CESA handed out a preview edition of its Video Game Industry Report 2026 at Tokyo Game Show on 17 September, and one line from it went round the world by the weekend: 85.8% of Japanese game developers use generative AI at work. Last year’s figure was 51%. That is a thirty-five point jump in twelve months, which is the kind of number that gets a headline whether or not anyone reads the second paragraph.
The second paragraph is where the survey design lives, so let me start there.
Two surveys, two sample sizes
The 85.8% comes from CESA’s annual developer questionnaire — the one it runs under the title Game Developers’ Employment and Career Formation — which was open online from May to August 2026 and closed with 1,349 valid responses. Inside that total, 63.0% said they use generative AI daily and 22.8% said they use it occasionally. Add those and you get the headline.
The breakdown of what they use it for does not come from those 1,349 people. It comes from a separate strand answered by 50 CESA member companies, and the shape of it is worth writing down in full:
- internal business operations — 32%
- supporting programming — 25%
- proposals and idea generation — 23%
- translation and localisation — 22%
- producing visuals — 16%
- QA and testing — 15%
- marketing and publicity — 14%
- scenario and text — 10%
- music and audio — 9%
Coverage is not unanimous on this point. Some outlets ran the task table as part of the preview; at least one reported that the preview did not disclose the most common uses at all. My reading is that both are describing different strands of the same document — the developer survey gave a percentage and no task list, the member-company section gave the task list. The full report lands in early December. Right now the methodology is an inference from press summaries.
Fifty companies is a small base. Percentages off fifty move a full two points every time one respondent changes their mind, and nobody should treat 16% and 15% as meaningfully different numbers at that size. What survives the error bars is the ordering, and the ordering is clear enough: the top four uses are paperwork, code assistance, pitching and translation.
What “uses generative AI” measures
The headline cannot carry this gap. A survey question that asks whether you use generative AI at work returns a yes from the programmer who had Copilot finish a loop, from the producer who had a model tidy a schedule mail into Japanese business register, and from the studio generating final character art. Only the third is what people picture when they read 85.8%.
I have watched the same conflation happen at smaller scale in every studio conversation about this. Somebody runs a model over the meeting notes and the question in the room becomes whether the art department is next. The two things share a tool and share nothing else.
The number that a working artist would want is not in this survey. How many of those fifty companies shipped a generative asset into a build? How many used it in blockout and threw the output away? How many ran it through an internal review board first? CESA’s managing director has been public about balancing experimentation against copyright and creator protection, and several of the named respondents — Capcom, Konami, Sega, Level-5 — have internal policies requiring human review and forbidding raw output going out unchanged. That policy layer is invisible in a yes/no question, and it is the entire difference between a tool and a replacement.
The benefits list is a schedule argument
The three benefits developers most often expected, in order: improved operational efficiency and productivity, shorter development cycles, reduced development and operational costs.
All three are arguments about time rather than quality, which matches what Japanese studios have been saying out loud. Akihiro Hino spent part of a Level-5 broadcast apologising for AI use and then, in the same show, restated the ambition — turn artwork straight into polygons and compress five-year projects into two. That is the same sentence as “shorter development cycles”, delivered by a studio head rather than a survey form.
Whether the compression is real is a different question. Blockout and iteration speed up. Review, integration, art direction sign-off and the long tail of fixing what the tool got subtly wrong do not, and in my experience the last of those eats a good share of what the first one saved. A 22% localisation figure I believe without argument. A claim that the same tooling halves a five-year project I would want to see land before I priced a schedule against it.
Where the constraint comes from
The counterweight comes from contracts, and rarely from inside the art team. Saber’s Space Marine 3 is shipping with no generative AI anywhere in it because the Games Workshop licence forbids it, which settles the question before a single artist has an opinion about it.
That is the mechanism to watch over the next year, more than the adoption percentage. Adoption is going to keep climbing — a survey that counts “did you use it once” always climbs. What decides whether generative output reaches a finished game is licensing, platform policy, and whichever publisher first has to explain in court where a texture came from.
December will tell us how much of the 2026 report holds up. Until then the honest summary of 85.8% is that most Japanese developers have opened the tool. Almost none of what matters to an artist is in that sentence.