Eva Vital on training AI to recognise insults and not rise to bigotry

In a talk from Pocket Gamer Connects Helsinki in September, Eva Vital from Fortis Games discusses training voice-recognition AI for games, particularly rudeness.

Eva Vital on training AI to recognise insults and not rise to bigotry

In case you missed Pocket Gamer Connects Helsinki, our largest ever event in the Nordic regions, we are still bringing you some of our highlights of the event. This piece covers the new, burgeoning field of utilising voice AI in games, specifically training them to recognise conversational phrases…via insults. You may be wondering how swearing at a video game character can create an interesting way to explore gameplay options, so let’s explain how.

“Today I’ll be mostly talking sh*t. Or rather, training AI on how to respond to insults in games”, is a rather striking way to open a talk. But Eva Vital of Fortis Games has a background with AI voice games, such as The 3% Challenge, to back her humour. She first introduced us to the basics of how AI is trained. Whereas many devices, such as Amazon Alexa or Google Assistants, use pre-trained phrases as simple as ‘yes’ or ‘no’, for a video game with a conversational setting, this simply isn’t adequate.

Talking Heads

As Eva pointed out, in many games with optional dialogue choices, such as Zaum’s Disco Elysium, you will usually have unvoiced protagonists where players have the luxury of reading through dialogue options in full before selecting them. Whereas in fully-voiced RPGs, like Mass Effect, you’ll have characters respond much more extensively based on only a single-sentence prompt, as the former would be less accepted in a fully voiced game. In the past, text-based games allowed players to write out their own dialogue, but if these weren’t recognised the game would simply cut them off.

Whereas, in a game that utilises voice input in full, AI is an integral part of making dialogue seem natural, and enabling an AI to recognise a wide variety of phrases is key. In this case, the intent is to make conversation with the AI seem real. As Eva pointed out when she worked on The 3% Challenge, they noted that one naturalistic part of human dialogue is insults. Thus, the team programmed in the ‘Rude Intent’ update. Whereas previously the ‘Fallback Intent’ (the AI recognising the closest possible match to their pretrained programming in a conversation) was used for insults, the game was updated so that being rude or insulting would affect player’s interactions with characters in-game.

As Eva pointed out, this can be not only a fun easter egg for players, but actively aid in building the personality of characters in story-based games. Having in-game characters she described reacting in different ways, helped inform the player of their roles in the narrative even though it did not affect the plot of the game. Not only that, but benefits such as immersion and accessibility through the use of voice commands can be extremely beneficial to the overall gameplay.

No Reaction

Eva goes on to discuss another key point, recognising bigotry - or rather not recognising it. According to her reasoning, whereas a person should not be under obligation to ignore bigotry, a character can, and given that “It felt to us that when bigoted people used slurs it was usually to get a reaction.” By refusing to offer reactions to that, whilst the interesting narrative element of rudeness can be maintained, it would not encourage bigoted behaviour. “By deliberately ignoring bigotry, we wanted to give no satisfaction to this type of users.”

PocketGamer Connects has many more events coming up, with just as many informative and fascinating talks such as this. Our next event is in Jordan this November, tickets are available now.

Staff Writer

Iwan is a Cardiff-based freelance writer, who joined the Pocket Gamer Biz site fresh-faced from University before moving to the editorial team in November of 2023.