Can AI Build a Video Game? A Feature-by-Feature Reality Check

The honest answer: it depends what you mean by "build"

Can AI build a video game? The question gets asked as if a game were one thing, and answered as if there were one answer. There isn't. A game is not a monolith you either can or can't automate — it is an assembly of dozens of separate crafts, and AI is wildly good at some of them and almost useless at others. Ask whether AI can build a game and you get a shrug. Ask whether AI can build the specific parts your specific game is made of, and you get a real answer.

So we took the question apart. In our Games-as-Recipes study we broke a finished game down into roughly 45 reusable ingredients — every distinct piece of work that goes into shipping something playable — and grouped them into five kitchens: art, audio, narrative, code, and the polish-and-feel work that happens last. Then we rated each ingredient on a simple 1-to-5 scale for how much today's AI tooling can actually do. A 1 means AI can carry it close to end-to-end; a 5 means expert, real-time craft that AI barely moves.

Two things to be clear about before any number lands. This difficulty scale is our editorial assessment of what current tooling can do, not a benchmark or a measurement — it is a judgement call, made deliberately and consistently, but a judgement. And it measures automatability of the parts, not whether the result is any good. "Buildable" is emphatically not "fun." AI can generate a perfectly competent something-shaped game and have it be lifeless. Hold both ideas at once and the rest of this piece makes sense.

Five kitchens, and a clean split down the middle

When you average our difficulty ratings by kitchen, an order falls out immediately. Audio is the most automatable kitchen at 2.2, followed by art at 2.57, narrative at 2.78, code at 3.0, and polish-and-feel dead last at 3.75. That is not random. The kitchens that score low are the ones whose work is generation — producing an asset from a prompt or a spec. The kitchens that score high are the ones whose work is judgement under real-time constraints, where the thing being made only reveals whether it's right once a human is holding the controller.

Inside each kitchen the same pattern repeats at the ingredient level. The generative tasks sit at difficulty 1 and 2; the craft tasks sit at 4 and 5. Voice acting is a 1 — in our assessment it is the single most automated craft in all of game development, because text-to-speech has quietly become indistinguishable enough for most uses. 2D backgrounds, concept art, and flavor text are 1s. Music, SFX, sprite art, branching dialogue, inventory and crafting systems, save and menu plumbing — all 2s. This is the half of game-making AI genuinely near-solves today.

Then the floor drops out. Game feel and juice — the moment-to-moment tactile satisfaction of a jump or a hit — is a 5. Real-time authoritative netcode is a 5. Level, puzzle, and encounter design are 4s. The whole "is it actually fun?" balancing loop is a 4. These aren't hard because they involve a lot of typing; they're hard because they require taste, feedback, and iteration against a human's nervous system. No prompt closes that gap.

The four things AI can't do (yet)

If you remember nothing else from our buildability model, remember the four ceilings — the ingredients sitting at 4 and 5 that resist automation hardest. The first is game feel and juice. You can generate a character and animate a jump, but whether that jump feels good — the anticipation frames, the screen shake, the landing thud, the half-second of input forgiveness — is hand-tuned by feel, and AI has no nervous system to tune against.

The second is real-time authoritative netcode: keeping many players in sync, fairly, under latency, while resisting cheating. It is the only ingredient we rate a flat 5 in the code kitchen, and it's why fast multiplayer remains a specialist's game. The third is design — level, puzzle, and encounter design — the work of shaping space and challenge so that it teaches, surprises, and paces the player. AI can fill a level with assets; it cannot yet decide where the interesting bit goes.

The fourth is the balance loop: the slow, taste-driven cycle of playtesting and tuning that turns a working game into a good one. This is the through-line behind all four ceilings — they are the parts that only resolve through iteration against human perception. Generation is a one-shot act AI excels at. Feel, fairness, design, and balance are loops, and loops are where craft lives. That distinction is the whole thesis.

A genre is just a recipe — and the recipe decides buildability

Here is the move that makes all of this practical. If a game is a recipe of ingredients, and each ingredient has a difficulty, then a genre's overall buildability is just the average difficulty of the ingredients its core demands. Genres aren't equally hard to automate because they don't draw from the same kitchens. A visual novel is almost entirely art, audio, and narrative — the low-difficulty kitchens. A fighting game lives or dies on frame-perfect game feel and netcode — the ceilings. Same engine, completely different automatability.

So we scored 30 genres by averaging the difficulty of their core ingredients, and sorted them into four bands. AI-buildable (below 2.4) is where a small team or a solo developer can lean on AI for most of the heavy lifting. AI-assisted (2.4 to 3.0) is where AI does real work but a human owns the core. Systems-craft (3.0 to 3.6) is where the defining work resists automation. Expert-only (above 3.6) is where the genre's whole identity is one of the four ceilings.

The spread is dramatic. Visual novels score 1.80 in our model; battle royales score 4.17. That is not a statement about which is the better game or the bigger opportunity — it's a statement about how much of the build AI can carry. The table below lays out the full ladder. Read it as a map of where, today, you can move fast with machine help, and where you still need a craftsperson.

Where the buildable genres meet real demand

Buildable is not the same as worth building. A genre AI can mostly assemble is only interesting if players actually want it. So we crossed our buildability scores against demand signals from the MythicLemon Games Catalogue — our own compiled dataset of roughly 82,000 released games — looking for genres that are both automatable and alive. To be clear about what those demand figures are: they are banded popularity counts and momentum, a relative-popularity proxy, never unit sales or revenue. We have no sales figures and we never present one.

