On 2 August 2026, the transparency obligations of the European AI Act came into force. A few months earlier, in January, Valve had rewritten its disclosure rules on Steam. In between, a study put a number on what the « AI » label actually costs a store page. For an independent studio, the question is no longer whether you use artificial intelligence, but where you put it in the pipeline and what you are willing to pay for that choice.
This article sorts it out. No promise of a fully automated production, no blanket refusal either: just what holds up today in a team of two to ten people, area by area, with the legal and commercial constraints that come with it.
What platforms and the law now require
For two years, disclosing AI use rested on a single checkbox and a fair amount of personal interpretation. That vagueness disappeared in January, and the European framework closed the remaining gap in early August.
Steam separated tools from content
On 16 January 2026, Valve replaced its single checkbox with a two-tier classification: pre-generated content shipped inside the game on one side, live generation during play on the other. More importantly, tools used purely to speed up development fell out of scope entirely — a code assistant, a debugging tool or an internal automation script no longer need to be declared. A game that generates dialogue on the fly, however, must both disclose it and document its safety guardrails. By March 2026, more than 7,300 games carried a disclosure on their Steam page. Valve, true to form, bans nothing: it labels, and lets the audience decide.
itch.io made disclosure mandatory for assets
The platform took a different route. The disclosure field is mandatory for asset packs and merely optional for games. A pack that answers « yes » is automatically given the « AI Generated » tag, the others get « No AI », and visitors can filter either way. The penalty is commercial rather than disciplinary: a pack built with AI and left untagged loses its indexing on browse pages, which amounts to vanishing from the catalogue. The rule applies even to elements hand-edited after generation.
The AI Act adds a European layer
Since 2 August 2026, Article 50 of the European AI regulation requires three things that bear directly on a game: informing people that they are interacting with an AI system when it is not obvious, marking generative outputs in a machine-readable format, and labelling deepfake-type content. The headline penalties — up to 15 million euros or 3 % of worldwide turnover — obviously target players other than three-person studios, but the text sets no size threshold. A companion character improvising its lines through an online model falls squarely within direct interaction.
What AI genuinely automates today
Code and internal tooling
This is the most profitable area and, since January, the only one explicitly exempt from disclosure on Steam. The real gains do not come from generating gameplay — results there remain poor — but from everything surrounding the game without ever appearing in it: import and data conversion scripts, in-house editors, build tooling, unit test generation, migrating from one engine version to the next. A small team recovers several days a month there with no commercial exposure. We already explored that shift in our piece on how AI is changing game development, and it is the part that has aged best.
Localisation
An indie game shipping in English only leaves a significant share of its market on the table, and professional translation costs long made multilingual releases unreachable below a certain budget. Machine pre-translation followed by native review changes that maths: you go from three languages to ten for the price of the review pass. Two precautions, learned the hard way. Freeze a glossary before you start, or the same object will carry three different names depending on the screen. And pull strings containing variables or interface width constraints out of the automated flow: those are the ones that break your layout.
Testing and QA coverage
Automated exploration bots, of which modl.ai is the most visible example, walk levels on a loop and surface what a tired human stops noticing: collision holes, geometry traps, localised performance drops, regressions after an engine update. That is coverage, not judgement. No bot will tell you the third level is boring or the tutorial arrives too late — that part stays human playtesting, and it remains the most valuable one.
Visual pre-production
Exploring twenty art directions in an evening, producing framing references, filling a blockout with placeholder textures: as long as nothing leaves the studio, nothing needs disclosing and the time saved is real. The switch happens at the precise moment a generated image ends up in the shipped build. That is where the calculation changes entirely.
What is better kept human
Visible assets are expensive in visibility
Ross Burton, of Game Oracle, studied nearly 10,000 Steam releases between January and October 2025. The raw figures already say a lot: games disclosing AI use averaged 4 reviews in their first month against 7 for the others; close to 20 % received none at all, against 15 %; and among titles passing 100 reviews, the average score dropped to 84.6 % against 88.3 %.
The causal model, which controls for developer experience, publisher backing, genre and release date, shows a gap of roughly 53 % fewer reviews. Since review counts serve as a sales proxy across the industry, two comparable games released the same week end up at 100 reviews on one side and 47 on the other.
The nuance that changes the decision
The effect is not uniform, and that is the study’s most useful result. Developers with no experience and no marketing budget lose almost nothing: their games were heading for invisibility anyway. It is the studios with talent, budget and know-how — the very factors that normally lift sales by 20 to 65 % — that take the sharpest hit. Put differently, the more you had to lose, the more the label costs. A studio five years into building its reputation is not in the same position as a first published project.
Voices and runtime dialogue
Voice synthesis is technically mature and lets a project with no dubbing budget exist at all. It is also the area the public watches most closely, the one that puts identifiable performers’ work at stake, and since August 2026 the one that triggers the most obligations. As for dialogue generated at runtime, it stacks everything: enhanced Steam disclosure, guardrails to document, user information under Article 50, and a dependency on an online service that may shut down or change its pricing during the game’s lifetime.
The professional climate has shifted too
In the State of the Game Industry 2026 survey, 52 % of the professionals questioned consider generative AI to have a negative effect on their field. Positive opinions fell to around 7 %, down from 13 % a year earlier. That is not a trivia point when you are hiring, looking for a publisher or pitching at a festival: the person across the table probably holds a settled view on the matter, and it is no longer a neutral one.
Three questions before opening a tool
First: will the result be visible to the player? If the answer is no, AI is a plain productivity gain and the discussion ends there. If it is yes, everything else applies.
Second: is this element part of the game’s signature? The art style, a main character’s voice, the musical theme that sticks in your head are exactly why an audience picks one independent title over another. Automating them erases what sets you apart.
Third: can I stand behind the disclosure publicly, store page included? If ticking the box makes you uncomfortable, the answer was usually already no.
How we work at the studio
At Itamde Studio, AI lives on the tooling, localisation and testing side, and does not enter shipped assets. This is not a moral stance: it is a visibility calculation, at a moment when the label costs more than it returns for a studio still building its reputation. Our two prototypes from the GMTK Game Jam 2026 were produced under that rule, and the time saved on internal plumbing went into finishing playable content.
To go further on cross-cutting uses, our selection of the best free AI tools for developers in 2026 covers the code and productivity side. For culture and video game news, our colleagues at SubwayPress follow how these releases land with audiences. And if you would rather dig into technical foundations than into this season’s tools, reference books on game AI remain the better investment: the behaviour and pathfinding algorithms they describe have no disclosure problem.
In short
AI today automates, at no commercial cost, everything the player never sees: tooling, data conversion, test coverage, a first translation pass, internal visual exploration. The moment it crosses into the shipped build it changes nature and becomes a positioning decision, with a disclosure to stand behind and a measurable visibility gap attached. The dividing line is not ideological, it is economic — and for an independent studio in 2026, it leans clearly to one side.
0 Comments