A Video Game Is a Recipe of ~45 Ingredients. AI Has Solved Some Kitchens and Barely Touched Others.

Half the story everyone repeats is the wrong half

There is a comfortable story going around that AI is about to build games. It is half true, and the wrong half is the half people repeat.

So we took a different approach. We broke a game down into its parts — the roughly 45 reusable ingredients that every commercial game is assembled from — and scored each one by how much of the heavy lifting a generalist can hand to AI tools today, not in principle. Then we wrote each major game type out as a recipe and asked a flat question: given what AI can build right now, who could actually ship this?

The answer is not 'games.' The answer is 'some games, and very specifically not others.' Which ones depends entirely on which kitchen the core ingredients come from.

A game is cooked in five kitchens, and AI has only learned two of them

Every game draws from five kitchens: Art and visuals, Audio, Narrative and content, Code and systems, and Polish and production. AI has not made uniform progress across them. It has made spectacular progress in a narrow band and almost none in another, and the line between the two is sharp enough to plan a business around.

In the strong zone — call it difficulty 1 to 2, near-solved — sit 2D backgrounds, concept and marketing art, voice acting, music, sound effects, ambient audio, flavour text, branching dialogue, worldbuilding and lore, inventory and crafting code, save and menu systems, procedural-generation systems, and platform integration. Voice acting is the single most-automated craft in games. If your game is mostly pictures, sounds, words, and well-trodden plumbing, AI does most of the work and a human curates.

In the weak zone — difficulty 4 to 5, barely moved — sit game-ready 3D characters, skeletal animation, combat systems, level design, puzzle design, encounter balance, performance optimisation, and live backend. And then the two that score a flat 5, the hardest things in the medium: game feel and juice, and real-time authoritative netcode. These are the only two ingredients in the entire ledger that resist AI almost completely.

The pattern underneath is clean. Generation of artefacts — a picture, a track, a paragraph, a boilerplate class — is largely solved, because an artefact can be judged on its own. Synthesis under real-time, fairness, and feel constraints is not, because it can only be judged in motion, against a player. AI compresses the production of a game by an order of magnitude. It compresses the design of one barely at all.

The buildable genres and the unbuildable ones are 2.4 points apart

When you score each game type as a weighted mean of its ingredients — lower meaning more automatable — the spread is wider than most people would guess. The most buildable archetype sits at 1.80. The least sits at 4.17. That is a 2.37-point gap on a 5-point scale, and it falls along a single fault line: content versus feel.

At the easy end, every core ingredient is in AI's strong zone — the full ranking is in the first table below. A visual novel at 1.80 is the most completely automatable commercial game format that exists: style-locked illustration, backgrounds, branching dialogue, worldbuilding, voice, music, and a save-and-menu shell — not one core ingredient outside AI's reach. The only human job left is taste: which fifty of five hundred generated portraits to keep, and whether the writing has a soul.

At the hard end — the second table — the score is dragged up by the ingredients no tool touches. A fighting game at 4.09 and a battle royale at 4.17 live or die on frame-perfect feel and authoritative netcode, the two ingredients scoring 5. AI shaves the edges of each. None of the cores collapse.

Between the two ends sits a smooth ramp: 2D platformers at 2.91, metroidvanias at 3.12, horror at 3.14, roguelikes at 3.25, survival at 3.38. The platformer is the trap of the set — its art and music generate themselves, so it looks AI-friendly, but its whole soul is controller feel, level design, and juice, the three things AI is worst at. A jump that does not feel good kills the game no matter how good the art is.

'Buildable' is not 'good,' and the data knows it

This is the caveat that holds the whole study together, so we will say it plainly. These scores measure the automatability of the parts, not the quality of the whole. You can now generate every asset and most systems of a competent platformer in a weekend, and it will still feel dead without the human game-feel pass that no tool automates. The score tells you where the labour collapses, not where the craft does.

That is also why the easy list should make you suspicious rather than excited. If a generalist can build a visual novel in a day, so can everyone else. Buildability alone is not an opportunity. It is a commodity.

The opportunity is the overlap, not the easy column

The useful question is not 'what is easy to build.' It is 'what is easy to build and has proven, growing demand.' Pull the catalogue for archetypes that score easy, carry real volume, and are still being made — at least 40% of their titles released in 2023 or later — and a short, specific list survives. It's the third table below.

These are not dead niches you would be automating into a graveyard. Walking simulators have 369M owners and over half the genre arrived in the last few years. Interactive fiction — pure writing, AI's single strongest domain — has 3,137 titles and a live, growing audience. This is the quadrant where the thing AI builds best meets the thing buyers keep paying for. Just outside it sits farming sim at 2.38 with 59% of its titles recent, the hottest momentum among the easy genres, held back from the table only by a smaller catalogue.

The most defensible bet is the genre AI half-builds and can't finish

Now the contrarian read, and it is the most actionable line in the whole study.

The fastest-growing archetype in the catalogue is deckbuilding — 64% of its titles released in 2023 or later, faster than anything else we measured. And it scores 3.18. Hard. Not because the art is hard; card art generates instantly. It scores hard because a deckbuilder's soul is balance, and balance is the AI-hardest content work there is. Compare it to fighting games, frozen at 17% recent — an expertise-gated niche that small teams do not break into. Deckbuilding is the opposite: demand is shouting, and the hard part still resists automation.

Sit with that for a moment. The highest demand sits on top of the hardest-resisting core. That is not a problem. That is the most defensible position on the board. Let AI generate every card, every illustration, every line of flavour text for free — and spend the entire human budget on the one thing it cannot do, which is making the numbers sing.

Because that is the shape of the whole map. There are exactly four things AI still cannot do: game feel and juice, real-time authoritative netcode, level and puzzle and encounter design, and the 'is it fun?' balance-and-cut loop. As AI commoditises production, all the value migrates into those four. The winning move is not to pick the easiest genre. It is to pick a genre whose content AI builds for nothing and whose craft AI cannot touch — and to spend your whole budget on the craft.

The fine print

Two different kinds of number live in this piece, and they were produced two different ways. The per-feature difficulty scores and the genre buildability scores are our editorial assessment of current tooling as of early 2026 — judgement calls about what a generalist can hand to AI today, not measurements. They will move; the fastest-shifting frontiers are 3D characters, skeletal animation, and combat systems. The demand and momentum figures — game counts, owners, and the share of titles released in 2023 or later — are pulled live from our local mirror of the Steam catalogue (81,957 games, 1.14M tags). Every figure here was recomputed from our own full crawl. And the load-bearing caveat once more: these scores measure the automatability of the parts, not whether the result is any good. Buildable is not good.

The most AI-buildable archetypes

Buildability is a centrality-weighted mean of each genre's core ingredients' AI-difficulty (1-5, lower = more automatable). Editorial assessment, early 2026.

The least AI-buildable archetypes

These genres' identity rests on game feel and netcode — the two ingredients AI has barely moved.

AI-easy and in demand

Genres scoring easy to build that also carry real volume and recent momentum. Demand figures are live counts from the Steam catalogue mirror.

FAQ