The practice of building intelligent narrative technologies is implicitly grounded in a sociopolitical context whose values are reflected in the artifacts we build and our narration of their use. This position paper attempts to make these implicit values explicit by naming them and conceptualizing alternatives. We argue for adopting a Critical Technical Practice of narrative intelligence in which explicit critical reflection on the values and metaphors a system's assumptions embed is interwoven into the design process. We demonstrate a broad application of this practice by identifying four ideals underlying narrative intelligence systems that impose and reflect systems of power: the Hero's Journey, Amalgamated Human Authorship, Social Believability, and the Holodeck. We then identify values that are marginalized by these ideals, such as narrative pluralism, amplifying underrepresented voices, representation of non-normative identities and relationships, and opportunities for reflection and interpretation. Finally, we propose imagined alternatives to existing ideals that bring these marginalized values to the center.
When building tools and technologies to support the human practice of storytelling, our training leads us to focus on technical innovation: what data structure will we use to hold knowledge about the world? Will that knowledge be handauthored or extracted from a dataset? What algorithms for model construction, search, and selection will determine the components of generated stories?
However, maintaining this focus comes at the expense of
asking other questions. For instance, what are stories for,
and what is the social value of creating technology that can
replicate or model the human processes around
understanding and creating them? Both the technology and mass-media
storytelling industries exhibit dramatic inequalities across
gender, race, and other axes of privilege in terms of whose
voices are centered and whose stories are told. Without
taking seriously the questions of how the technology we create
can shift or amplify existing power imbalances, the
narrative intelligence community is in danger of perpetuating and
deepening these inequalities. We may operate under the
belief that the systems we construct are sociopolitically
neutral, but in doing so, we willfully ignore our own biases. As
Elizebeth Sampat puts it in her book Empathy Engines,
“Instead of striving for neutrality, the first step to truly affecting
game design is to realize that nothing is neutral: what we see
as a lack of bias is our own blindness to the circumstances
we live within”
The good news is: we don’t have to stop writing code and building systems in order to confront these issues. In fact, the practices of technology-building and critical reflection can mutually support one another. This position paper argues for a critical technical practice of Narrative Intelligence in which technical innovation is tightly coupled with an examination of implicit values throughout the design process.
Critical Technical Practice (CTP) is a proposal by Phil
Agre, originally in the context of Artificial Intelligence more
broadly, to pair the creation of software artifacts with critical
reflection on the underlying (often implicit) metaphors and
values used by its creators
Motivated by the continued relevance of Agre’s 1997 observations, we present a position paper arguing that Narrative Intelligence as a field needs a Critical Technical Practice. We make an effort to “defamiliarize” familiar concepts in the field of Narrative Intelligence and imagine alternatives to them, listing a broad (but shallow) set of examples specific to this field.
According to
2. Noticing what, when working within those metaphors, remains marginalized; 3. Inverting the dominant metaphors to bring the margin to the center;
We demonstrate the first three steps of the above outlined
process by taking a broad view of Narrative Intelligence as
a field, considering problems as diverse as narrative
cognition, interactive story systems, and expressive simulation.
However, our aim is not to take the fourth step by
presenting an embodiment of our imagined alternatives as a new
technology. Instead, we outline several potential directions
for narrative intelligence research to more closely examine
marginalized values, aiming to create space for critical
reflection and ideation without the pressure to develop
validated solutions to each problem as it arises. This approach
is consistent with Agre, who “believe[s] that building things
is an important way of learning about the world,” but also
points out that “even if the value of critical reflection is
proven only in its contribution to improved technical
systems, many valuable criticisms will go unpublished if all
research papers are required to present new working systems
as their final result”
Agre’s initial writing defining CTP was in the context
of artificial intelligence, and he lamented being unable to
get across to his colleagues the importance of examining
metaphors and values
We are hopeful that AI for entertainment’s similarly
interdisciplinary makeup bodes well for CTP. There is
certainly precedent for openly confronting methodological
differences between the disciplines that comprise us, such as
the dialogue between “Science Considered Harmful”
Finally, efforts to address implicit assumptions in AI as a
field at large are resurfacing. In “Decolonial AI”,
In this section, we identify several common metaphors and values prevalent in the past and present of Narrative Intelligence as a community and body of work. For each metaphor or value, we identify ideas and values that are marginalized as a result.
Some of the earliest commonly-cited work in narrative
intelligence is the Russian formalists’ attempts to identify
common structural patterns in Russian folktales, such as Propp’s
Morphology of the Folktale
For those of us with computer science training, the concept of interchangeable parts within a constant structural framework appeals naturally to our ideal of abstraction. By generalizing from examples, we can extrapolate patterns in the same sense as a sewing pattern, something that can be transmitted through language, then traced and replicated to materialize in different contexts, each instance still recognizable as having a common origin.
