Generative Art in Big-Budget Games and the Price of Cheap Assets

Every publisher earnings call in the last two years has developed a stock phrase for generative AI, and the phrase is always some variant of “efficiency.” Electronic Arts’ leadership has talked publicly, in investor calls since 2023, about generative tools cutting the cost of asset production; Ubisoft has spoken about “Ghostwriter,” an internal tool built to draft first-pass NPC barks — the incidental lines background characters shout that a studio previously paid writers to hand-craft one at a time — for a human writer to edit rather than originate from scratch. None of that is secret or disputed. The companies say it themselves, in the specific, careful language finance departments use when they mean “this got cheaper.” The part worth examining is what “cheaper” is actually a euphemism for, because a texture or a line of dialogue that costs nothing to generate didn’t get more valuable. It got produced by a process with different failure modes than the one it replaced, and those failure modes show up later, in places a spreadsheet doesn’t track.

What “efficiency” is actually describing

The efficiency case for generative tools is genuinely strong at the margins the industry talks about most: rough concept iteration, first-pass barks for crowd NPCs nobody will remember individually, placeholder textures during early production that get replaced before ship. Used that way, generative tools are doing exactly what a junior artist’s rough sketch or a temp recording used to do — filling a gap cheaply so the team can judge the shape of a scene before spending real budget finishing it. That use case has existed in some form for decades and generative tools just make it faster.

The trouble starts when “first pass” quietly becomes “final pass” because the budget that would have paid for the human finishing step got cut on the assumption the generative output was already good enough. That’s not a hypothetical slippery slope; it’s the specific mechanism finance departments are incentivised to pursue once a line item shows up as reducible. A studio under budget pressure doesn’t need anyone to make an explicit decision to ship unfinished machine output — it only needs the finishing budget to be the first thing cut when a project runs over elsewhere, because that budget is the one place on the sheet that now has a cheaper alternative sitting right next to it.

What a texture generated without reference actually loses

The specific failure mode worth naming in detail is coherence — not visual quality in isolation, which generative tools can produce at a startlingly high baseline, but coherence with everything around a given asset. A hand-painted texture for a stone wall in a specific game’s specific dungeon was made by an artist who had seen the lighting rig, the adjacent assets, the game’s overall palette rules, and who made hundreds of small, consistent decisions informed by all of that context at once. A generatively produced equivalent, prompted in isolation against a general training distribution, has no access to any of that context by default — it can be prompted toward a described style, but it wasn’t produced with an artist standing in the actual scene checking whether this particular crack pattern reads correctly against that particular light source. The result, at scale, across hundreds of assets, is a specific kind of visual noise: individually plausible textures that collectively fail to cohere into a world that feels considered, because considering a world as a whole is precisely the part of the job that got skipped.

This is a design problem, not a purity argument. It matters to a systems reader for the same reason a poorly tuned difficulty curve does: the player’s felt experience of coherence is downstream of thousands of small decisions, and a process that strips out the decision-maker doesn’t strip out the decisions’ consequences. It just moves who absorbs them — from a budget line during production to a player’s sense, during play, that something in the scene doesn’t quite belong there, without necessarily being able to say why.

Marketing art became the canary before gameplay assets did

The clearest documented cases of generative output shipping into a finished product without adequate scrutiny have, so far, mostly hit marketing rather than in-game assets, and that pattern is itself informative. Wizards of the Coast publicly acknowledged in 2023 that promotional artwork for Magic: The Gathering had included AI-generated elements sourced from a contracted studio, following fan identification of telltale generative artefacts — warped hands, incoherent background text — in official marketing materials, and apologised for the lack of disclosure. Marketing art is exactly where this would surface first, because it moves fastest, gets the least individual scrutiny relative to its visibility, and is furthest from the core team whose reputation depends on the game itself. In-game asset pipelines have more structural friction against the same failure — a texture that breaks coherence with its surroundings gets caught by an art director doing a pass over the whole scene, in a way a single marketing banner reviewed by a different department often doesn’t. That friction is exactly the kind of review step a cost-driven pipeline is under the most pressure to shorten.

