Ubisoft’s AI NPC project can now deliver sarcastic GMs, helpful team-mates and hidden lore – and is being playtested now

Ubisoft’s first AI showcase was all about conversation. Its second, Teammates, is all about combat, and a group of three personalities trained up by an in-house narrative team for you to go into combat with.

It’s the product of an expanded internal team (now up to 80 people) developing in-house middleware that connects a wide range of external AI models with the company’s Snowdrop and Anvil game engines, and supports internal teams in building that into their games. Concurrently, it’s developing its own “playable experiences” – the team is careful not to call them “demos” – that highlight potential applications. The company let me squad up with the Teammates at an event in Paris this week, and revealed that it’s being put through its paces by both internal studios and external testers, touting feedback from “hundreds” of users on its Ubisoft Connect program.

Teammates is a notable change from the team’s Neo NPC debut at GDC 2024, which was based on conversations with AI NPCs which unlocked RPG-style relationship impacts. This had a supportive but subdued reception from internal teams, said Manzanares, who couldn’t see it fitting the types of games they were making. In response to this, Teammates feels like something that could easily wear a “Tom Clancy’s” prefix.

It’s based on a couple of AI-powered cyborg soldiers you can command through a few simple encounters with small squads of bullet-sponge robots, guided by a third AI companion which has a more GM-like control over the experience. The soldiers, Pablo and Sophia, have their own personalities (stoic and bubbly, respectively, which the demo offers a few different flavours of for both) and background lore which can be surfaced through conversation as you play. The companion, Jaspar, delivers GM-like insights with a Ryan Reynolds-esque smarm. The whole thing is powered by Google’s Gemini platform.

At the start of the mission you have team-mates but no weapon, and need to direct them from cover for the first encounters. Using voice commands brings back unwelcome memories of Xbox’s late and unloved Kinect, but this works in a way Kinect rarely managed – I was able to give snappy instructions like “stand on those pads to open the door” or “flank those soldiers and stay in cover” and Pablo and Sophia would do it, with varying degrees of sass depending on which personality had been assigned.

Both would periodically volunteer suggestions for plans of attack or abilities like being able to concentrate fire, using what’s termed the Initiative system. This became increasingly useful as the level progressed, because even when granted a weapon it was notably useless and each enemy squad took time to grind down. Botched tactics lead to Pablo or Sophia being downed and requiring resuscitation, and it’s possible to demand them do the same to you.

Beyond just bossing them around, picking at conversational asides unearthed differing backstories for each; I was able to tease out Pablo had seen a previous squad wiped out, and Sophia cheerily glossed over the practicalities of cyborg dining habits. I didn’t want to spend my demo time laboriously mining them for lore, but I did serendipitously discover it, which is the result of what the R&D narrative team said was extensive preparation, briefing and training.

Jaspar, meanwhile, was authentically annoying: self-satisfied and occasionally belittling, although seemingly responsive to requests to be less so. It had wider control over the game – I could ask it to highlight soldiers on the HUD, change the colour of the HUD, or tell me what I’d just picked up so I didn’t have to go into the D-pad menu and find out.

It’s also in charge of giving out achievements, the ten or so in that the demo contained being unlocked when the AI deemed the requirement to have been met, rather than when the player satisfies a fixed criteria. This “sounds simple,” says Manzanares , “but it’s quite complex. We have generative AI produce content that is not just talk but a game system – like a complex systemic quest system. How can an LLM play the live producer role? That’s something that we’ve been working on.” He also points to the significant accessibility potential, with a range of applications stretching from adjusting the game experience for players with disabilities through to reminding them what they’re supposed to be doing after returning to a long-lost save.

I did not enjoy spending time with Jaspar’s arch wisecracks (apparently the playtesters find it better company), but that was because of its character and that does represent a success for the project: Narrative Director Virginie Mosser was delighted when I described it as “sassy” because that was a key characteristic on what the team call the “character sheet” for each team-mate.

“We want to create characters,” she says. “The worst is when somebody says ‘I don’t like him, I don’t hate him.’ I want a player to say “I hate him” because then it’s a feeling. They need to provoke something.” She’s keen to emphasise the amount of work this takes on the behalf of the creative team, at a time when there’s escalating anxiety in the industry about creatives losing work to generative AI.

“We have to work a lot to prepare, almost as if they were actors,” she says. “Giving them the rules of improvisation takes a lot of time. Whatever you write in the backstory has an impact when you play afterwards”. The key element here is the creation of the “character sheet” which outlines the NPC’s backstory and personality, which is then fed into the prompt – and then tuned, tested, and iterated, alongside working with voice actors. “It’s a lot of creative work and it’s good news for us,” says Mosser, “because it’s good news for creators.”

Manzanares sees this as becoming a key differentiator for Ubisoft in a potential future where AI-powered NPCs become ubiquitous. “The character sheet and the prompt is a big, big thing for the team,” he says. “That’s where we put a lot of effort into the tools, how to write with them. Because in the future this is where it will make the difference, versus a generic model that maybe everyone has.”

Teammates is just one output of the team’s ongoing efforts to create a central platform that enables Ubisoft developers to plug AI tools into their games. “We have to build a middleware that connects to the engines first so that those teams don’t have to do it,” says Manzanares “Everything around gen AI – the tools, the models, what happens with Gemini, whatever new announcement – they don’t have to care about that. We’ll take care of that. Our goal is to be plug and play so that they can focus on their game.”

Meanwhile, it will continue to develop samples like Teammates to demonstrate what’s possible. “We have to continue building small experiences, things that could be playable,” says Manzanares “so that creative directors and maybe even players can start to check them out and maybe and go further.” Teammates’ resemblance to Ubisoft’s existing output is expected to help with this. “It’s FPS and squad-based and even the settings are familiar,” says Remi Labory, Data & AI Director, “and I think that talks to the team. It’s a vocabulary they have. It helps reflection on how to create new gameplay bricks using AI, if you already have the vocabulary of the game you’re trying to work with.”

The R&D team’s focus is rapid iteration: the new build was in players’ hands in October, and the roadmap calls for monthly updates through January. “We will add more features to Jaspar, so you can do more stuff,” says Manzanares. “We’ll add the emotion system: if you talk badly to Pablo, he will not be as happy, we will change emotion in real time.” The expectation is to test, learn and respond to user feedback – and to roll with whatever changes arrive with new AI models, because the platform means that they can be swapped in and out as required.

There’s an expectation that Ubisoft will develop its own in-house AI modes, but none were named or mentioned; the team wouldn’t be drawn on plans or costs, preferring to focus on the benefits of picking and choosing from external models currently available. Given the significant financial and environmental costs of those models, that reckoning will come due at some point, but for now the priority is identifying potential use cases.

The initiative as presented appears a lot more creator-focused and opt-in than the reported AI mandates of Krafton and EA. Ubisoft’s approach is to build a central resource, and leave it to the internal game development teams to find an application for it (tempted, the team hopes, by “playable experiences” like Teammates.) While the development of the tool is moving rapidly, and the internal reception is described positively, its implementation in shipped games will take a while longer.

“A lot of internal people are like, ‘okay, does it match with my game vision? I need to think about it.’ Some people are more like, ‘I need to ship soon so I’ll come back to you next year,'” says Reynauld Francois, Creative Director on the project. “Lots of people see the potential – whether that potential fits in their vision for their game today, I don’t know. But maybe the next one.”