Generative AI is no longer a fringe experiment inside Japanese game development. According to new CESA survey data, 85.8% of surveyed game developers say they use generative AI in their work.
That number will grab headlines. The more useful part of the report is what comes immediately after it.
CESA says the most common management practice around generative AI is still human checking, correction and supervision. Member companies also reported restricting which tools can be used and avoiding the direct use of raw AI output without review.
For solo developers, indie teams and studios trying to build an AI-assisted pipeline without turning production into chaos, that is the part worth stealing.
The Number: 85.8%
The Computer Entertainment Supplier’s Association, better known as CESA, published the preview edition of its 2026 Game Industry Report on September 17. For the first time, the organization included its own surveys focused specifically on generative AI usage in game development.
CESA says 85.8% of game developers surveyed are using generative AI in their work. The developer survey included 1,349 valid responses and primarily targeted commercial game developers working across roles such as engineering, art, technical art, design, production, sound, QA and management, while also including some educators and students.
A separate CESA member-company survey asked companies what they expected from generative AI. The most commonly cited goal was operational efficiency and improved productivity.
That is important because it cuts through one of the loudest misconceptions around AI in games. The practical use case is not simply press button, get game. Studios are looking at the boring, expensive, repetitive middle of production and asking what can be accelerated without giving up control.
The More Important Number Is Actually Zero
Zero should be the amount of unreviewed AI output that quietly ships because somebody assumed the model knew what it was doing.
CESA’s survey found that human verification, correction and supervision was the most common management practice reported around generative AI. Companies also described restricting approved tools and avoiding the use of raw output as-is.
That sounds less exciting than a giant automation promise, but it is exactly how a serious production pipeline should think about AI.
An AI model can be excellent at generating options, drafting code, producing reference material, summarizing documentation, finding repetitive patterns and accelerating iteration. It can also confidently give you a broken API call, a licensing headache, a bad design assumption or a piece of content that does not match the rest of your project.
The answer is not to ban every useful tool. The answer is to build gates.
What Small Teams Can Learn From the Survey
1. Decide which jobs AI is allowed to touch
Begin with a list of tasks where speed creates obvious value and mistakes are easy to catch. Good candidates include documentation search, boilerplate code, test generation, localization drafts, asset tagging, build-log analysis, marketing variations, first-pass concept exploration and repetitive editor tooling.
High-risk areas should get stricter review. Final legal copy, licensing decisions, security-sensitive code, irreversible data migrations and anything that ships directly to players without a human pass deserve a different standard.
2. Restrict the toolchain
CESA says tool designation and restriction is already a common practice among surveyed companies. That prevents a team from having five people feeding project data into five different services with five different privacy policies.
For an indie studio, a simple approved-tool list is enough. Record what the tool can access, whether project data is retained, what licenses apply to generated output and whether confidential source material can be uploaded.
3. Keep humans at the acceptance gate
The highest-leverage AI workflow is often not AI finishes the task. It is AI gets the task to a reviewable state faster.
A coding agent can prepare a change, but tests and review decide whether it enters the branch. An image model can produce concepts, but the art direction decides what becomes canon. A writing model can draft copy, but the editor decides what the audience actually sees.
AI should make the queue move faster. It should not silently become the person who signs off on its own work.
4. Measure saved production time, not prompt count
If a tool generates ten things in five minutes but you spend two hours repairing them, you did not save time. You created a very enthusiastic intern who set the office on fire.
Measure the full workflow: time to produce, time to review, time to correct, defect rate and whether the result can be reused. That tells you whether AI is actually improving production rather than simply making activity look faster.
This Does Not Settle the Copyright Debate
CESA itself acknowledges that generative AI raises rights issues, including copyright concerns. The organization’s statement says the industry needs to balance the possibilities of the technology with rights protection and human-centered game creation.
An 85.8% usage figure does not mean every developer uses the same models, uses them for the same tasks or agrees on where the ethical and legal lines should be drawn. It also does not tell us that 85.8% of game content is AI-generated. Those are completely different claims.
The survey tells us that generative AI has entered the workflow. It does not tell us that humans have left it.
Why This Matters for Indie Developers
Large studios can absorb inefficiency with headcount. Small teams cannot. If a tool turns a four-hour repetitive task into a 45-minute task plus 20 minutes of review, that is meaningful production capacity.
But small teams are also more vulnerable to bad automation. One corrupted project, one licensing mistake or one badly generated system that nobody fully understands can erase the time saved everywhere else.
That is why the governance part of CESA’s findings is arguably more valuable than the adoption number. The emerging pattern is not replace the team. It is use the tool, constrain the tool, verify the output.
What We Still Do Not Know
The preview report is just that: a preview. CESA plans to release the full 2026 Game Industry Report in early December.
The current release includes two different surveys with different respondent groups. The developer survey had 1,349 valid responses, while the member-company survey had 48 valid responses to the relevant question out of 220 member companies targeted. Percentages from the two surveys should not be blended together as if they describe the same population.
ZerotoBeast Take: the interesting future is not a studio where nobody does any work. It is a studio where people spend less time on the parts a machine can accelerate and more time deciding what deserves to exist. The teams that win will probably not be the ones with the most AI. They will be the ones with the best acceptance gates.
