According to Automaton, a preview of the 2026 CESA Video Game Industry Report found that 85.8% of surveyed Japanese game developers use generative AI in their work. Of those respondents, 63% said they use it daily, while another 22.8% use it occasionally. That is a significant increase from the 51% recorded in the previous year’s report. The headline is obviously important, but it is also incomplete. The real production question isn’t whether developers are using generative AI anymore. It is what they’re using it for, which decisions it influences, how generated material is verified, what knowledge the team retains, and who remains accountable when the output enters the game. Adoption can happen quickly. Responsibility takes considerably more design.
What The CESA Survey Actually Says
The survey targeted game developers and companies belonging to Japan’s Computer Entertainment Supplier’s Association. CESA’s membership includes major developers and publishers such as Capcom, Sega, Level-5, and Konami, although the published coverage doesn’t provide enough detail to treat the figure as a perfect measurement of every developer working in Japan. The preview offers several useful signals:
| Survey Finding | What It Suggests |
|---|---|
| 85.8% report using generative AI | AI use is widespread among the surveyed developers |
| 63% use it daily | For many respondents, AI is becoming part of ordinary production |
| 22.8% use it occasionally | Other teams may be experimenting or limiting use to particular tasks |
| The 2025 figure was 51% | Reported adoption has increased rapidly |
| Efficiency and productivity are expected benefits | Studios are primarily evaluating AI as a production tool |
| Shorter development cycles are a priority | Schedule pressure is encouraging adoption |
| Lower development and operating costs are expected | Commercial incentives are shaping usage |
| Human verification is the most common safeguard | Teams recognise that generated output can’t simply be trusted |
The survey also reports that companies are restricting which tools can be used and avoiding the direct use of generated output. These are sensible controls. They demonstrate that adoption doesn’t necessarily mean developers are placing an AI tool beside the build server and politely asking it to finish the game before lunch. However, the preview doesn’t disclose the most common 2026 use cases. The previous report identified visual asset generation, story and text generation, and programming assistance among the leading applications. Without current task-level data, 85.8% tells us how widespread usage is, but not how deeply it affects final games. The full 2026 CESA industry report is expected later in the year. Until then, the percentage should be treated as a strong adoption signal rather than a complete map of Japanese development practice.
Insider Tip: A percentage describing tool usage doesn’t reveal production influence. A developer using AI to summarise meeting notes and a studio generating final character assets may both answer “yes,” but the creative, legal, and operational risks are completely different.
Adoption Isn’t The Same As Approval
A developer can use a tool without believing it should determine creative direction. They may use it because the company provided it, because it is embedded inside existing software, because it accelerates a repetitive task, or because production pressure makes experimentation difficult to avoid. “Uses generative AI” can describe very different relationships:
| Type Of Use | Example | Level Of Production Influence |
|---|---|---|
| Administrative assistance | Summarising meetings or organising notes | Low |
| Research assistance | Comparing documentation or locating information | Low to moderate |
| Internal prototyping | Temporary text, images, or code used to test an idea | Moderate |
| Production assistance | Code suggestions, localisation drafts, or asset variations | Moderate to high |
| Final content generation | Art, dialogue, music, animation, or marketing assets | High |
| Player-facing generation | Live dialogue, characters, quests, or reactive content | Very high |
Collapsing these uses into one number creates confusion. Supporters can use the percentage to claim that adoption is inevitable. Critics can use it to suggest the industry has already replaced enormous amounts of creative work. Neither conclusion is justified without knowing which tasks were affected and how the resulting material was used. The important distinction isn’t simply use versus refusal. It is assistance versus substitution, experimentation versus dependency, and temporary material versus content that reaches players.
Insider Tip: The important distinction isn’t simply use versus refusal. It is assistance versus substitution, experimentation versus dependency, and temporary material versus content that reaches players.
Human Verification Isn’t A Magic Shield
The most common safeguard identified by the CESA survey is human verification, correction, and supervision. That sounds reassuring because it places a person between the generated output and the final decision. The problem is that “a human checked it” can mean almost anything.
A senior artist may examine an image for composition, anatomy, visual consistency, provenance, and compatibility with the wider art direction. A programmer may review suggested code for security, performance, architecture, licensing, and maintainability. A producer under deadline pressure may look at an output for twelve seconds, confirm that it hasn’t visibly caught fire, and move it into the approved folder. All three processes contain a human. They don’t provide the same protection.
| Verification Claim | The Reviewer Must Understand |
|---|---|
| “The code works” | Architecture, dependencies, security, performance, and long-term maintenance |
| “The art looks correct” | Art direction, anatomy, style consistency, provenance, and production requirements |
| “The writing reads well” | Character voice, context, theme, pacing, continuity, and cultural sensitivity |
| “The translation is accurate” | Language, intent, tone, terminology, and regional context |
| “The output is legally safe” | Training-data risks, licences, ownership, contracts, and jurisdiction |
| “The result is unbiased” | Data limitations, excluded perspectives, representation, and likely harm |
Verification requires time, expertise, authority, and access to enough information to judge the result. If the organisation removes specialists because a tool can generate material, it may also remove the people capable of evaluating that material. A human in the loop isn’t useful when the human has become a ceremonial approval button.
