Embark Studios has again defended its use of generative AI to create voice lines in Arc Raiders, an extraction shooter released earlier this year. Chief creative officer Stefan Strandberg told viewers in a new audio-focused video that the studio is not in the business of replacing actors, but that feeding existing performances into generative models was the only practical way to meet the volume of dialogue required. The comments arrive as the studio also confirms it has begun swapping some of those AI lines for fresh human recordings.
The game drew criticism both before and after launch for training models on real voice-actor performances and then synthesizing additional lines from the same data. Strandberg acknowledged the scale problem directly, stating that actors “would just go crazy if they had to read all of that.” Chief executive Patrick Söderlund has separately confirmed that the studio is now replacing some generated lines with newly recorded ones, a shift the company links to stronger-than-expected commercial performance.
The original approach and its defense
Embark’s initial pipeline took recorded performances from voice actors and passed them through generative systems to produce further dialogue. The studio has framed this step as a response to the sheer quantity of lines needed for an extraction shooter with procedural and player-driven scenarios. Strandberg repeated in the recent video that the goal was never to sideline performers, only to manage output that would otherwise have required performers to deliver an impractical amount of material.
Both Rock Paper Shotgun and Eurogamer published coverage of the new statements on the same day. The reports note that the latest remarks closely track earlier comments from Söderlund, who described the post-launch replacement of AI lines with human recordings. The studio has not disclosed how many lines have been re-recorded or how many remain synthetic.
Commercial performance and course correction
Söderlund has tied the decision to increase recorded dialogue to revenue the game has generated since release. The studio has described that performance as strong enough to support additional recording sessions. This sequence—initial reliance on generative tools followed by partial reversion to conventional recording—has become visible only after the title reached market and began earning.
The two source reports differ slightly in emphasis. Rock Paper Shotgun highlights the volume argument and the studio’s insistence that it is not replacing actors, while Eurogamer focuses on the claim that the AI step ultimately benefited voice actors by sparing them exhaustive recording sessions. Neither report contains new data on the exact proportion of lines still generated by model.
Why it matters
The episode illustrates how generative audio tools move from experiment to production pipeline and how studios later adjust when revenue appears. Embark’s position rests on the assertion that full human recording of every line was never feasible at the required scale. Yet the same studio is now able to increase recorded content once the game is earning. That timing undercuts any claim that the choice was driven solely by artistic or technical necessity rather than cost.
For voice actors and their representatives, the case adds a concrete example to ongoing questions about consent for training data and the durability of work once models exist. For other developers, it shows both the short-term appeal of scaling dialogue with existing recordings and the longer-term cost of public criticism once revenue allows a partial reversal. The pattern suggests that commercial results, more than stated principle, will determine how quickly studios move away from synthetic voice work after it has already shipped.
---
Sources:
No comments yet