Part 3 of 4: What the Studios Know (and Won’t Say)
AI adoption inside established game development — the real picture
Here is a number worth holding in your head while reading this piece: 52%.
That’s the share of game developers who told GDC’s 2026 State of the Game Industry survey that generative AI is bad for the industry. Up from 18% two years earlier. At the same time, 96% of studios told Unity’s 2025 report that they’re integrating AI tools into at least some workflows. Google Cloud’s survey puts 90% of developers already using AI in some capacity.
Those numbers don’t contradict each other. They describe the same situation from two angles. Studios are adopting AI broadly and quietly while a majority of the people inside them think it’s a net negative for the craft. That’s not hypocrisy. It’s an industry caught between economic pressure and professional identity, trying to figure out what it actually believes.
Part 2 showed what AI-native game development looks like when one person with professional instincts runs it with intention. The studio picture is messier. More political. And in some ways more revealing about where this is actually heading.
What Adoption Looks Like in Practice
Owen Mahoney, who ran Nexon for a decade, organized the studio-level picture into three layers in his essay “Deus Ex Machina.” It’s a useful frame because it separates what’s actually happening from what gets conflated in most industry coverage.
The first layer is power tools that help humans build faster. This is where most visible studio adoption lives right now. Ubisoft’s Ghostwriter AI generates NPC dialogue across eight major franchises, with a claimed 40% reduction in narrative development time. Embark Studios, building ARC Raiders and The Finals, rebuilt its environment pipeline around procedural generation: satellite laser-scanned terrain, AI-generated trees, weather, and rivers, instead of hand-painted assets. Slower to set up, faster and more flexible forever. All first-party Microsoft studios are required to use AI in production, with $2 billion invested in supporting tools. Sony’s PlayStation Studios has a dedicated AI initiative across its first-party titles.
None of this is vibe coding. It’s targeted AI integration into specific, defined parts of a production pipeline, with human review at every step. The GDC data supports this: the top applications are playtesting and balancing (47%), localization and translation (45%), and code generation and scripting support (44%). Repetitive, high-volume tasks where AI provides clear leverage and humans retain creative direction.
The second layer is what Mahoney calls things humans can’t do: AI systems that act autonomously at a scale and speed no human team could match. Nexon’s work is the clearest documented example. Starting in 2017, their Korea team used AI to detect bots in MapleStory — identifying scripted characters with more accuracy and fewer false positives than manual review, at massive scale. Players noticed immediately. With bot farmers gone, the world felt more alive because every other character was a real person. The team then applied the same approach to matchmaking in FPS games, using AI to analyze keyboard and mouse patterns in real time to assess skill more accurately than kill-death ratios, and to form teams based on playstyle compatibility rather than raw metrics. The result was better matches, higher retention, stronger revenue. No human team could have evaluated and adjusted player dynamics across thousands of simultaneous sessions.
This layer is genuinely different from the first. It’s not AI assisting human decisions. It’s AI making decisions humans couldn’t make at all.
The third layer — AI changing what a game fundamentally is — is where things get speculative but interesting. Whispers from the Star, from Anuttacon (founded by former miHoYo leadership), puts the player as the sole contact for a stranded astrophysicist. Communication happens through voice, text, and video, and the player’s words shape how she survives. Talking to a character who responds to what you actually say, rather than selecting from a dialogue tree, is a qualitatively different experience. It’s a small game. It’s pointing at something large.
The most significant move at this layer, announced in June 2026, came from Roblox. The company acquired Morpheus AI and brought in the founders of Dynamics Lab and Lucid AI, three separate research teams assembled around a specific technical thesis: that the future of multiplayer gaming is photorealistic worlds generated by AI in real time. Roblox calls it Roblox Reality, targeting 4K resolution at 60 frames per second.
Each acquisition addressed a different technical problem that has blocked every other company working on generative world models. Morpheus AI’s Xun Huang invented Self Forcing, a technique that converts slow offline video generation into real-time interactive generation — solving the latency problem. Dynamics Lab’s Joe Chen built a world engine where users upload any image and step into it as a live interactive world — solving the creation interface problem. Lucid AI’s Alberto Hojel invented what Roblox calls the game cartridge harness, which combines AI-generated video worlds with the deterministic logic of a traditional game engine — solving the consistency problem that makes multiplayer unreliable in pure world models.
The Roblox blog was direct about what standalone world models can’t do: “they fundamentally lack the long-term memory, consistent logic, and structured user input required for a massive multiplayer platform.” The three acquisitions are a direct answer to that gap.
Whether Roblox Reality delivers on its technical promises is still ahead. What isn’t speculative is the research talent now assembled to pursue it, and the fact that Google, Alibaba, and Tencent are all working on the same problems. Roblox just acquired named researchers with published solutions to the two hardest ones.
The Politics of Adoption
The technical picture is one thing. The human picture inside studios is another.