In addition to the general use of patterns appealing to
particular computational metaphors, there is also much to
be gleaned from the choice of patterns employed (and the
source material that gives rise to them). For example, a
common “plot template” is the Hero’s Journey, as popularized
by
This pattern embeds a number of culturally relative assumptions, particularly in Campbell’s more detailed breakdown of each segment of the journey, including gender as binary, social functionalism, and male as primary/default. The eponymous Hero is always male (male as default), and women appear only in elements of his journey in specific roles, such as the Goddess who aids his quest, and the Temptress who tries to derail him. The central male character takes on all the active roles in the story, solving problems with decisiveness, aggression, and violence; women play much more passive roles, such as fleeting assistants or irrelevant distractions. This prescription of role to gender is a form of social functionalism, a philosophy that argues that the moral function of society relies on people following the roles that match their innate characteristics. This use of gender roles both embeds the cultural values of the authors who created these stories, and reinforces those values as cultural standard by embedding them in trans-generational myths.
The Hero’s Journey does not exist in a vacuum: it
comprises the mythologies that precede it and informs the
stories that follow it. The same cultural assumptions can be
discovered in modern (literary and film) media, as revealed
by David Robertson’s analysis of Mark Riedl’s Wikipedia
Plot Summaries dataset
A second set of values reflected in the Hero’s Journey and
its fellow narrative templates pertains to their embedding in
Western culture, taking Western assumptions about narrative
for granted as universal. For instance, the assumption that
all narratives must have conflict, which drives, for instance,
the CPOCL project
Along similar lines,
In summary, narrative templates, especially those based on culturally dominant pretexts, marginalize the following values: • Narrative Pluralism: the co-existence of differing human cultures and the non-universality of any given culture’s approach to storytelling • Social Role Autonomy: the idea that one’s function within a story, as in life, is not pre-determined by attributes like gender and race that one is born with.
A number of computational narrative researchers
explicitly cite Roger Schank’s concept of social scripts
Once the challenges of extensive domain authoring,
maintenance, and debugging became apparent for creating
narrative templates as above, some people got the idea to try
to assemble this kind of “commonsense knowledge” about
human behavior through crowdsourcing and statistical data
analysis. For instance,
This last example, the Scheherazade project, is especially
interesting in the central metaphor the authors use in the
title of the first paper: “Learning to Tell Stories about Social
Situations” Li et al. (2012a). They state (emphasis added):
“By leveraging the crowd and its collective understanding
of social constructs, we can learn a potentially unlimited
range of scripts regarding how humans generally believe
real-world situations unfold. We seek to apply this
scriptlike knowledge to the generation of believable stories that
involve common social situations or the direct engagement
of virtual characters in social behaviors.” The form of social
intelligence they aim to replicate computationally is exactly
that found in the statistical average of “the crowd,” along
with whatever ingrained biases and cultural assumptions the
humans making up this crowd may carry—which is evident
in the plot graph that emerges for a prompt in which “John
and Sally” go on a date (see Figure 3). Although the authors
suggest that one could use such a model to subvert
common expectations, generating a more narratively-interesting
form of surprise, in practice, their uses of the crowd-learned
model for story generation and interaction straightforwardly
replicate the model’s statistical distribution
In summary, the amalgamation of human authorship into statistical models marginalizes the following values: • Underrepresented Storytellers: amplifying the voices and creative decisions of people whose stories are seldom told. • Authorial Voice: the ability to attribute meaning to the identity and labor of human story authors.
One common approach to interactive storytelling is social
simulation: the use of AI to simulate story characters that
embody a systemic model of social interaction. Examples
include commercial games such as The Sims, Rimworld, and
Animal Crossing, in addition to academic work such as the
Comme il Faut engine
The concept of simulated societies necessarily embeds
sociological theories chosen by the programmer. When these
theories are based on intuition rather than research, we
might call them folk sociological theories: for example,
Rimworld
In summary, the ideal of social believability marginalizes the following values: • Non-Normative Representation: representing nonsocially-normative people and their experiences with chafing against norms (e.g. gender/sexual minorities, people of color, people with disabilities) • Procedural Inquiry: purposefully limiting social simulations to legible systems of expression as a means of interrogating what you can and cannot say within that system.
While not unique to interactive narrative, the obsession with
“immersion” that sees virtual reality as its logical
conclusion pervades discourse in our community as well. In
particular, the Star Trek fiction of the Holodeck has long been a
metaphor for the goal of intelligent interactive narrative
experiences
Critique of this ideal is not new or unique to our field,
either.
The immersive fallacy marginalizes the following three
values, the first two of which come from Mitu Khandaker’s
AIIDE 2015 keynote
What are stories for, and why are they so important to being human? They help us imagine alternatives to the way things were in the past or might be in the future; they help us make sense of our own lives and difficult situations we have faced; they help us see alternative versions of ourselves, and by placing ourselves in the shoes of a story character, we learn new ways of emotionally processing and problem solving.
Intelligent narrative technologies have the potential and opportunity to do all of these things in new and powerful ways: authoring tools that democratize technology, enabling more authors to tell stories about the world through their unique perspectives; supporting the human imagination through procedural generation of unexpected configurations; narrative generation systems that ask players to make choices and reflect on their consequences. To achieve this potential, we posit that we need to reexamine what human forms of narrative intelligence we really want to emulate with our technology. For example, a college literature professor might assert that narrative intelligence means the ability to interpret ambiguous stories, to situate the meaning of the story with respect to concepts of authorship (who wrote the story?), representation (who is shown in the story doing what kinds of things?), and historical context (what was going on in the world when and where the author wrote this story?). We might value a reader’s ability to draw parallels with present-day events, with other narrative works, and with their own personal experiences.