Insider Tip: Don’t measure verification by whether somebody viewed the output. Measure it by whether the reviewer understood the risks, had authority to reject it, and could explain why the result belonged in the project.
AI Governance Is Production Design
When generative AI becomes part of daily development, policy can’t remain a paragraph written by the legal department and stored somewhere nobody visits until something goes wrong. The policy becomes part of the production system. It affects which tools developers can access, what data can be entered, how contractors work, which assets require provenance records, how generated code is documented, who reviews output, and what the studio tells players. It also needs to survive schedule pressure, staff turnover, outsourcing, software updates, and the sudden arrival of a new tool that promises to solve every existing problem after watching a four-minute demonstration. A useful governance structure needs several connected layers:
| Governance Layer | Question It Must Answer |
|---|---|
| Purpose | What specific production problem is the tool solving? |
| Scope | Which teams, tasks, assets, and development stages can use it? |
| Data | What information can be entered into the system? |
| Rights | Who owns the input, output, and underlying training material? |
| Verification | Who reviews the result and against which criteria? |
| Documentation | Can the studio trace where generated material entered production? |
| Capability | Does the team retain the skill needed to understand and revise the work? |
| Disclosure | What should developers, contractors, platforms, and players be told? |
| Response | What happens when output creates a legal, ethical, or production problem? |
This isn’t bureaucracy added after the creative work. These decisions shape the creative work. If a concept artist receives generated material with unclear origins, that affects authorship and art direction. If a junior programmer accepts code they can’t explain, that affects maintenance and training. If a writer is asked to repair generated dialogue rather than construct the character’s voice, that affects how narrative knowledge develops across the team. The tool changes the production environment, and the production environment changes the game.
Insider Tip: The tool changes the production environment, and the production environment changes the game.
Faster Output Can Create Capability Debt
Generative AI is usually justified through efficiency. The tool can produce an answer, draft, variation, function, or placeholder more quickly than the existing process. Speed is useful, but production doesn’t only generate assets and code. It generates knowledge. An artist learns the visual language of the game by solving visual problems. A programmer understands the architecture by tracing relationships and repairing failures. A writer discovers a character’s voice through drafts that reveal which language feels wrong. A designer learns the simulation by testing interactions that initially fail. If generation repeatedly removes those moments, the studio may create more output while developing less understanding. That creates capability debt: the task is completed today, but the team becomes less capable of judging, adapting, or repairing similar work tomorrow.
| Immediate Efficiency | Possible Future Cost |
|---|---|
| Generated code completes a task quickly | Fewer developers understand the implementation |
| Rapid visual concepts increase options | The art direction becomes harder to defend or reproduce |
| Generated dialogue fills content gaps | Character voices lose ownership and consistency |
| Automated documentation saves time | Errors are repeated with professional-looking confidence |
| Generated prototypes accelerate testing | Temporary solutions quietly become production dependencies |
| AI handles junior-level tasks | Fewer opportunities exist for juniors to develop senior judgement |
This doesn’t mean every automated task is educationally sacred. Developers don’t need to manually perform repetitive work forever just because somebody once suffered through it. The question is whether the removed work was waste or practice. Automating file organisation is different from automating the moment when an artist discovers the visual language of the project. Generating basic test data is different from outsourcing the decisions that define how the game’s systems relate.
Insider Tip: Before automating a task, ask whether the task produces only output or whether it also produces understanding. If it creates understanding, decide how the team will retain that learning.
Internal Use Can Still Reach The Player
Studios sometimes separate internal AI use from final content. A generated image may be used only as concept material. Generated text may be used for brainstorming. Suggested code may be rewritten before release. That boundary matters, but internal tools can still influence the finished experience. Concept material shapes composition, costume, mood, and expectations. Brainstorming suggestions affect which ideas receive attention. Generated prototypes influence which mechanics survive. Code assistance can alter architecture even when the final lines have been revised. A tool doesn’t need to place an untouched asset inside the game to affect what the game becomes. This is why “none of the generated output appears in the final game” isn’t a complete explanation. The useful question is whether the tool influenced important creative or technical decisions, and whether the team understood that influence. Players don’t need a record of every internal experiment. Studios should still know where major decisions came from and which people remain responsible for them.
