Two years ago, AI in game development was mostly a demo reel. In 2026 it is in the pipeline of almost every studio that ships, and the numbers behind that shift are more dramatic than most people realise. This guide from Maxio Fun Studio covers where generative AI is genuinely being used, what AI 3D modeling can and cannot produce, and why half the industry is unhappy about all of it.

The Numbers: How Fast This Actually Happened

Japan's CESA industry survey found that 85.8% of Japanese game developers now use generative AI, up from 51% a year earlier. Among those who use it, 63% use it every single day. That is close to a complete sweep of an industry in twelve months.

The picture in the West is different and worth understanding. The GDC 2026 State of the Industry report found 52% of companies use generative AI, but only 36% of individual developers personally use the tools. That 16 point gap is the whole story in one statistic. Companies are adopting faster than the people doing the work.

The sentiment number is the one nobody saw coming. 52% of developers now say generative AI is harmful to the industry, up from 30% the year before. Opposition nearly doubled in the same year adoption surged. Both things are true at once.

Where Generative AI Is Actually Being Used

Among developers who use these tools, the split is clear and it is not what the headlines suggest.

Research and brainstorming leads at 81%. This is the quiet, unglamorous use. Naming things, outlining systems, summarising research, drafting store copy, working through a design problem out loud.

Code assistance sits at 47%. Autocomplete, boilerplate, writing tests, explaining unfamiliar code, converting a shader from one engine to another.

Content creation is substantially lower. Despite being what everyone argues about, actually shipping AI generated art and text into a game is the least common use.

The Japanese survey breaks the content side down further: visual assets and image generation first, then story and text generation, then programming support. Notably, 32% of companies reported using AI in their in-house engine development, which is a far more technical use than the art debate usually covers.

AI 3D Modeling: What Text-to-3D Really Produces

This is the area that changed most in the last year, and it is the one mobile studios care about.

The leading AI 3D model generators are Meshy, Tripo and Rodin. The workflow is genuinely as simple as it sounds. You type something like "weathered wooden treasure chest, stylized, low poly" and roughly forty seconds later you have a textured 3D model you can download.

What you get back is more complete than people expect. Geometry ranging from a few thousand to several hundred thousand polygons. A full PBR texture set with base colour, normal, roughness and metallic maps. Auto-rigging for humanoid and some animal shapes. Export in .glb, .gltf, .fbx and .obj, so it drops straight into Unity or Unreal.

For a small studio, that is a real change. Background props, crates, barrels, street furniture, roadside objects, the hundreds of small things that fill a 3D world and that nobody notices individually. Those used to eat weeks of artist time.

What AI 3D Models Still Cannot Do

Here is the part the tool marketing leaves out, and it matters more than the generation speed.

Scale is wrong. Drop a generated model into a scene and it might come in the size of a building or the size of a grain of rice. Every asset needs manual scaling against your actual character.

There is no collision geometry. A visual mesh does nothing physically. Your player walks straight through that beautiful crate. Collision has to be authored separately, and for a vehicle or a physics game that is not a small job.

Materials rarely match. Generated textures often read flat, too shiny or too dark once they are under your game's actual lighting. They need adjusting to sit with everything around them.

Auto-rigging breaks on anything unusual. It works on clean humanoid meshes. A four legged creature, a multi armed boss, a mech, a truck with moving parts, all of those still mean rigging by hand.

Topology is not production ready. Polygon counts are often far too high for mobile, and the edge flow is not built for deformation. Retopology is still a human job.

The honest summary is this. A generated chest is forty seconds of work. A generated chest that is actually in your game, at the right size, that the player cannot walk through, and that matches the lighting around it, is the real finish line. AI moved the starting line, not the finish line.

AI in 2D Art, Concept and UI

Two dimensional work is where AI has been useful the longest. Concept art and mood boards are the clearest win, because those are meant to be thrown away. Generating forty directions in an afternoon and picking one beats sketching three.

Icons, background textures and placeholder UI are the other common use. These let a small team keep building while the real art is still being made, instead of staring at grey boxes.

Store screenshots and advertising creative are a quieter but commercially important one. Large publishers produce fifty to a hundred ad variants a month, and generative tools are a large part of how that volume is now possible.

AI in Code, Testing and Localisation

Code assistance is the use developers themselves are most comfortable with, and it is the least controversial. Writing boilerplate, generating unit tests, explaining a legacy system, translating a shader between engines.

