One Checkbox Saying a Game Used AI Tells You Nothing

The Problem

The AI disclosure on a Steam store page is one box. Used, or not used. Below it the developer writes a few lines of free text. Three completely different kinds of game currently sit inside that single box: games where AI is the mechanic itself, games shipping generated images as art assets, and games that used AI only to improve their localization. To a player, the badge says nothing.

The numbers show how overloaded the box has become. Per PC Gamer’s roundup, the share of releases carrying a disclosure went 10.9% in 2024, 19.9% in 2025, and 30.8% in 2026 as of July. PC Guide reports more than 17,000 games now carry one, and 30% of those list two or more uses. Sulka Haro’s full census of 53,597 Steam releases from July 2023 to July 2026 gets more specific. Of roughly 9,400 AI-flagged titles in that window, 87% were commercial non-performers, and 72% of those non-performers listed AI art, images, or textures. Titles that did perform skewed the other way: voice appeared in 24% of them versus 8% of the failures, localization in 18% versus 6%.

So the same badge carries opposite signals. A player avoiding asset-farm shovelware and a player who welcomes good dubbing and translation are looking at the same mark and have to guess. Studios lose too. A team that used AI only to ship five extra languages gets filed next to the generated-art flood, which is exactly why some teams now write the thinnest disclosure they can get away with.

Why Now

Three things line up.

First, the meaning of the flag started drifting. Valve revised the disclosure form to specify that it covers AI-generated content consumed by players, not efficiency tools used behind the scenes. The definition moved, so a 2024 checkbox and a 2026 checkbox no longer describe the same thing. Any analysis over the accumulated data breaks at that seam.

Second, the vocabulary for the argument is settling. A July 27 feature in Korea’s Inven makes the case that writing code with Claude does not make something an AI game. The paper it cites, “AI-Native Games: A Survey and Roadmap,” analyzes 53 released games and prototypes and carves out the category where removing AI removes the game. There is now language for the difference between AI as a tool, AI as material, and AI as the rules.

Third, there is something to port. In photography and video, provenance signing under Content Credentials has started shipping inside cameras and editing tools. Games have no equivalent yet, and they are actually the easier case. A photo has no single chokepoint, but a game build does. The pipeline already knows which file passed through which tool on its way into the package.

flowchart LR
  A[Editors and gen tools<br/>art, voice, text, code] --> B[Build pipeline hook<br/>records tool per asset]
  B --> C[AI use manifest<br/>machine readable]
  C --> D[Store label and verify API]
  D --> E[Player filter<br/>screen for what you want]

How to Build It

Start with a thin recorder that attaches to the build. A Unity or Unreal build hook walks the assets that made it into the shipped package and cross-references each file against its metadata and the project log to work out which tool produced it. Six columns is enough: art, voice, text, music, code, and runtime model calls. Each column gets three states: not used, machine-assisted with human editing, generated and shipped as is. That whole step has to be a few clicks for the developer or it dies on adoption.

The output is a manifest that ships alongside the game. It has to be a machine-readable table rather than a paragraph, because stores and community tools are the consumers. Generate the Steam disclosure text from that table so the studio never writes one again. Verification comes next: begin with self-report plus build logs as evidence, and tighten toward signing and reproducible builds only when false reporting actually becomes the problem. Perfect verification on day one is how this idea never ships.

Do not chase the big store first. Land somewhere policy moves fast, itch.io or a single indie publisher, and get the badge onto real store pages. Free for developers; charge stores and publishers for the verification API and for portfolio audits. Publishers are already a live buyer here, because their legal teams cannot currently answer which AI touched which part of their own catalogue.

Success Criteria

Two things decide it.

One is whether studios have a reason to bother. You need evidence that a granular disclosure sells better than a vague one. Haro’s finding that voice and localization use clusters among titles that performed is the starting argument, but it is correlation and should not be sold as more. Run an A/B with early partner titles and measure what the badge does to conversion directly.

The other is adoption. If no store honors it, the label is decoration. Which means the early design must not try to replace the Steam form. It has to be the tool that fills the Steam form in automatically, so no store has a reason to object, and standardization becomes a later step rather than a precondition. Plenty of provenance standards over the past few years tried it the other way around.