7 Min Read

Last week, Ubisoft became the latest publisher to hitch itself to the generative AI bandwagon with a view to creating previously unimaginable experiences. At least, that’s the pitch. During a behind-closed-doors briefing attended by Game Developer, the company unveiled an R&D experiment called ‘Teammates’ that uses generative AI technology—Google Gemini, in this instance—to turn NPCs into “living companions” that can recall player information, follow contextual commands, offer insights about tasks, dispense lore, tweak settings, or simply have a chinwag. 

After a detailed briefing, hands-on demonstration, and series of panels, we had the chance to sit down the key members of Ubisoft’s generative AI team to learn more about how Teammates works, why Ubisoft is leaning on third-party LLMs like Gemini, and whether it’s possible to leverage the technology in a way that feels ethical and intentional. We kicked off the conversation by asking why Ubisoft chose to use technology like Gemini and ChatGPT when owners Google and OpenAI are both facing litigation (thanks, Bloomberg) for how they’ve trained and deployed their AI tools. 

There’s a sense that major AI companies have adopted an ‘ask forgiveness, not permission’ approach when training their models on information scraped from the internet—including the work of numerous creatives, who aren’t exactly thrilled with the situation (via Reuters).

Related:Unity developers will soon be able to bring games to Fortnite

As acknowledged by Ubisoft itself during our hands-on demo, being tethered to third-party LLMs can also cause technical headaches. For instance, the company suggested we might encounter a few hiccups during our session because of the Gemini 3 update that was rolled out a few days prior. 

Ubisoft’s director of generative AI gameplay, Xavier Manzanares, said worries about being tied to one specific model prompted the company to create a robust middleware platform that allows it to pivot on a dime. 

“It’s a modular platform, because we didn’t want to be bound to one model or another. So basically, we wanted to have the choice of whatever one to use based on the needs that we have. Today, it could be a model that is based on Open AI, it could be Claude, or it could be Gemini,” said Manzanares. “Tomorrow, it could be something else. It could be linked to the innovation of some models that are really cool that we want to swap with. Or it could be just because we want to have it in-house and we want to build a model ourselves. If we don’t do that, it means that, basically it will just fumble, because if there’s something new coming and it destroys everything that we thought [we knew] when it comes to prompts, it’s just a fumble.”

Related:Dispatch’s episodic format – alongside sharp storytelling – is key to its success

Manzanares said LLMs each have a “different form of intelligence” and are unique in terms of how they handle performance, which is why Ubisoft feels it’s vital to build a system that’s adaptable. Ubisoft’s director of data and AI, Remi Labory, added that the company isn’t betting on a model, but rather a broader “creative vision.”

“Tomorrow, there will be new models. State of the art. Open source. They will open possibilities and we are going to identify whether smaller models are capable of solving our problems,” adds Labory. “So we have to be pragmatic about where to start, but also making sure that it’s our vision and that we can adapt choices and focus on smaller layers.”

“It’s a super cool tool, but it can do bad stuff”

But what about those aforementioned ethical and creative considerations? Ubisoft narrative director Virginie Mosser, who worked on characters present in the Teammates experiment, said she had lengthy discussions with the AI team about how to merge her writing nous with their technical expertise to “keep the creator at the center of the development team.” 

In one instance, she explained that meant being upfront with voice actors about the nature of the work being done and compensating them fairly. It also, she added, meant acclimatizing to a “new way of writing.” 

Related:Lumines Arise developer Enhance is forever chasing God

Mosser admitted she still has some concerns about the use of generative AI, but said she wouldn’t be using the technology if she thought it was incapable of stirring emotion. “I don’t want my work being flat and cold,” said Mosser, who also claimed she held a lot of misconceptions about generative AI that came from people viewing it as a shortcut that can be used to quickly generate content. 

By contrast, she believes that if creatives work together with AI teams they will be able to create characters and stories that truly resonate. “This is a real focus for me because this is how we can make the difference in the future,” she continued. “Put a lot of creative people with AI teams and merge them today to ask ‘what does the future look like for players?’ We have only just scratched the surface and we are aware there is a lot of work to do with emotion systems and  relationship systems—but we are focused. 

“For me, we still have the choice and capability to say ‘we want to add more emotion here.’ This is an iterative process. We work a lot on the character sheet and content, and sometimes it doesn’t raise emotion.”

Manzanares said that process lowers the risk of churning out “generic slop.” Labory agrees, and underlines the importance of feedback from internal teams and playtesters, who will be candid about when something needs reworking.

“We have internal tests where we can ask ‘what do you think about this character? How do you describe this character? What did you feel when playing with this character?’ We check they feel something, otherwise it’s back to the drawing board and we get together with the machine learning experts and the narrative teams and ask ‘Okay, what’s wrong? Is it a question of instruction? Do we need a new process to ensure what we have more examples and more evaluations to ensure that it’s something that we can measure—or something that we can’t.”

Labory said the AI team would love to have clear and obvious criteria for success but acknowledges that sometimes characters and gameplay elements either click or fall short because of intangible feeling. That aspect of the process, he added, is what unites the AI team. “By having this interaction, and by working on something that is close to design and creation, we are discovering something new about our jobs and what we create,” he continued.

To that end, Manzanares said Ubisoft approaches all of its AI experiments by starting with “something that matters.” 

“It’s cool to have potential prototypes, but if we don’t see the code behind it and if we don’t know how it integrates into a potential project or our pipelines, it’s not for us,” he explains. 

The aim, he claims, isn’t to simply produce glitzy showreels, but to provide Ubisoft’s internal teams with tools that will help them ship interesting video games. 

“It’s a wild world right now. There are a lot of extremes [around generative AI]. Actually, on our side, it’s not just about building experiences and middleware, but also trying to teach everyone what is happening behind the scenes,” he said.

“It’s a super cool tool, but it can do bad stuff. Everyone needs to know that and we need to be humble about that fact. If we didn’t do the work that we did, it would just create generic content that will not be distinguishable between experiences. It will reduce creativity to something very generic, and this is very scary because it means that everything we did in video games–and I’ve been in this industry for more than 20 years–would all just go down the well.” 

Game Developer attended Ubisoft’s press event in Paris with travel and accommodation provided by the French publisher.

About the Author