There is a version of the AI debate that runs entirely on vibes: a six-fingered zombie Santa in a loading screen, a suspiciously smooth texture, a Discord thread full of people zooming in on hands. That version is exhausting and mostly unprovable.
There is also a paper trail. Publishers talk to investors four times a year, and investors ask about costs. That is where the admissions live, in progress reports, earnings calls and Steam store disclosures, written in the flat language of a company that has already made the decision.
This is that paper trail, studio by studio, with the document and the section each claim comes from. No inference, no zooming in on hands.
EA: "at the very core of our business"
EA said the quiet part out loud at EA Investor Day 2024 in September 2024. CEO Andrew Wilson told the room that "AI is at the very core of our business," and EA put a number on it: more than 100 active AI projects, sorted into three buckets it called efficiency, expansion and transformation.
Thirteen months later it signed the partnership. The Stability AI announcement of October 23, 2025 names two first initiatives explicitly. One is generating physically based rendering materials, the 2D textures that have to hold their colour and light response across a whole environment. The other is AI systems that pre-visualise entire 3D environments from prompts, so artists can creatively direct generation instead of blocking it out by hand.
That release also confirms Stability is "embedding our 3D research team directly with EA's artists and developers," which is a much deeper arrangement than buying a licence. Stability CEO Prem Akkaraju framed it as innovation beginning "with the creator," and EA VP of creative innovation Kallol Mitra said the pair are "amplifying that creativity."
The counterweight came from inside. A Business Insider investigation in October 2025, summarised by VGC and others, reported that staff were being pushed to use AI for close to everything, and that EA's internal chatbot ReefGPT was hallucinating and writing code developers then had to unpick. Mandated adoption and useful adoption are not the same metric, and EA is the clearest case of the gap.
Krafton: the full "AI-first" conversion
Krafton went furthest. In an investor document dated October 23, 2025 and repeated in its Q3 2025 results, the PUBG and Subnautica publisher declared itself an "AI-first company," with the transition targeted for completion in the second half of 2026.
The spending is specific. Krafton committed roughly KRW 100 billion, about $70 million, to an NVIDIA GPU cluster, and partnered with SK Telecom on a five billion parameter in-house foundation model. Its first Co-Playable Character, PUBG Ally, was slated for the first half of 2026. Game Developer's write-up quotes the earnings language directly: the pivot is meant to enable "automation of new and existing tasks" and "large-scale creative experimentation and organizational transformation."
Then came the part nobody puts on a slide. On November 12, 2025, weeks after the announcement, Krafton opened a voluntary resignation programme to all staff, with severance running from six to thirty-six months of salary depending on tenure. Inven Global later reported that 271 people left the group in January 2026. Whatever you want to call that, it is not a workflow change.
Square Enix: 70% of QA, by the end of 2027
The bluntest target in the industry sits on page 5 of Square Enix's Progress Report on the Medium-Term Business Plan (FY2025/3 to FY2027/3), published on November 6, 2025.
Under the strategy pillar "Roll Out Initiatives to Create Additional Foundational Stability," a section headed "Promoting AI utilization in Japan" describes a joint development project with the Matsuo-Iwasawa Laboratory at the University of Tokyo. The goal is printed verbatim: "Automate 70% of QA and debugging tasks in game development by the end of 2027. Through the use of automation technology, aim to improve the efficiency of QA operations and establish a competitive advantage in game development."
The same page lists the structure: a joint research team of more than ten members drawn from the lab and from Square Enix group engineers. Note what it is not. It is not concept art, and it is not dialogue. It is QA, already the most outsourced, most contract-heavy and least protected job in the building.
Take-Two: hundreds of pilots, zero in GTA VI
Strauss Zelnick spent a while as the industry's designated AI sceptic. That framing did not survive Take-Two's Q3 FY2026 earnings call in February 2026, where he said the company is "actively embracing generative AI" and has "hundreds of pilots and implementations across our company, including with our studios."
He drew one hard line, and it is the interesting one. Speaking to GamesIndustry.biz, Zelnick said "generative AI has zero part in what Rockstar Games is building," describing Rockstar's worlds as "handcrafted" and built "building by building, street by street, neighborhood by neighborhood." Not procedurally generated, in his words, and they should not be.
Read those two statements together and you get the real corporate position: AI everywhere in the cost base, nowhere near the crown jewel. If you are tracking the other half of that story, our GTA VI presale and release countdown breakdown covers where the game itself currently stands.