Testing is quietly becoming a bigger one. Automated agents that play through a level repeatedly can find a geometry gap or a softlock far faster than a human tester playing it twice. For a mobile game that has to run on a thousand device combinations, that is genuinely valuable.

Localisation has shifted from a luxury to a default. Translating a game's interface into fifteen languages used to be a budget decision. Now it is a first draft plus a human review, which puts it within reach of a small studio.

AI NPCs and Procedural Content: Less Than the Hype

The demo everyone shares is an NPC you can have a real conversation with. The reality in shipped games is much smaller, for reasons that are practical rather than philosophical.

Running a language model for every NPC costs money per conversation, adds latency, and gives you a character who can say anything, including things that embarrass your studio. On mobile specifically, it also means the game stops working offline, which for a lot of players is the whole point.

Procedural generation, meanwhile, is decades old and was never really an AI story. What has changed is that the inputs to procedural systems can now be generated faster, not that the systems themselves became intelligent.

Why Half the Industry Is Against It

It would be dishonest to write this article without covering the other side properly.

The jump from 30% to 52% of developers calling generative AI harmful happened in a single year, and it is not mainly about quality. The concerns that come up repeatedly are training data and copyright, work being done by tools trained on art that nobody licensed, and the gap between management adopting tools and the workers expected to use them.

CESA's own executive director said plainly that concerns about infringement arising from generative AI must be addressed as the industry expands its use. The most common safeguard reported by Japanese studios was human verification, correction and supervision of every output, with some restricting which tools can be used at all and avoiding unreviewed AI output entirely.

That is a reasonable position and it is where most careful studios have landed. Use the tools, verify everything, ship nothing you have not checked.

How a Small Studio Should Actually Use AI

At Maxio Fun Studio we are a small independent team self-publishing on Google Play, so our view is practical rather than ideological.

Where it earns its place is in the boring middle of production. Placeholder assets so a level can be tested before the real art exists. Background props that fill a scene. First draft localisation. Boilerplate code. Store listing variants. Research and reading.

Where we do not lean on it is the part players actually feel. Core mechanics, level design, the difficulty curve, the moment to moment handling of a truck on a hill road. Those are design decisions, and games that outsource them feel hollow in a way players notice even when they cannot name it.

Our games are also built to run fully offline, which rules out anything that needs a live model at runtime. That constraint has been useful. It keeps the AI in production where it helps and out of the game where it would get in the way.

What Is Coming Next

Three things look likely rather than speculative.

Retopology and optimisation get automated. The gap between a generated mesh and a game ready mesh is a well defined technical problem, which makes it the next thing to be solved.

Engine integration gets deeper. Generating an asset inside the editor, already scaled and with collision, rather than downloading a file and fixing it, removes most of the friction that exists today.

Disclosure becomes standard. Stores and platforms are moving toward requiring developers to declare generative AI content. Expect that to be a normal part of a store listing rather than a controversy.

Frequently Asked Questions

How many game developers use AI in 2026? It depends where you ask. Japan's CESA survey found 85.8% of developers use generative AI, up from 51% a year earlier. The GDC survey found 52% of companies and 36% of individual developers in a largely Western sample.

Can AI generate 3D models for games? Yes. Tools like Meshy, Tripo and Rodin produce a textured model with PBR maps in under a minute. But the output needs scaling, collision geometry, material adjustment and usually retopology before it is usable in a real game.

What is AI used for most in game development? Research and brainstorming at 81% of users, then code assistance at 47%. Actually shipping AI generated content into games is far less common than the debate suggests.

Will AI replace game developers? Not on current evidence. It compresses the repetitive parts of production, which is real value, but the parts that decide whether a game is good, the design and the feel, are not what these tools do well.

Do developers actually want to use AI? Many do not. 52% of developers say generative AI is harmful to the industry, up from 30% the previous year, even as company adoption rose. Copyright and training data are the most cited concerns.

Do AI NPCs work in mobile games? Rarely, so far. A live language model costs money per conversation, adds latency, and breaks offline play. For a game meant to work with no signal, that is a dealbreaker.

The Short Version

Adoption of generative AI in gaming went from a minority to a majority in roughly a year, and developer opposition grew just as fast. The tools are genuinely good at the repetitive middle of production and still weak at everything that decides whether a game is worth playing.

At Maxio Fun Studio we use AI where it saves time and keep it away from design. Our games stay free, offline and sign-up free on Google Play.