their associations of gameplay objectives, goals, design, challenges, obstacles and so forth – cannot help but assign, even if only tongue-in-cheek, some degree of intentionality to the actions being streamed.

In turn, we propose that the dissonance between human investment in outcomes and the absence of the same in nonhumans further motivates the audiences of these streams. Nonhuman actors enact their agency in a way that is impartial to the result, while human spectators choose to wait and see what occurs. We propose that human perception and appreciation of nonhuman play is thus attributed heavily to the apparent unpredictability of what takes place and hence based on a false dichotomy between human-controlled and random outcomes. In truth nothing within these streams is unpredictable, but humans may perceive such because the processes underpinning the nonhuman decision-making are invisible and do not adhere to human logic, or they may lack the required information to understand the processes at play (cf. Johnson, 2018). One of the challenges facing game design is the gulf that persists between programmed games and human perception (Petrovic 2018; Ouellete and Conway 2020). Like a livestream spectator who watches a streamer play their favourite game to see what they do differently, here spectators gather to witness – or participate in – a different approach to play in the hopes of evaluating its legitimacy through eventual success or failure. Not only do the human viewers give a sense of intentionality to these nonhuman processes, none of which ‘know’ they are playing a game, but they also become invested in and concerned by these outcomes in a way these nonhuman biological or algorithmic broadcasters cannot.

Yet, the reality of this perceived unpredictability varies between technological and biological nonhumans. We previously noted the incidental nature of Grayson’s play: as a living being his movement is motivated by his environment, comfort, and survival. This is not at all ‘random’ even though the associated movements in the game seem unpredictable as they are oriented towards no particular goal. In the case of technological nonhumans, meanwhile, Giddings (2005: 126) is correct when he states that ‘whatever agency these simulacra exert, it is unguided by any moral or epistemological purpose’, yet even this is not to imply that it is entirely unguided. When no outside data is involved any ‘play’ is controlled algorithmically. Algorithms produce at best pseudorandom numbers, which are in fact not at all random as they produce outputs based on things like the system time. So while a human could not predict whether the next run of Lucky Draw is going to be successful or not, or whether the next perfectly timed Wonder Trade will connect with them or another player, these decisions are made algorithmically. Again, without exposure to this information, a spectator may perceive chance instead. Viewers therefore can attribute intention and investment where none exist, yet while also attributing outcomes to chance when that is not, necessarily, the case. This confusion of intentionality and agency and chance is a distinctive element of these streams.

The Agency Gap

All of this brings us to our second point – the concept we call the ‘agency gap’ which can be observed in these channels. We use this term to refer to the observation that in these channels lacking a human agent – an agent who can talk, engage, interact, joke, as well as playing the game itself – the audience increasingly seems to ‘take up the slack’ in generating ideas of intentionality within that channel. In the previous section, we characterised nonhuman agency in part through an investment in outcomes present in humans and absent in nonhumans. Fizek (2018:208) similarly suggests that ‘human players are driven, immersed, or frustrated. AI and machine just are’, and in the same vein Ouellette and Conway (Ouellette and Conway, 2020:11) point towards the importance of affective intentionality in consideration of agency, and that ‘how we feel about phenomena impacts how we perceive phenomena as significant, inconsequential, interesting, etc’. In both cases, these authors ask whether these nonhuman agents can ‘feel any particular way’ about their play, in this case in the way that (presumably) human streamers and spectators do.

In most games ‘the role of the human player is to actively participate in gameplay, and that of the machine to enable, sustain, and facilitate play’ (Fizek, 2018:206). This is very much the case in live streaming, where the player is not just the one performing these roles, but is also ‘actively participat[ing]’ in the wider play of the channel. Much of the appeal of watching Twitch broadcasts stems from the back-and-forth between the streamer and the viewers (Sjöblom and Hamari, 2017; Hilvert-Bruce et al, 2018) as well as streamers’ abilities to commentate and critique the game as being played (Wulf et al, 2020), or even cracking jokes and generating norms of community humour. None of these nonhuman actors can do this – their actions can provide amusement or interest, especially when incompetent or counter-productive, but any direct and intentional engagement with the audience is hard-coded (as in Pokémon trading streams), and certainly not any kind of verbal or affective back-and-forth. Nevertheless, the nonhuman streamers discussed in this paper do operate in relation to other nonhuman elements present within most streams. This is not necessarily via a controller or keyboard and mouse, but certainly as a computer that reads inputs of some sort (whether from the Twitch chat, a camera or a script within the game itself) and performs the associated in-game actions. Yet, when there is no human streamer present to perform these usual affective roles: what happens?

