Larian draws a clear line on GenAI for Divinity's art and writing - and keeps testing it elsewhere
Larian won't use generative AI for concept art or writing on the next Divinity. After community pushback and mixed headlines, CEO Swen Vincke clarified the studio's stance and tightened the guardrails.
Here's the useful part for writers: they're still experimenting with AI for speed and iteration, but not for final creative output without full control of training data and consent. That balance is the signal through the noise.
What changed - and why
After announcing a new Divinity, Vincke told Bloomberg the team had tested AI to explore ideas, build early presentations, generate placeholder text, and rough visual prompts. None of that was meant for shipped content.
The backlash was predictable: even early-stage use can look like replacing human ideation. Vincke responded that Larian isn't swapping artists for prompts and highlighted the team's size - 72 artists, including 23 concept artists, with more hiring underway.
To remove doubt, he set a firmer line: "There is not going to be any GenAI art in Divinity," and the studio will refrain from using genAI tools during concept art development at all. The goal is simple - no questions about provenance.
Writing stays human
Writing director Adam Smith said there's no text generation touching dialogue, journal entries, or other in-game writing. A small group has tested tools, but the quality landed "3/10 at best," and that work stays in a research sandbox.
Translation for narrative teams: AI isn't clearing the bar for voice, consistency, or nuance. Human drafting remains the standard.
Where AI might still help
Larian is still running controlled experiments to speed iteration across departments - faster prototyping, tighter feedback loops, less waste. The intent is a more focused dev cycle and, ideally, a higher-quality game.
On in-game assets, there's a hard rule: they won't generate anything without knowing exactly what the model was trained on and having consent for that data. If they do use a model, it will be trained on data they own.
Why this matters for writers
- Set non-negotiables: no AI in shipped narrative content unless you control training data and rights. Make provenance part of your style guide.
- Use AI for speed, not voice: concept notes, outlines, research patterns, and iteration ideas - then rewrite by hand.
- Sandbox everything: keep experiments out of production until a human approves and adapts the output.
- Quality bar first: if the tool isn't hitting 8/10 or better, it's not saving time once edits start. Kill the tool or the use case.
- Be transparent with collaborators and audiences: clarity defuses speculation and protects your team's reputation.
- Invest in people: Larian's move doubles down on artists and writers while using AI to reduce waste elsewhere. That's the sustainable path.
The bigger picture
Generative AI's footprint in games is growing fast. Steam disclosures of AI use reportedly jumped ~800% year over year, with thousands of titles now flagging some level of AI involvement. The market is experimenting; the best teams are pairing that with strict consent and clear boundaries.
Bottom line
Larian's stance is the pragmatic middle: human-led creative, AI-assisted iteration. For writers, that's the workable model - protect voice, guard rights, and use tools to think faster, not to speak for you.
If you're building your own AI guardrails or testing prompts for research workflows, this quick hub can help: Prompt Engineering resources.
Your membership also unlocks:
