The Four Things AI Still Can't Build in a Game

If AI can build a game, why does my game still not feel like one?

Here is the question that should be bothering anyone who has actually tried to ship a game with AI tools beside them. The asset pipeline genuinely collapsed. You can generate a wall of concept art, a soundtrack, a voice cast, a worldful of flavour text and most of the boilerplate plumbing in an afternoon, and it is competent. Then you press play, and the thing on screen is still inert. The jump feels mushy, the hits land on nothing, the enemies cluster stupidly, and no amount of regeneration fixes it. So what exactly is the part that didn't get easier?

We went looking for a precise answer rather than a vibe. In our Games-as-Recipes study we broke a commercial game into roughly 45 reusable ingredients and gave each one an AI-build difficulty score on a 1-to-5 scale, where 1 means AI does it near end-to-end and 5 means expert, real-time craft. That score is our editorial assessment of what tooling can do today, not a measurement, and crucially it rates how automatable a part is, not whether the finished game is any good.

When you lay all 45 ingredients out, the strong zone and the weak zone separate cleanly, and four ingredients sit at the very top as the things AI has barely touched. They are game feel and juice, real-time authoritative netcode, level and puzzle and encounter design, and the 'is it fun?' balance loop. This article is about those four, why they resist, and what to do about it.

The five kitchens, and the two AI never learned to cook in

Every game is cooked in five kitchens, and in our buildability model they do not score anywhere near each other. Audio averages 2.2 across its ingredients, art 2.57, narrative 2.78, code 3.0, and polish and feel 3.75. That spread is the whole story in one line. The kitchens that produce artefacts you can look at and judge on their own are largely solved. The kitchen that produces the feel of the thing in your hands is not.

Look closer at the easy end and you see why. Voice acting scores a 1 in our assessment, the single most-automated craft in games. Backgrounds, concept art, flavour text all sit at 1; music, sound effects, ambient soundscapes, branching dialogue, PBR texturing, inventory and save systems all sit at 2. Every one of those is an artefact. A picture, a track, a paragraph, a boilerplate class can be evaluated in isolation, which is exactly the condition under which a generator can be trained and graded.

Now look at the ceiling. Game feel and juice scores a flat 5. Real-time authoritative netcode scores a flat 5. Level, puzzle and encounter design all score 4, and the QA-and-balancing loop scores 4. Combat systems, performance optimisation and live backend round out the 4s. These are not artefacts. None of them can be judged sitting still on a screen. Every one of them is only legible in motion, against a player, in real time.

Ceiling one and two: game feel and netcode, the only flat 5s

Game feel is the millisecond coupling of input, animation, camera and sound into a single sensation that the controller is connected to the world. A jump that feels right is dozens of tiny, correlated choices: input buffering, coyote time, the easing curve on the launch, a few frames of squash, a hitch of camera lag, a screen shake scaled to impact, a sound that fires on the exact frame of contact. Change any one by a few milliseconds and the whole thing reads as wrong without a player being able to say why. There is no artefact here for a generator to produce. The output is a feeling, and the only test for it is a human hand on a pad, which is why it sits at 5 in our assessment.

Real-time authoritative netcode is the other flat 5, and it resists for a related reason: correctness is defined by an adversary. A server must reconcile inputs from many clients across unpredictable latency, predict and roll back, resist cheating, and stay fair under packet loss and contention — and 'works' means works against players actively trying to break it, on the worst connection in the room, at the worst possible moment. The failure modes are emergent and only appear under live, hostile load. You cannot generate your way to that any more than you can generate your way to a building that stands up in an earthquake; the property being satisfied lives in the interaction, not in any line of code you can hold up on its own.

This is also where having precise eyes on the running editor stops being a luxury. The whole reason these two resist automation is that they are only legible in motion, so the tools that earn their keep are the ones that let you watch the live state rather than guess at it. Mythic Dev Assist exists for exactly this seam: it gives an AI coding agent a queryable, causal view into the running UE5 editor over MCP, so the agent can observe what actually happens frame to frame instead of pattern-matching a static file. It does not build the feel for you — nothing does — but it closes the gap between the agent and the only place the answer lives, which is the game while it is moving.

Ceiling three: level, puzzle and encounter design

Level design, puzzle design and encounter balance each score 4 in our buildability model, and they fail automation as a group for one shared reason: a level is not a layout, it is a sequence of experiences paced against a player who is learning. AI will happily generate a geometrically valid room, a maze with a solution, or a fight with the right number of enemies. What it cannot reliably author is the teach-then-test rhythm, the safe place to breathe before the hard part, the readable sightline that pulls you toward the objective, the one corridor that turns a fair fight into an ambush. Those are judgements about how a human mind will move through space over time, and they are graded by playing, not by looking.

