At this week’s Game Developers Conference, Ubisoft unveiled a tantalizing vision of a future filled with AI wonders for gamers. The company showcased a groundbreaking prototype that utilized Nvidia’s Ace microservice to create “smart NPCs” with fully voiced capabilities, allowing players to engage with them through spoken interaction. Despite initial skepticism, the hands-on demo left a lasting impression as I found myself engrossed in a nuanced conversation with an NPC discussing the complexities of eco-terrorism, all made possible through the marvels of AI technology.
While this demonstration showcased the immense potential of this new technology, it also shed light on a significant challenge that even Ubisoft is grappling with: linguistic bias.
AI’s subtle prejudices
In the brief demo, I assumed the role of a space-faring adventurer entangled in a rebellion against a powerful megacorporation. Interacting with two meticulously crafted NPCs from Ubisoft, I delved into the intricacies of a sci-fi realm, learning about the characters’ backgrounds and engaging in strategic planning for a daring heist.
Upon completion, I took the opportunity to converse with two Ubisoft developers involved in the project to inquire about any lingering issues with the tool. While they expressed enthusiasm for the technology as a whole, their expressions betrayed a sense of frustration as they acknowledged a series of challenges that still needed to be ironed out before widespread adoption. Chief among these concerns was the inherent bias rooted in the English language, a human flaw that AI is currently absorbing by default.
Invisible barriers
The experts pointed out specific instances that had eluded my notice during gameplay. For example, when I inquired about a character’s least favored team member, the NPC initially professed affection for all comrades before singling out a character named Iron as problematic. The catch? Iron was not intended to be male.
According to the developers, the AI tool appeared to associate the term “iron” with masculinity, leading it to assume that the character Iron was male. A similar issue arose with a character named Bloom, portrayed as a nurturing, maternal figure. This reversal of the Iron problem highlighted the machine’s tendency to ascribe gender stereotypes to characters based on linguistic cues, raising questions about how generative AI dialogue would navigate issues of race and identity.
These intricacies of language reflect a larger trend in AI development, where language models like the one used here draw from vast datasets that may inadvertently perpetuate biases and stereotypes. While Ubisoft’s NPCs boast rich, hand-crafted backstories, they are still influenced by broader language learning sources, complicating efforts to combat ingrained prejudices within AI systems.
Navigating uncharted linguistic waters
The complexity deepens when considering the challenges of applying generative AI to languages beyond English. Developers note that current AI models are heavily skewed towards English, presenting substantial localization hurdles for non-English speakers. As nuances and conventions vary across languages, the task of adapting a machine designed for one linguistic framework to another poses a formidable obstacle in the quest for inclusive AI solutions.
Ubisoft and Nvidia remain actively engaged in addressing these limitations, underscoring the importance of collaboration with Diversity, Equity, and Inclusion (DEI) teams to identify and rectify problematic language output generated by AI characters. While the prospect of engaging in profound discussions with AI NPCs is undeniably enticing, it is evident that these machines still require significant human intervention to rectify the unintentional biases ingrained in them.
Support our work ❤️
If you enjoyed this article, consider leaving a tip to help us keep publishing great content.
