Let’s get the hot potato out of the way without preamble: Embark’s extraction shooter Arc Raiders doesn’t include any gun models generated from Youtube videos, executive producer Aleksander Grøndal has told RPS in an interview about the game’s usage of generative AI and machine learning technologies.
This clarification follows the partial online publication of an Edge magazine interview in which Embark CEO Patrick Söderlund made various claims for the studio’s in-house tech, including the suggestion that the developers “can take a video from YouTube, feed it through our tools and pipelines, and [produce] a 3D model of the weapon you had in that video.” According to Grøndal, this particular technology is not actually used in Arc Raiders. “That’s a research project, and that’s not something that we’re using in the game now,” he told me over a video call this Monday, following a sprawling and generally enjoyable hands-on.
If Arc doesn’t use AI generation for gun models, however, it does make use of generative AI. Ahead of our early access review this week (which I’m not writing, though I may put together a Take or two), I asked Grøndal to clear up some details.
According to Grøndal, Arc’s in-game generative AI implementation is limited to voiced barks – that is, exclamations such as “let’s go to objective B” – and similar, shorter pieces of more functional audio. These were produced by applying software models to voice data supplied by actors, a practice known as “text to speech”. As in the developer’s previous The Finals, the voice actors in question have given consent for their recordings to be used for this purpose.
“We find the voices that we like, the actors that we like, and then we make a contract with them about recording them,” Grøndal explained. “So we use them for the dialogue, generally speaking, that’s live recording. Then we also create a clone out of them, and we use that for small, shorter things in-game, like item names and stuff like that, or callouts in the game.”
The AI generation of videogame voicework has been a focus in discussion of whether generative AI adoption will lead to job cuts and a general erosion of working opportunities, with big budget tools such as Microsoft’s Copilot often described as a way of magically increasing productivity without raising costs. In July last year, the US actors union SAG-AFTRA called a strike, demanding that videogame publishers agree to proper constraints on what companies can do with recordings, to protect the livelihoods of performers.
The Finals has come under fire for its usage of text-to-speech, with Embark protesting back in 2023 that “making games without actors isn’t an end goal.” In my interview about Arc Raiders, I asked Grøndal for more specifics about Embark’s voice actor contracts – are there any limits on what the developers can do with that voice data? Grøndal was unable to go into specifics, partly because he didn’t have the information to hand, and partly for the usual confidentiality reasons, but I’ve followed up over email to ask if Embark can share some of the broad strokes.
Grøndal was, however, able to tell me more about other practices and technologies used in development of Arc Raiders that he thinks need to be explicitly distinguished from “generative AI”. “Apart from [text-to-speech], we don’t use any generative AI in the sense that we create content around that – maps or meshes or textures,” he continued. “We use photogrammetry, which is pretty common in the games industry today, where we take multiple photos of objects, and then we put them into software, they help us create a model from it. But that’s not [generative] AI – that’s been around for 15 years or so.”
He also discussed the developer’s use of machine learning for animations. To give some brief context, machine learning is an umbrella term for the study and development of algorithms that can perform tasks without explicit instructions, which dates back several decades. Much like procedural generation – whereby developers devise rules for generators, rather than having the software derive its own rules from patterns in datasets – machine learning informs today’s better known generative AI tools, such as ChatGPT. But developers often use it to refer to technologies that pre-date the current genAI craze.
In his Edge interview, Söderlund discussed the use of “AI” to create animations for certain enemies. “What if, instead of having to hand-animate or use mocap for every single frame, we created a model, gave it physical attributes, then used AI to train it to walk?” he said. Last week, I linked this remark to a Medium article by Embark software engineer Tom Solberg, posted back in 2019, which describes how in-game spiderbots can be set up to walk believably by subjecting a software agent to different stimuli in a 3D test environment.
“That’s machine learning that we have done, and that’s something that has been a research [project] for us for a long time, and that helps us to get some of that physicality into our robots,” Grøndal explained, when I asked about the relationship of this tech to Arc Raiders today. The technology is designed to “emulate what we have already created from hand,” he added, and the output isn’t a fully functional arachnobot adversary – it needs further work before it can form part of a videogame. “It’s nothing like, ‘hey, make this thing do it, and then off it goes’,” Grøndal said. “That’s far from it. It’s more that it can help support creation of some of the robots.”
While Solberg’s 2019 post does cite machine learning as a way to create fancy animations with a smaller team, Grøndal also insisted that Embark’s machine learning experiments are not capable of “replacing” developers – or at least, not for the moment. “That research that you’ve seen on the Medium post is from a long time ago, and none of that is necessarily applicable in runtime, as of a game engine, as of right now,” he said. “I can’t remember exactly the context in which that article was written, but it is not our ambition to replace any developer.
“Working with that type of technology is very hard,” he continued. “It doesn’t solve any immediate problems for you. So you will see some of our robots using machine learning, like you talked about, and just getting them to the point where [it doesn’t] completely break the illusion of some sort of robot – that was a massive amount of work. There weren’t any shortcuts there, to put it mildly.”
Current discussion of generative AI is both justifiably heated and very confused. I think that’s partly due to ignorance on the part of journalists – I’m not sure my own reporting on generative AI has been very accurate or coherent, though I stand by this piece on megabudget genAI “copilots”. The videogame specialist press is structurally ill-equipped to deal with subjects like this: few of us have either the technical expertise or the thinking time to keep up with and critique such a rapidly changing sector. But a lot of the confusion is also down to freewheeling rhetoric on the part of executives and investors, who aim simply to hype generative AI as a concept, conflating new and older technologies that have very different applications.
If Embark were to eventually use generative AI to fill out the game’s arsenal, it would make for some bleakly satisfying irony, because Arc Raiders is all about recycling scavenged parts. The game takes place in a world overrun by homicidal robots, ranging from waspish flying drones equipped with rapid-fire turrets to pouncing arachnid monstrosities. The surviving humans live in underground cities, with certain brave souls venturing to the surface via fortified elevators to harvest crafting components such as old batteries and cloth.
As such, all the game’s gun designs are DIY fantasies, rather than the flashy military boomsticks you find in similar shooters. There are functional correspondences to real-world light machineguns, scoped rifles and such, but everything broadly looks like it’s been hammered together out of fridge motors and coathangers. “We wanted them to have this makeshift feel, as a contrast to this perfect ARCs who have almost polished surfaces,” Grøndal told me. That improvised aesthetic also helps the weapons feel disposable, which is useful for a game whose hook is looting. “It’s part of the understanding with the Raiders that a weapon isn’t permanent,” Grøndal went on. “It’s not you and this rifle for the rest of your life. If it breaks you cobble something else together.”
It’s perhaps too tempting to read Arc’s sheeny hunter-killer bots as metaphors for current fears about generative AI. Their smooth and seamless design, together with the mystery of their origins, makes me think of how opaque companies like OpenAI can be about which data they’re using to “train” their all-singing, all-dancing artificial companions. Arc Raiders at least grants me the catharsis of blowing these wily bots to bits.
- Security Camera Installation – indoor/outdoor IP CCTV systems & video analytics
- Access Control Installation – key card, fob, biometric & cloud‑based door entry
- Business Security Systems – integrated alarms, surveillance & access control
- Structured Cabling Services – voice, data & fiber infrastructure for new or existing builds
- Video Monitoring Services – 24/7 remote surveillance and analytics monitoring
Author: 360 Technology Group
.jpg?format=1500w&ssl=1 "Oathbound release \"Set Adrift\" music video")