Procedural generation already ran this experiment once

The industry has actually run this exact test before, with a different technology producing the same shape of trade-off, and the results are worth remembering before repeating them with a new tool. No Man’s Sky’s procedural planet generation in 2016 promised effectively infinite worlds from a compact set of generative rules, and delivered exactly that — quintillions of technically distinct planets, each individually the product of a real algorithm rather than a hand-placed decision. What players actually encountered, within hours, was the specific texture of proceduralism: recognisable rock formations recombined in slightly different arrangements, creatures assembled from a limited part-library that started to rhyme with each other after enough hours, biomes that were locally varied but globally homogeneous in a way that became legible faster than the promotional material had implied it would. The lesson Hello Games eventually absorbed, across years of subsequent updates, was that infinite generation isn’t the same achievement as infinite interest, and that a smaller number of hand-considered set pieces dropped into the generated framework did more for a player’s sense of discovery than any increase in the raw combinatorial space. Generative AI art tools are being sold on precisely the same promise — effectively unlimited variation, produced cheaply — and the same lesson applies before a single player has to learn it the hard way again: variation without curation reads, eventually, as noise rather than richness, no matter how technically impressive the generation process behind it was.

Who actually catches the mismatch, and when

The deeper structural problem is timing. An art director reviewing a scene built the traditional way is reviewing work produced by people who already made hundreds of contextual judgment calls before the review ever happens, which means the review is mostly confirming coherence that was already built in. A director reviewing a scene assembled from generatively produced assets is reviewing work with none of that judgment pre-applied, which means the review has to catch every contextual mismatch cold, asset by asset, across a volume of content that was specifically generated because reviewing it individually the traditional way was judged too expensive. That’s not a theoretical tension — it’s the exact bottleneck that determines whether the savings from generation actually survive contact with a finished, coherent product, or simply relocate the labour from creation to inspection without reducing the total amount of human judgment a scene actually needs to look right. A studio that cuts the art-direction review pass to match the generation savings hasn’t found an efficiency. It’s found a new place to hide the same cost, one level closer to the player.

The tell that separates the honest use from the quiet one

The honest version of this technology, used the way EA and Ubisoft describe it publicly, keeps a human finishing pass as a non-negotiable step regardless of how good the generative first draft is — the tool accelerates iteration, the studio still owns the final coherence check. The quiet version treats the generative output’s baseline quality as sufficient justification for skipping that check, on the theory that the savings are real and the coherence cost is diffuse enough that no single reviewer will be blamed for it. The tell, as a critic, is the same one that flags cheap assets in a rushed release long before generative tools existed: repetition that reads as procedural rather than intentional, environments that are individually detailed but collectively indistinguishable from each other, texture work that’s high resolution but low information. Procedural generation has always run this same risk — a system that can produce infinite content still needs a human decision about which of that infinite content is worth showing a player, and skipping that decision is the actual cost, whether the content came from a hand-coded algorithm or a trained model.

What the saving is actually buying

None of this argues that generative tools have no place in game production; plenty of the internal, iteration-stage use cases are a genuine improvement on what came before, and pretending the technology is uniformly bad would be as dishonest as pretending it has no costs. The argument is that “efficiency,” as a word in an earnings call, describes a saving on one line of the budget without describing where the corresponding cost lands, and the corresponding cost — a world that reads as considered rather than generated — is exactly the quality that separates a game a player remembers from one they finish and forget. The day-one patch and the death of the finished build taught the industry that a deferred cost doesn’t disappear just because nobody pays it at launch. Generative asset pipelines are teaching the same lesson in a different currency, and the invoice, when it eventually comes due, gets paid in the felt texture of the world itself, one uncanny wall panel at a time.

The honest question a studio should be asking isn’t whether generative tools belong in the pipeline at all — that ship has sailed, and pretending otherwise wastes energy better spent on the actual question, which is where in the pipeline the human judgment gets preserved and where it quietly gets priced out. A pipeline that keeps a director’s coherence pass intact, and uses generation only to reach the point where that pass has something worth judging, is a faster version of the same craft that always shipped good games. A pipeline that treats the generation step as the finish line has built something cheaper to produce and more expensive, eventually, to be believed in by anyone who plays it.