What we see in these streams – addressing this agency gap – is viewers increasingly taking up some of the intentional and affective work of human streamers. The absence of the central human streamer democratises stream participation, as no human actor’s agency is prioritised, or at least any prioritisation is moderated by an impartial nonhuman actor. While participation and interaction can be understood as democratising in their own right, the presence of a human streamer directing the flow of conversation and having gameplay choices potentially influence by selected viewers somewhat undermines games streams as being democratic spaces in terms of open, collective interaction. The nonhuman streamer’s indifference to any affective relationships between their actions and in-game consequences, by contrast, removes that investment and so they either perform entirely at the directives of members of chat regardless of the outcome (as in Pokémon trading streams or other examples not considered in detail here such as Twitch Plays Pokémon), or without any consideration to those directives (FPP or ‘Lucky Draw’). Viewers are no longer beholden to the structure imposed by a human streamer and nor as some singled out for attention and responses. We also note these streams tend not to possess the elements of social hierarchy most channels exhibit as a result of interactions between the streamer and their favourite viewers, again setting all viewers on a (relatively) equal standing.

It is also viewers who begin talking about the channel or sending it to games journalists who might be interested, rather than the streamer attempting to advertise themselves or promote their channel directly, as many do. Human live streamers often exhibit an aspirational orientation but with no humans to pursue visibility, these channels mostly spread by word-of-mouth instead. It is the viewers who discuss among themselves, now with no reference to a human streamer, defining and framing events as they choose rather than using the streamer’s commentary or jokes to drive the interaction in the chat window (cf. Wulf et al, 2020). This curatorial work does play a role in the eventual direction or ‘success’ of a channel, but falls entirely to viewers in these channels. The nonhuman streamer does not guide the emergence of the codification of a channel’s culture, and is indeed selection of actions and interactions into which viewers write their own interpretations. The traditional hierarchy disappears as there is no human streamer to establish a channel’s authority, rules, norms, and so forth—with viewers now instead doing all of these things themselves. Without direction from a clear ‘leader’ in a channel – normally a human streamer – direction and meaning are more widely spread.

This also extends to the different responses viewers seem to have to the nonhuman agents in these channels: some embrace the strangeness of these broadcasts, while others attempt to use language and existing norms to contain and define a game playing itself in an infinite loop, or the bizarre spectacle of ‘one pet fish’s struggle against chaos’ (Statt, 2014). As Švelch (2020) observes, players ‘welcome a degree of unpredictability, variation and mystery afforded by AI routines – or by random number generators or pet animals – yet, many also wish to ‘completely account for and understand the behaviour’ of nonhuman actants. In these streams (as in TPP), we see viewers eagerly hanging on every movement of the fish, typing a storm of messages in chat trying to lure it one way or another (akin to cheering in a stadium); alternatively, we see people using it as incidental background entertainment, or checking back from time to time to see whether any progress has been achieved through apparent chance. Regardless of how spectators engage, they do engage, and this active participation sustains these streams, filling the gap that emerges in the absence of a human streamer and producing distinct dynamics for nonhuman streams. The presence of the agency gap thus extends Egliston’s discussion of ‘the way that relations between non-human objects create a channel for various organic and inorganic affects’ (Egliston, 2020:250) in terms of intentionality, whereby the lack of affective intentionality exhibited by the nonhuman streamer is compensated for by human spectators who project their own values and investment in success (as they understand it) onto the stream.

