Neverwinter Online Twitter
“I really liked writing the bard, ” says Stanton, who handled the bot’s narrative voice while Kazemi programmed the custom graphics needed to produce the personalized headstones. “I ended up thinking of her as sort of like Mallory Archer from Archer, but as a [British] graveyard tour guide. That was sort of the voice I ended up kind of landing on. You know, very droll.”
Feel Train itself defies a catch-all label. Essentially, Stanton and Kazemi’s work hopes to bring a shred of humanity to how people interact with things on the internet, and something like a character-based D&D bot is a perfect match for their philosophy.
“There’s a big difference between an interesting tech demo and something that really resonates with people, ” Stanton says. “And what’s missing is that narrative component. A lot of what we do is that translation between the thing that they already really care about and making something that lots of people on the internet might really care about.”
Neverwinter wasn’t the first time Feel Train took on a D&D character, either. Earlier this year they created a fortune-telling bot, Madame Eva, that took over Wizards of the Coast’s official Twitter account for a few months leading up to the release of Curse of Strahd, a new installment of the company’s Gothic horror series Ravenloft.
Madame Eva's debut in the 1983 D&D campaign Ravenloft marked the first time in the series where story locations could be randomized.
Longtime players may remember Madame Eva from the original 1983 Ravenloft campaign, a soothsayer who would randomize quest locations by drawing from a Tarot-like deck of Tarokka cards to read players’ fortunes — at the time a first-ever design for the series. For Curse of Strahd’s launch, the company was interested in bringing the character back while using social media to interact with fans.
“We wanted to kind of use that mechanic, that kind of randomness of fortune-telling on Twitter, ” says Wizards’ communication manager Greg Tito, who noted the initial idea was to have the company’s social media manager get online for hour-long sessions where he would physically draw from a deck of cards to read fortunes.
“People would get a kick out of that, ” he says.
When the idea was officially proposed, the suggestion that the fortunes might somehow be automated was brought up.
“That got my brain going about Darius’ work, because I’d been following him on Twitter for a long time, ” Tito says. “I remember fondly his Sorting Hat bot that used the four houses from Harry Potter — if you replied to the bot it would tell you which house you were in, and did a rhyming couplet along with it that was randomized. I was very impressed with that.”
Wizards of the Coast released Curse of Strahd, a new Ravenloft adventure, earlier this year.
When Tito reached out to Kazemi and Stanton, they were very excited about the prospect. Wizards sent Feel Train a design document and everyone started working on what parameters the bot would operate in.
“We worked with them on creating mad-libs style phrases that matched up with the Tarokka cards [as well as] with the theme and feel of Ravenloft and Curse of Strahd, ” he says.
Essentially how Madame Eva (and later, the Neverwinter bot) functions is by pulling words — adjectives, verbs, locations and characters — from a series of Google spreadsheets, which are then randomized and placed in one of several tailored narrative templates; these are then run through a parsing program that produces one of countless combinations using proper grammar.
Aside from the challenge of writing in character, a bot like Madame Eva or Neverwinter’s is can’t draw from anything that isn’t written ahead of time.
Though Perfect World's D&D MMO Neverwinter was originally released for PC in 2013, it recently received a PS4 port.
“[A] bot can’t ever put two words next to each other that you didnt at some point say, ‘this word is allowed to be next to this word, ’” Kazemi says. “So when Courtney’s sitting there on the spreadsheet and filling in adjectives and verbs and phases and coming up with, she is going, ‘okay, all of these words would look okay next to all of these words.’”
That aspect of the writing process is intuitive, but it still takes a lot of editing.
Multiple times on the Neverwinter project, I would just generate Courtney this list of 300 [sample] tweets, just for her to go and read through, ” he says. “Like, this template where we talk about someone falling to their death, it doesn’t sound quite right. It looks like we had a ‘the’ where we were supposed to have an ‘a’ in one of the locations.”
Even if a bot has 10 million possible responses from its word combinations, the advantage of a closed bot with a pre-scripted narrative element is that there’s no chance of it becoming unpredictable like Microsoft’s failed experiment Tay, a bot which quickly evolved into a sexually questionable Trump-supporting bigot earlier this year by drawing in new information from myriad unsavory responses.
Microsoft's experimental bot Tay started out innocuous but quickly evolved into a monster as it consumed content from internet trolls.