Puzzles are the sharpest case. A generator can produce a puzzle that has a solution; it has a much harder time producing one with a single intended insight, no unintended shortcut, and a difficulty curve that lands the 'aha' instead of either spoiling it or frustrating it. Encounter balance is the same problem wearing combat clothes — the question is never 'is this winnable' but 'is this winnable in a way that feels earned at this point in the player's growth.' Both are content work, and our study is blunt that content built around a real-time human response is the AI-hardest content there is.

The practical consequence is that procedural and AI-generated levels are excellent at the substrate and poor at the spine. Use them to fill, to vary, to populate — and keep a human owning the critical path, the pacing beats, and the encounters that have to teach something. The cost of building the level dropped; the difficulty of designing it did not move.

Ceiling four: the 'is it fun?' balance loop

The fourth ceiling is the quietest and the most fatal, and in our assessment the QA-and-balancing loop scores 4. It is the iterative cycle of play, notice it is not fun yet, change one number or one rule, play again, and repeat until the thing sings. Nothing about that loop is an artefact either. Its input is a felt verdict — too grindy, too swingy, too easy on turn three, dead in the mid-game — and its output is a tuning change validated by the same felt verdict on the next pass. There is no static target to optimise against, because the target is whether a person enjoyed it.

This is exactly why our contrarian pick in the study is deckbuilding. It is the fastest-growing genre we measured, with 64% of its titles released since 2023, and yet it scores 3.18 — systems craft, not buildable — specifically because its core is balance. AI will generate every card, every illustration, and every line of flavour text on that game for nothing. It will not tell you that the deck is broken on turn four, because knowing that requires having played a hundred games and felt the break. The content is free; the balance is the entire job.

The takeaway generalises past deckbuilders. Whenever a game's identity rests on tuning — an economy, a progression curve, a combat triangle, a difficulty ramp — AI changes your cost, not your difficulty. It clears the production work off your desk so that you can spend the whole human budget on the loop that decides whether anyone has fun. That is not a consolation prize. It is the actual job.

What the four ceilings have in common

Step back and the four are not a random list, they are one property wearing four costumes. Game feel, netcode, level and encounter design, and the balance loop are all judged in motion, against a player, in real time — never in isolation. That single shared trait is the entire reason they resist automation while art and audio and flavour text fell over almost completely. A generator can be trained and graded on artefacts that stand on their own. It cannot be trained on a verdict that only exists in the interaction between a system and a person playing it.

This is why the strong-zone-to-weak-zone split in our model is so clean rather than a gentle gradient. Generation of artefacts is largely solved because an artefact can be evaluated by itself. Synthesis under real-time, fairness and feel constraints is not, because the only available test is to run it and have a human respond. AI compresses the production of a game by something like an order of magnitude. It compresses the design of one barely at all.

So the honest one-line summary of where we are: buildable is not good. You can now assemble every asset and most systems of a competent game in a weekend, and it will still feel dead until a human does the four things on this list. The labour collapsed. The craft did not.

What to actually do with this

The strategy that falls out of the four ceilings is almost embarrassingly direct. Let AI take everything in the strong zone — the art, the audio, the writing, the boilerplate — and refuse to spend a human hour on any of it that a generator can do at 2 or below. Then aim the entire human budget at whichever of the four ceilings your game's identity actually rests on. If you are making a platformer, that is feel. A competitive shooter, netcode. A Metroidvania or a puzzle game, level and encounter design. A deckbuilder or any economy-driven game, the balance loop.

Pick your genre with that math in mind. The least defensible thing you can build is a game whose whole soul is something AI also builds for free, because then so can everyone else — buildability alone is a commodity, not an opportunity. The most defensible is a genre whose content AI generates for nothing and whose craft AI cannot touch, where you let the production go to zero and pour everything into the part that resists. That is the shape of every winning move on this map.

And because all four ceilings live in motion, instrument the motion. Measure rather than eyeball: a tool like Lumen Meter reads scene brightness so you light to spec instead of guessing at exposure, and Fast Chart Widgets drops live graphs and stat overlays straight into your UI so you can watch frame time, damage curves or progression pacing as you tune. Pair that with an agent that can actually see the running editor through Mythic Dev Assist, and you have closed the only loop that matters — the one where a system and a player meet, which is the exact place AI still cannot reach for you.

The four ceilings AI hasn't moved

AI-build difficulty is our editorial assessment of current tooling (1 = AI does it near end-to-end, 5 = expert real-time craft). It rates automatability of the part, not whether the game is good.

The five kitchens, by average AI-build difficulty

Average ingredient difficulty per kitchen in our buildability model. Lower = more automatable. Editorial assessment of current tooling.

FAQ