The Automation of Play

Lastly, these case studies suggest interesting observations about the ongoing automation of play (cf. Taylor and Elam, 2018), and point towards a number of hypotheses for the future of this area. With the rise of pay-to-win games (Lelonek-Kuleta et al, 2021) and the ability to use real-world money to pay for advancement (or simply guaranteed success) in many blockbuster or mobile phone gaming titles, a similar trend can be observed on Twitch. Confronted with an endlessly growing number of games to play, many viewers turn to platforms like Twitch (and YouTube) to get a sense of a game without having to commit the time and effort to actually playing it themselves. Such channels and videos can be watched in the background while doing something else, or (if broadcast by a particularly skilled player) might actually enable a viewer to experience more of a game’s content than they would on their own. This automation of viewing or experiencing elements of play without the embodied human player is also comparable to the concept of the ‘Tool-Assisted Speedrun’ (or TAS) in which a digital program is created to play a game as perfectly and as rapidly as possible, often producing outcomes that are simply impossible for a biological actor such as a human to produce themselves. Such speedruns are noted for essentially reducing games to their purely mechanical or algorithmic components (i.e. without any affect, story, theme, etc) and using those systems to achieve extreme gameplay outcomes. Given these developments, we can reasonably regard Twitch streams as having become one element of this trend – this is the case for streams with human broadcasters but is especially true of the channels described in this paper, where not only are viewers not intentionally engaging in play but nor is the streaming agent.

Similarly, when Anne Allison conceptualises so-called ‘Pokémon capitalism’, she notes that ‘the currency of play here, pocket monsters, is at once traded and accumulated, building capital for the player but also relationships with others’ (2006:339). In her examination of children's engagement with the virtual creatures she found that the exchange of the Pokémon converted them to a commodity status, but also connected players through the act of trading. Fifteen years later, when these trades are now performed across the globe and facilitated by the Internet, this human-to-human sociality is somewhat compromised. Moreover, as the streams we discuss here demonstrate, it is possible that the trader on the other end is not even human. On the other hand, however, is the democratisation of this ‘Pokeconomy’. Both Wonder Trading and Link Trading enable players’ access to Pokémon that they may never have otherwise caught, and whether this comes down to a lack of time to play the game, poor luck or missed time-sensitive events, the existence of these streams opens up new possibilities for players. The second author this paper has these streams to thank for his own ‘shiny’ pink Heracross (a normally blue creature based on the Hercules beetle), given that he is apparently too busy writing about such a Pokémon to actually catch one. The tensions around legitimate forms of play hence persist here, which in turn raise questions about the actual value of these Pokémon, a subject worth further study which has echoes in other popular games (cf. Jarrett 2021). Nevertheless, these nonhuman streams enable unprecedented levels of access and opportunities for players to live up to the ultimate challenge of the Pokémon games by aspiring to, indeed, catch ‘em all.

More broadly, such an ‘automation of leisure’ (cf. Baranowski and Mroczkowska, 2021) represents an important domain and framing for future study in the live streaming context. Just as ‘repetitive and mechanical actions’ in factories or on production lines can be automated through technological systems, so too can this take place for digital actions (de Paoli, 2013b:13). This is of course the case for systems like automatic spam filters and the like, but has rarely been applied to leisure, since the point of leisure is, ordinarily, to do the activity, and hence derive enjoyment or satisfaction or fulfilment. Yet, just as the ‘tasks required by Amazon’s low-price anonymous [Mechanical Turk] labourers are infinite’ (Zylinska, 2020:118) one cannot help but wonder – what other gaming tasks might be assigned to computers instead of the humans who are, in principle at least, supposed to be getting something out of their play? Since Marx, it has been noted that industrial machinery devised by a capitalist class has, at least in part, its objective in automating repetitive tasks (de Paoli, 2013a). Although Fish Plays Pokémon is perhaps different and the repetition of the fish’s near-hopeless moves is in a sense the very point of the stream, it is clear that the creation and trade of high-quality Pokémon creatures and the endless attempts required to complete ‘Lucky Draw’ are both highly repetitive in nature and have a clear instrumental goal in mind (from the perspective of human viewers), that is, removing a task from a human and delegating it instead to a nonhuman as a source of cheap labour. De Paoli (2013a) points to the ‘deskilling of players’ as being a ‘key aspect of the automation of gameplay’, and this transference of the skills humans would normally possess – to play a level, navigate a game world, catch and train high-quality Pokémon for trade, and so forth – to automatic software may well signal broader changes to come in the consumption of leisure. In the coming years – as more and more games have these sorts of systems, game live streaming presumably continues to grow, and integration between streaming and the games themselves develops further – we anticipate a growth in these sorts of automated responses.