The ambition to create “real”, believable characters has been a cornerstone of literature since the 19th century. The Russian philosopher Mikhail Bakhtin praised Fyodor Dostoyevsky for his ability to give each of his protagonists their own sets of beliefs, “as if the character were not an object of authorial discourse, but rather a fully valid, autonomous carrier of his own individual word”.
Fast forward a hundred years and writers such as Joseph Conrad and Virginia Woolf were even more enthralled with the idea of character autonomy.
During the 1950s, the British novelist Henry Green grew increasingly frustrated with the tyranny he had over his characters, eventually writing novels composed almost entirely of dialogue. “Because if you want to create life,” said Green in a lecture for the BBC, “the one way not to go about it is by explanation.”
The digital landscape of the 21st century has created opportunities for Green’s ideas to be explored from a different perspective.
From virtual assistants such as Siri to conversational artificial intelligence (AI) such as Cleverbot, to games that learn from thousands of online players, the idea of characters who speak, think and feel has slid into reality.
Clever conversations
Launched in 1997, Cleverbot is a web application that imitates conversation; it is able to “learn” what to say through algorithms that call on a large database of stored responses from millions of previous exchanges.
“Things you say to it today may be used in its conversations with others in future. It can then learn how to respond to those things from the next person,” explains the software’s creator, Rollo Carpenter.
He explains that there is currently data for 195m conversational lines for Cleverbot to call on. “To reply using all that data is not a small computational task, especially if considering every possible way to reply contextually,” he says.
“The context it knows is the past of your current conversation versus the past and ‘future’ of all of the conversations that have come before. That’s where any perceived ‘intelligence’ it might display comes from.”
While the “perceived” intelligence of Cleverbot has spawned films and stage-plays, perhaps better for the spontaneous quality of their script than the psychological depth of their characters, recent developments are bringing newer breeds of AI into the fray.
The recently released racing game Forza Horizon 2 is one example of how videogames are signalling the ability of AI to learn in increasingly sophisticated ways. Developed by researchers at Microsoft, the game uses Drivatar technology – a system first introduced in last year’s Forza 5 – to simulate genuine human behaviours.
Other games let players race against exact recordings of previous runs, but Drivatar allows the game to repeatedly observe the driving style of each player; how they handle corners, how they overtake, and then it creates a “ghost” version for online friends to race against. Even more than this, once the AI has been “trained” it can take over and complete certain game events in place of the player, approaching races in the same way the player would.
Speaking to the principle researcher of the technology, Thore Graepel, it becomes clear that the possibilities of the technology extend beyond racing. “Car racing was a good application because it is more of a low-level AI problem, which does not require symbolic reasoning or language capabilities,” he says.
“In the future, I anticipate that we will see machine learning methods applied to more complex game scenarios such as in first-person shooters or role-playing and adventure games.”
Avatars with memories
Replacing racing tracks with conversations and stories might not be an arbitrary step, but the proliferation of online gaming and cloud computing are offering immense new sources of data for machines to learn behaviours.
Graepel explains that Drivatars learn from supervision, in that the AI observes the behaviour of the player and attempts to imitate it. Another type of learning is through reinforcement, where a computer can explore a game and either be rewarded or punished, through algorithms, for its actions.
Both this “carrot and stick” method and the imitation method are important for the future of AI, says Graepel, and the scale of learning offered by online gaming has the potential to teach AI characters increasingly sophisticated behaviours.
“One can imagine equipping in-game characters with an entire cognitive architecture including short- and long-term memory, drives and desires, attention, cognition, and even emotions,” says Graepel.
“They would then roam the game world, interact with other in-game characters, both of their own kind and human controlled, and slowly learn how to behave in more complex ways in accordance with their drives and desires.”
For authors who’ve fantasised about their characters leading a life of their own, could the situation of AI avatars roaming a fictional world, interacting based on their own individual drives, offer greater freedom of expression? Or is it likely to result in hundreds of dead-eyed puppets walking repeatedly into the sides of houses? Don’t stories require direction?
“I don’t see AI replacing human authors in the creation of existing forms of literature — novels, short stories, poetry and the like,” says artificial intelligence expert Malcolm Ryan from Macquarie University. “Rather I envision new literary forms that will be enabled by narrative AI.”
One of Ryan’s students, Margaret Sarlej, recently caught public attention with her project, The Moral Storytelling System, which generates moral tales structured around characters who are able to experience up to 22 emotions. The project is a good example of the complexities required for an AI to author even simple formulaic stories, let alone psychologically intricate novels.
Other developers who have experimented with creating AI-structured fiction include Emily Short and Richard Evans, who built the interactive fiction engine Versu.
Digital stories built using Versu, such as Blood & Laurels, put the AI characters in a position where they have to “improvise” around the choices made by the reader. In Blood and Laurels; a pulpy story of Roman intrigue, possible actions involve everything from sipping wine to trying to kiss other characters.
“The point is not that a single playthrough of a Versu game would match the experience of reading a human-authored story,” notes Ryan. “Rather, it is that fact that Versu can be played multiple times and generate a variety of stories out of the same raw materials, which makes it a fundamentally different kind of work.”
The computer that wrote games
If, as Ryan suggests, these ventures introduce a different form of storytelling, one project which pushes the idea of authorship further still is Michael Cook’s Angelina. A researcher at Goldsmiths College, Cook has over the last few years been developing an AI which can independently design its own videogames. A fuller look at how Angelina works can be found here but to give a basic idea, a word is given to Angelina, such as “fishing” or “alone”.
The AI program will then search databases and twitter to find associations for the word. These associations in turn go towards finding multimedia assets which will make the scenery of the game. More associations are found to obtain feelings people express in relation to the words, which in turn leads Angelina to choose the soundtrack.
“We’re getting better and better at building hand-made custom AI systems for characters, as we saw with Versu,” says Cook. “But what I’m seeing more of lately is us getting better at building systems which can go and find out about the world to fill in gaps in their knowledge.”
Angelina has gone through multiple iterations, moving from simple 2D games to 3D environments. Its autonomy is still in its infancy, but it shows a potential direction for technology to take and raises again the ghost which haunted a great deal of literary discourse in the twentieth century, that of authorship. It may not represent the death of the author or the rise of the machines, but the way Angelina crowd-sources assets, comparable to the way Chatbot or Drivatar learn from the ‘cloud’, raises the idea of whether Cook, Angelina, or us are responsible for what it creates.
For all this technological progress, in the end we may be no closer to truly “real” characters. Examining and indexing the methods AI use to learn and interact is fascinating, but something that 20th century authors like Green understood is that autonomy does not necessarily come from understanding characters; that life is perhaps best expressed through inaccessibility.
Throughout his novels, Green’s characters constantly misunderstand each other. We might hear their voices but there’s the impression we don’t know the people speaking.
In her essay on the subject, Dr Julia Jordan at UCL points to this impenetrability as a way for writers to make their characters resistant to the author’s control: “The ultimate authorial not-knowing is a refusal to presume one knows one’s own characters, so that they can, in flight, become entirely invisible”.
For Cook, opacity plays a different role: “It is a very important topic with Angelina and has been for the past few years,” notes Cook. “The more people learn about Angelina, the less creative it appears to them. In the real world, we can’t know what someone else is thinking, which might lead us to think they’re being clever when they’re not.”
Whether inaccessibility to a character’s thoughts is used to grant them an inner world or to mask an empty hole, the inability to fully, truly understand a video game character may one day be the thing that makes them the most real. They may become like the people we meet every day.
Comments (…)
Sign in or create your Guardian account to join the discussion