In this section, we return to the lists of marginalized values at the end of each subsection above and identify alternative ideals and research goals that would center them.
In support of the values undermined by using common story patterns as a basis for narrative generation, we propose to redefine what we consider the “canon” of narrative reference work, and to rework our definitions of narrative on the basis of these findings. We need to surface the implicit canon (e.g. the Hero’s Journey, Plotto, and Aesop’s Fables are just a few), and we need to recognize their contingency by seeking counterexamples to their structural assumptions. What would a practice of narrative intelligence that centers the value of narrative pluralism look like?
One potential approach is to work actively to make narrative intelligence a more international field, forming more collaborations with researchers from Africa, Asia, South America, and members of indigenous nations across the world. Another is to encourage inquiry about our own tastes and influences as narrative researchers, consider the cultural origins of the stories we love, and surface the contingency of the patterns we take for granted. A third strategy is to change our approach to pedagogy when we are teaching students from narrative cultures that differ from ours: rather than stating the Western canon as a given that they must learn, we must give students the pedagogical “permission” to invent new typologies and definitions based on the stories they grew up with.
What would it look like to take a crowdsourced dataset, or
a model like GPT-3 that is compiled from a huge number
of different English text resources, and model their tropes
and biases for the sake of analyzing and subverting them?
Would it actually work to take Scheherazade’s plot graph
model and simply invert the optimization function,
choosing the least expected event over the most expected, and
argue that the result inverts common social expectations, or
would we need a more systematized model for tropes and
expectation? Supposing we could identify bias in datasets
or text input, could we build a mixed-initiative system with
an interface similar to Tom Forth’s “gender bias calculator”
for recommendation letters
More generally, we propose a model of narrative intelligence that takes into account the fact that our understanding of any given story depends on the thousands of stories we have encountered before it. Through our lives, we develop not just linguistic literacy, but also media literacy, expectations about the conventions and tropes present in a given genre or artistic movement. Authors depend on this literacy in order to subvert it: if they know their audience expects one thing, they can strategically pull the rug out from under them to say something new about that trope. What would it look like to design a computational narrative generation system that could reason about and implement similar strategies?
In interactive narrative experiences, neither the player nor the game designer has sole creative ownership over the story. However, most systems attempt to lean strongly one way or the other, either asking the player to (merely) bear witness to a well-defined story constructed by the designer, or providing a fully sandbox-mode simulation in which the player is free to tell any story they can with the simulation mechanics.
We envision a model for granting a player a sense of
autonomy without turning the dial all the way to
“omnipotence.” This kind of experience requires a push-and-pull
between player and system in terms of asserting authorial
voice; the story emerges from a negotiation over their
contested overlap. Models for this kind of collaborative
negotiation can be found in tabletop storygames, particularly those
without a central story-manager or game-master, such as
Fiasco
Even within the more conventional
player-as-storycharacter mode of interaction, however, we have
opportunities to enable more active participation in shaping the story.
One potentially fruitful avenue is to prompt the player
towards autobiography, as recently explored for the purpose
of mental health recovery for post-traumatic stress
disorder
The concept of legibility, as defined by
Mitu Khandaker describes wrestling with how to
“represent love” in the game Redshirt, and unlike the author of
Rimworld, opts to make some of its workings based on a
hidden variable with no connection to the explicit
character traits
“Simulation can’t simulate everything, of course,” she
argues. ”The things we leave out when we’re designing a
game are often more interesting and more expressive than
the things that we put in. Play operates on that gap. It
entices us to fill it in. It’s the same principle [of cartoon
abstraction as identified by] Scott McCloud in
Understanding Comics
In this position paper, we have argued for narrative intelligence as a field to engage in Critical Technical Practice, and against the requirement of proof in the form of a working system as a process for resisting the “fallacy of alternatives.” We outlined the steps for doing so and demonstrated the first three. We took a shallow but broad pass at identifying dominant metaphors and values in the current-day technical practice of narrative intelligence and discussed several alternative values and metaphors, which we believe have the power to expand our imagination of what this field can achieve.
While this paper attempts to overview the breadth of ways in which narrative intelligence can benefit from CTP, to really do justice to the critical thought required, an entire paper could be written about each identified value (and corresponding alternatives) described above. Likewise, any given technical project could also go through a much more indepth application of the steps of Critical Technical Practice. We therefore conclude by asking researchers in the INT community to learn the vocabulary of critique, build relationships with their humanist colleagues, and ultimately to engage differently with their technical work, creating a culture in which considering the values reflected by each technical project is not an afterthought, but a built-in and continual part of the design process.
The authors gratefully acknowledge the reviewers for engaging thoughtfully with this work and helping us identify interesting lines of related work. Intelligence and Interactive Digital Entertainment Conference.