Activision: the disclosure it had to make
Activision's admission was not a press release. It was a checkbox.
Valve's AI disclosure policy, live since January 2024, requires developers to declare pre-generated or live-generated AI content on the store page. The Call of Duty page duly picked up the line "Our team uses generative AI tools to help develop some in game assets," which players found and screenshotted after months of arguing about Black Ops 6 calling cards and loading screens. Black Ops 7 shipped with a similar disclosure, with Activision saying the tools help empower its teams.
That is a materially different kind of admission from EA's. Nobody chose to announce it. A platform rule made staying quiet impossible, which is worth remembering every time a publisher describes its AI use as transparent.
Capcom: the split-the-difference position
Capcom has been running gen AI experiments in the open for a while. Technical director Kazuki Abe described a prototype to Google Cloud Japan that used Gemini Pro, Gemini Flash and Imagen to mass-produce and then score the "hundreds of thousands" of small environment ideas a modern game world needs, the fake TV brands and signage and packaging that fill a street without anyone ever looking directly at them.
The formal policy arrived later. In a Q&A summary from an online investor briefing held on February 16, 2026 and published on March 23, Capcom drew its line at output: the company "will not implement assets generated by AI into our game content," while planning to "actively utilize this technology" to improve development efficiency and productivity, with tests running across graphics, sound and programming departments.
It is the most legible position of the eight studios here. Use it upstream, keep it out of the shipped file.
Ubisoft and Microsoft: the R&D lane
Ubisoft has been iterating publicly since Ghostwriter in 2023, the tool that drafts throwaway crowd barks. Its Neo NPC prototype followed at GDC 2024 with roughly 25 people on it. By November 2025 that had become Teammates, a first-person shooter demo built by around 80 people using Google Gemini plus in-house middleware on the Snowdrop engine, with squadmates that respond to live voice commands. Ubisoft has been consistent that Teammates is an R&D experiment handed to internal creative directors, not a product with a store page.
Microsoft's entry is a research model. Muse, announced on Xbox Wire on February 19, 2025 and published in Nature, is a 1.6 billion parameter World and Human Action Model trained with Ninja Theory on more than a billion images and controller actions from Bleeding Edge. It generates gameplay sequences rather than assets, and Microsoft has pitched it at prototyping and at getting old games running on hardware they were never built for.
What the developers actually doing the work say
Here is the twist that makes the executive quotes worth reading twice.
GDC's 2026 State of the Game Industry survey, covering more than 2,300 professionals, found 36% personally using generative AI tools. Split by employer, that is 30% at game studios against 58% at publishing, marketing and PR outfits. Sentiment went the other way entirely: 52% now say gen AI is having a negative impact on the industry, up from 30% and 18% in the two prior years, while positive responses fell from 13% to 7%. Visual and technical artists were the most negative group at 64%, with design and narrative at 63% and programming at 59%.
The Game Developer Collective survey with Omdia, reported in March 2026, found usage actually falling, from 36% to 29% year on year, after two straight years of increases.
Meanwhile Steam disclosures keep climbing. Sulka Haro's analysis of roughly 53,600 releases from mid-2023 to mid-2026 puts AI-flagged games at 10.9% of 2024 releases, 19.9% in 2025 and 30.8% so far in 2026. Worth remembering that the flag is binary, so one upscaled texture and a fully generated art set look identical in the data.
The read
Top-down adoption is accelerating while bottom-up adoption stalls. Executives are describing AI as a cost lever to investors, and the people actually holding the tools are using them less than they did a year ago, and liking them considerably less than that.
Notice how narrow the admissions are, too. Square Enix names QA. Capcom names productivity and then explicitly excludes shipped content. Take-Two names hundreds of pilots and then fences off Rockstar. Only Krafton has restructured the company around it, and the first visible consequence was 271 people leaving. Nobody has stood up and said generative AI is making their next game better, because that is the claim they would have to defend in reviews.
That gap is where the next two years of games industry news is going to come from. It is also the difference between the studios above and something like World of ClaudeCraft, where the AI was the entire point rather than a line item in a cost-reduction deck. And if you want the sober version of what any of this does to a shipped game's odds, our 2026 genre study remains a better predictor than any of these press releases.
Keep the receipts. The next earnings call is never far away.
Written by
MML Tech →