AI Nemesis

AI Storytelling and Its Future in Games

As in the wider field of technology to which it is subject, revelations and disruptions occur frequently in gaming. The Xbox 360 ushered in the online age of gaming, the explosive popularity of the Nintendo Wii and smartphones introduced novel input methods, and consumer-ready virtual reality headsets are set to offer something fundamentally different for gamers from next year on.

Recent years have also seen budgets ballooning at the top end of the development scale, which has resulted in some studios, including Firaxis and From Software, using procedural content generation to increase replayability. While the technique is effective in randomising environments and challenges, that is not the full extent of its usefulness.

Enter the Entertainment Intelligence Lab at Georgia Institute of Technology, led by Associate Professor Mark Riedl. Earlier this year, a small team from the lab released the first report on Scheherazade-IF (Interactive Fiction), a program capable of generating video game stories with “near-human level authoring” based on crowdsourced examples. The report was based on the results of the system’s performance in two narrative situations (domains): a movie date and a bank robbery. Riedl explains that this limited sample was selected from about a dozen story models that have so far been introduced to Scheherazade-IF because they were “a bit more mature than the others because more resources were invested in them (i.e., crowdsourcing enough to get them relatively correct; the more data the more accurate).”

Furthermore, these two domains show off the versatility of the system. Although they don’t quite prove the grandiose rhetoric of the initial press release’s claim for the possibility of “a Star Wars game using online fan fiction to let the AI system generate countless paths for a player to take”, they do reveal Scheherazade-IF’s diversity. Riedl says that this stems from the diametric differences between the two situations:

A date at a movie theatre shows that the system is learning sociocultural knowledge—it is the sort of commonly shared cultural experience that humans have that computers typically don’t have and can’t easily get by other means. Bank robbery was selected because it is a bit more action-oriented. The bank robbery also shows that we can crowdsource topics that have commonly shared beliefs but are not directly experienced by crowd workers.

These shared beliefs offer a loose set of “rules” for any given situation, which Scheherazade-IF learns in order to generate new stories. The examples that have so far been made available to the system have been written bespoke by contributors on Amazon Mechanical Turk, but Riedl says that it is theoretically possible to use any writing database. He qualifies this, however, by saying that “in practice you need a lot of people writing stories about very similar things. The less that stories align with each other the greater number of stories the machine learning will need.” Even with this limitation, Scheherazade-IF remains a far more versatile AI solution to narrative generation than other extant models.


Mark Riedl, Ph.D, Associate Professor at Georgia Institute of Technology.


One of the most visible examples of procedural storytelling in recent times is Monolith Productions’s Nemesis technology, which featured prominently in the marketing for Middle-Earth: Shadow of Mordor and proved to be among the best aspects of the game. For those unfamiliar with the title, the Nemesis system allowed enemies to remember past encounters with the player character, gaining new resistances and abilities for future battles and creating a personal narrative throughline alongside the overarching quest. As interesting and engaging as the system proved to be, it features one fundamental difference to Scheherazade-IF in that it offers emergent stories, rather than directed ones.

Another developer that has toyed with the same idea is Bethesda Game Studios with the Radiant technology, which has featured in its games since 2006’s The Elder Scrolls IV: Oblivion. Originally used to create the simulacrum of a living world by imbuing NPCs with routines, later iterations saw it expanded to be able to generate a purportedly infinite number of side-quests based on the player’s abilities and previous play experiences. While the system also created small, self-contained stories to accompany these quests, they were often of a rote, repetitive nature. Ultimately, like Nemesis, the narratives offered by the Radiant technology found their strength through the emergent structure of Bethesda’s games.

Although each of these solutions has proven effective at bolstering player engagement, Riedl doesn’t see AI becoming the standard for the creation of narrative content any time soon, saying that “worlds would have to be very large to reach the point where it was intractable to use human content creators.” In the short term, he argues, “game development studios can often afford to pay a lot of content developers and programmers. However, game development studios also strive to build tools for content creation into the hands of non-technical designers. In the future, one might envision non-technical designers ‘programming’ small social vignettes or quests into open-world role-playing games by telling stories.”


As for Scheherazade-IF, despite being built upon a core system that began development back in 2011, the program remains in its infancy. Creation of the core program (simply called Scheherazade) was predicated on the team’s desire to develop a system capable of generating stories in more than a single domain, as had been the case prior. The addition of the interactive branch came later, after the team realised that they could “let the user play a character in the story because story generation is fundamentally about making choices for characters. The user would make choices for one character, and the AI story generator could make choices for all the others.”

For the time being, Scheherazade-IF’s application is restricted to a textual interface whereby users must select their action from a limited number of options (as demonstrated in the video above). However, the next step, says Riedl, is to make it into “an improvisational story co-creator”, wherein the player is able to interact with the story through natural language inputs. At the same time, the Lab has also developed GameForge, which “creates 2D Zelda-like worlds from an input story” (see below), which could, in principle, be connected to Scheherazade to create simple adventure games with emergent narratives. Another technology that Riedl’s team has experimented with is automatic generation of 3D environments built from text-based descriptions. It was partially successful, with Riedl attributing its key difficulties to “story text not containing all the spatial details necessary to render scenes in which story plot points take place.”

As might be expected, these various technologies and experiments serve a convergent goal. The end result of Riedl’s efforts is envisioned to be “a system that not only crowdsources the interactive story, but also constructs the entire virtual world necessary to play the game graphically.” An ambitious goal, to be sure, but Riedl isn’t interested in bringing an end to video game development as we know it. His purpose is broader, and more academic and humanitarian than that.


A recent survey of more than five thousand smartphone users worldwide found that the majority believed that AI interfaces will replace current devices within five years. Siri and Cortana represent the steps already taken by Apple and Microsoft, respectively, to implement AI-based personal assistants into their operating systems, and there is an ever greater push to develop true artificial intelligence, such as the recent announcement of OpenAI. In its current form, however, the decisions made by AI can often seem inscrutable. In this respect, Riedl refers to it as “an alien sort of intelligence”, attributing AI’s shortcomings to “the fact that they cannot make sense of what we are trying to accomplish or why.” Riedl’s research into computational narrative intelligence is one potential way of overcoming that barrier by opening the possibility for AI to “create rapport with humans by sharing virtual vignettes.”

In keeping with this, Scheherazade-IF was not designed with entertainment-oriented games in mind and may not be suitable for that role.  Riedl says that this is the result of the median-based model of this system, which means that its stories would “largely avoid the dramatic twists that one would want in a strongly-story driven game.” The team has, however, discussed ways to make the system’s vignettes more dramatic, although practical application of such theories has not yet been attempted.

Rather, the system is tailored towards more serious purposes, with Riedl citing corporate training as one potential avenue of the system, where the scenarios could be drawn from real-world situations. Another possibility for a future iteration of the system is to use it as a tool to teach social skills to children and young adults on the autism spectrum by simulating particular scenarios in a graphical format. There is little to suggest that such situations may not come to pass. Studies have shown video games capable of increasing spatial awareness, memory recall, and other neurological factors, as well as aiding in physical rehabilitation, so, given the right stimulus, they could almost certainly help the social development of young people.

More than this, however, Scheherazade has been built with the goal of contributing to the integration of artificial intelligence with the wider world. The real reason behind its creation, says Riedl, “is to probe the bounds of questions of how to get a computer to reason about, create, and understand narrative.” Given that humans are raised on stories, and make them a part of our everyday lives, Riedl’s goal is vital for any artificial intelligence if we want it to understand us.

Click to comment
To Top