Insider Tip: Players don’t need a record of every internal experiment. Studios should still know where major decisions came from and which people remain responsible for them.
Disclosure Needs More Than An AI Badge
As generative AI becomes widespread, vague disclosure becomes less useful. “Made with AI” can describe a meeting summary, a generated texture, suggested code, an entire marketing campaign, or a live system constructing dialogue for players. “AI-free” can also hide uncertainty if contractors, software providers, or external agencies operate under different rules. Meaningful disclosure should identify boundaries:
- Which stages of development used generative AI?
- Which types of content or code were affected?
- Did generated output appear directly in the final game?
- Were private, licensed, or publicly available models used?
- How was the material reviewed?
- Were performers, artists, writers, and contractors informed?
- Who remains accountable for the final result?
Transparency isn’t about forcing studios to publish every awkward experiment. It is about giving developers and players enough information to understand the production promise being made. The higher adoption climbs, the less useful it becomes to ask whether a studio uses AI. The useful question is how.
Insider Tip: Good disclosure describes scope and responsibility. Bad disclosure uses “AI” as a magical noun and leaves everyone else to guess what happened.
This Isn’t A Story About Japan Choosing AI
The CESA figure shouldn’t be turned into a simplistic cultural conclusion. It doesn’t prove that every Japanese developer supports generative AI, that Japanese players approve of it, or that every surveyed company uses it for final creative work. It reflects responses collected from developers and CESA member companies within a particular industry context. The organisations involved differ enormously in scale, discipline, production model, and likely use case. The survey is still valuable because it shows how rapidly reported adoption has increased. It gives the wider industry a warning about the speed at which experimental tools can become normal production infrastructure. Normalisation is precisely why governance matters. When a tool is rare, every use receives attention. When it becomes ordinary, assumptions settle in, exceptions accumulate, documentation weakens, and “we’ve always done it this way” begins forming at remarkable speed for something the studio adopted eight months ago.
Insider Tip: Normalisation is precisely why governance matters. When a tool is rare, every use receives attention.
The Design Lesson For Studios
CESA executive director Tsutomu Masuda described the result as evidence that generative AI is being widely and steadily adopted, while also acknowledging infringement concerns and the need to protect intellectual property alongside human-centred development. That balance is the real challenge. Studios don’t need identical policies. A small independent team using an assistant for administrative work faces different risks from a publisher generating assets across several external studios. A private model trained on owned material creates different questions from a public service with unclear data practices. Live player-facing generation requires different safeguards from an internal prototype. Every studio should still be able to answer the same fundamental questions:
- What problem is the tool solving?
- What part of the process must remain attributable to people?
- What data, rights, and confidential material are exposed?
- Who understands and verifies the output?
- How is generated material documented?
- What happens to junior development and specialist knowledge?
- What will contractors and external partners be required to disclose?
- What will players be told?
- Who is accountable when the result causes harm?
Insider Tip: If those questions can’t be answered, the studio hasn’t adopted a production strategy. It has adopted a dependency.
Final Thoughts
CESA’s reported increase from 51% to 85.8% is significant because it suggests generative AI has moved rapidly into everyday development among the surveyed developers and companies. With 63% reporting daily use, the technology can’t be treated only as a future possibility or a debate happening somewhere outside ordinary production. The survey also suggests that developers understand the need for safeguards. Human verification, restricted tools, and avoiding direct output are useful starting points. They aren’t complete governance systems. Accountability needs scope, documentation, expertise, authority, provenance, disclosure, and a clear owner. Studios must understand not only what the tool generated, but what knowledge it replaced, which decisions it influenced, and whether the team remains capable of maintaining the result. The argument is no longer simply between using AI and refusing it. The harder question is whether studios can adopt it without weakening authorship, capability, creative intent, intellectual-property protection, and player trust. Adoption can make a tool normal. It can’t make the consequences disappear.
That’s it for this one! Subscribe to The Design Lab for more breakdowns and analysis. Please like, share, and comment if you found this article useful AND…
Need help applying these concepts to your game? Most games don’t fail because of bad ideas. They fail because the systems don’t work together to evoke immersion.
> Work with me:
https://thedesignlab.blog/services/
> Or go deeper into this framework:
Register for my upcoming book – Immersive Design Framework – below and get a second book, now, free!
Grab Myles’ FREE ebook now and find 15 indispensable design patterns that will equip you to craft exceptional Web3 gaming experiences. We’ll also notify you when his new book on immersive design is out!