Larian Studios, maker of Baldur’s Gate 3, became the year’s most prominent case study in how politically charged this territory is. CEO Swen Vincke told Bloomberg in December 2025 that the studio was using generative AI for ideation during development of their next game, Divinity. The backlash was immediate. Vincke spent weeks clarifying: the studio uses AI to explore concepts internally, not to generate final assets or replace their 23 concept artists. “Any ML tool used well is additive to a creative team or individual’s workflow, not a replacement for their skill or craft,” he said. By January 2026, he had posted a public commitment on Reddit: “There is not going to be any GenAI art in Divinity.” A studio with 530 developers, one of the most critically acclaimed RPGs of the decade on their record, had to make a public pledge about their internal tool usage to avoid losing audience trust.
Obsidian Entertainment, currently developing Avowed, stated flatly that the team isn’t using AI at all. That’s a meaningful data point in a different direction: for studios whose identity is bound up in handcrafted narrative and artistic integrity, the reputational risk of AI adoption may outweigh the productivity gain at this moment in the discourse.
Krafton went the other direction entirely. The South Korean publisher behind PUBG announced an AI-first strategic pivot, offering voluntary resignation packages to employees whose roles it expected AI to absorb. The backlash there came from developers, unions, and press simultaneously.
What this tells you is that “96% of studios are integrating AI tools” means something different depending on who’s doing it and what they’re integrating. A localization team using AI translation with human review looks nothing like a studio using AI to generate final assets. The number flattens a spectrum into a headline.
Where the Advantage Actually Sits
Mahoney’s investor read on all of this is the most clarifying thing written about AI and games in 2025. His argument: AI will lower the cost of producing high-quality games by an order of magnitude. A studio with the right tooling can deliver what was once a $300 million AAA production for roughly $30 million. Embark has already demonstrated this in early form. Others are following.
The implication for incumbents is uncomfortable. AAA publishers built their competitive advantage on the ability to finance and manage large teams to produce content at scale. As production costs soared over the past decade, fewer companies could compete, and that scarcity reinforced their dominance. AI removes the cost basis that made the moat. A 6-person team with the right tools and the right craft skills can now approach quality levels that previously required 200 people. Not in every genre, not for every kind of game, but in more contexts every month.
Mahoney’s position: go short on AAA incumbents whose advantage came from scaling large teams. Go long on small teams free of legacy methodologies.
That framing connects directly to what Adam Clegg described in Part 2: a nimble team of 5-6, each person paired with AI in their lane, no production management overhead, just implementing, playtesting, and iterating. Clegg arrived at that model from the inside of a $50 prototype. Mahoney arrived at the same model from a decade of running one of the largest online game companies in the world. The convergence is notable.
The Problem That Hasn’t Been Solved
None of this means the transition is clean or that the incumbents are simply going to be displaced.
The “gameslop” problem is real. Steam is already managing a flood of low-quality AI-generated titles, and the platform’s disclosure policies haven’t caught up to the volume. Valve is under pressure from larger studios to minimize consumer-facing AI disclosure, which creates a transparency gap that erodes trust across the whole category. When every AI-generated game looks like the worst AI-generated game, the legitimate ones suffer.
The GDC data on developer sentiment reflects something real. Game feel, the quality that separates a satisfying combat system from one that just technically works, requires iteration, taste, and judgment in ways that AI generation doesn’t address and may never fully address. The developers who are most negative about AI in games aren’t wrong about the risk. They’re worried about the right thing. The question is whether the risk materializes at the craft level or at the business level, and the evidence suggests it’s mostly a business-level disruption so far: it’s changing who can afford to make games and how, not yet changing what the best games feel like to play.
That distinction, between what AI changes about the economics of game development and what it changes about the experience of playing games, is where the industry’s real argument is happening. It’s also where the most interesting work is being done.
Part 4 looks at what that work points toward.
In Part 4: Where the path from vibe coder to game developer leads — and what the medium looks like when the tools that built it become available to everyone.
Sources and further reading
GDC 2026 State of the Game Industry survey https://reg.gdconf.com/2026-SOTI/
Google Cloud’s 2025 game developer AI report https://cloud.google.com/resources/games-report
Unity: AI in game development report, 2025 https://unity.com/blog/2025-unity-gaming-report-launch
Kotaku — “Larian Studios Leaning Into Generative AI Despite Internal [backlash]” https://kotaku.com/larian-studios-gen-ai-divinity-bg3-2000653850
Game Developer — “Larian CEO says any AI tool ‘used well’ is additive” https://www.gamedeveloper.com/production/larian-ceo-says-studio-is-more-or-less-ok-around-gen-ai-use
Embark Studios — “The Content Revolution to Come” https://medium.com/embarkstudios/the-content-revolution-to-come-f2432dc6a434
Nexon/MapleStory AI — Korea Herald https://www.koreaherald.com/article/3296675