We set three gates: a buildability score at or below 2.7, at least 1,000 catalogued titles so the audience is real, and at least 40% of titles released since 2023 so the genre is still growing rather than fading. Five genres clear all three. Interactive Fiction (1.83, 3,137 titles, 48% recent). Hidden Object (2.07, 4,172, 45%). Clicker and idle games (2.23, 2,174, 47%). Walking Simulator (2.59, 3,171, 51%). And Card Game (2.62, 2,180, 45%). These are the genres where AI's strengths and player demand line up cleanly right now.

Two honourable mentions sit just outside the gates and deserve a glance. Farming Sim has the hottest momentum of anything we looked at — 59% of its catalogued titles are from 2023 or later — but it slips under the 1,000-title bar, so treat it as a high-energy bet rather than a safe one. And the JRPG scores 2.77: technically a hard role-playing genre, but in practice a content, art, and music genre wearing an RPG costume, which is exactly why our model rates so much of it automatable.

The contrarian pick: build where AI builds, bet where it can't

There is one genre that breaks the pattern in an instructive way, and it is the most important entry in the whole study. Deckbuilding is the fastest-growing genre we track — 64% of its catalogued titles have appeared since 2023, the highest recent share of anything on the board. Demand is surging. And yet it scores 3.18 in our model, parked firmly in the systems-craft band, because its core is balance: the endless tuning of card interactions and economies that is precisely the AI-hardest content work there is.

Put those two facts together and you get the most defensible position in game development right now: highest demand crossed with the hardest-resisting core. Anyone can have AI stamp out the cards, the art, and the flavor text. Almost nobody can get the balance right, and balance is the entire game. That moat doesn't erode when the next image model ships, because the next image model doesn't tune a curve against a thousand playtests. The thing AI can't do is the thing worth owning.

That is the thesis the whole decomposition points at: build where AI builds, bet where it can't. Use generation to collapse the cost of everything in the low kitchens — the backgrounds, the voice, the music, the boilerplate, the menus — so your time and money pile up against the ceilings, where they actually create defensible value. The studios that win the next few years won't be the ones who refuse AI or the ones who hand it everything. They'll be the ones who know exactly which half is which.

How to put the split to work in your own pipeline

Practically, this reshapes how you scope a project. Before you start, run your game's feature list through the five kitchens and mark each ingredient's difficulty using the ledger above. Everything you tag 1 or 2 is a candidate to generate or accelerate: a text-to-image model for backgrounds and concept art, an AI voice for placeholder or even shipped VO, a music generator for the soundtrack, a code assistant for inventory, save systems, menus, and tools. Everything you tag 4 or 5 is where your human hours have to go, untouched by shortcuts.

On the code side specifically, the win is rarely the gameplay itself — it's the scaffolding around it. Save and settings, procedural-generation systems, platform integration, and in-house tooling all sit at difficulty 2 in our model, which means an assistant can stand up the unglamorous infrastructure fast and leave you free to hand-craft the controller, the combat, and the feel. This is the everyday loop our own MythicDevAssist is built for: keeping an AI assistant grounded in what your Unreal project actually contains so the boilerplate gets generated correctly and you stay focused on the parts that need taste.

And reach for the right narrow tools at the right difficulty tier. A 2D-heavy or sprite-driven game lives in the art kitchen, where something like our AI Flipbook Generator turns prompts into usable animated frames instead of hand-drawing every cell. A data-driven or systems game leans on widgets — our Fast Chart Widgets exist for exactly the dashboards and readouts those genres need — and almost anything that talks to a backend or live-ops service needs clean HTTP plumbing, which is the entire job of our EasyHTTP. None of these replace the craft at the ceilings. They clear the low kitchens out of your way so you can spend your one scarce resource, judgement, where AI can't.

The bottom line

So — can AI build a video game? It can build most of the parts of some games, the easy parts of every game, and the hard parts of none. The generation half of game-making, spread across the low kitchens, is genuinely close to solved for everyday use. The craft half — feel, netcode, design, balance — is barely moved, and shows no sign of being moved by more of the same generative tooling.

Which means the real question was never "can AI build a game?" It was "which game, and which parts?" A visual novel scores 1.80 in our buildability model and a battle royale scores 4.17 not because one is better, but because of which kitchens their cores draw from. Pick a genre whose ingredients land in the kitchens AI has mastered and you can move astonishingly fast. Pick one whose core sits on the ceilings and you'll spend your whole budget exactly where AI can't help.

Treat our Games-as-Recipes study as a planning instrument, not a verdict. Decompose before you commit, automate the low kitchens without guilt, and pour your scarce craft into the four things AI can't do. Build where AI builds. Bet where it can't. That's the whole strategy, and the decomposition is just the map that shows you the line between them.

The five kitchens, by how automatable they are

Average AI-build difficulty of each kitchen's ingredients. Lower means more automatable. These are our editorial assessment of current tooling, not a measurement, and they rate automatability of the parts — not whether the game is fun.

Genre buildability — the four bands

Genre score is the average AI-build difficulty of a genre's core ingredients (our editorial model). Titles and recent-share are banded popularity counts from the MythicLemon Games Catalogue — a relative-popularity proxy, not unit sales.

FAQ