Ten Lines of AI Code: Why the Games 'Purity Test' Breaks Down at the Engine

What this is about

Big Walk is the new co-op game from House House, the studio behind Untitled Goose Game. It launched on August 4, sits at a Metacritic score of 91 according to Inven Global, and sold more than a million copies in its first week. Late in development, the team hit a bug: at sunrise and sunset, the game ran about 50 percent slower for roughly 20 seconds. The cause wasn't in the game itself but in Unity's Universal Render Pipeline (URP), the part of the engine that draws the graphics. URP was treating shadow casters behind the player character as if they were infinitely far away. The team investigated on its own, brought in a Unity engineer, and finally asked a chatbot. According to the studio, its suggestion was "heavily modified." What remained were about ten lines in URP. House House puts the AI share of its total code at "well under 0.1 percent."

The studio had disclosed this when it submitted the game to the Independent Games Festival (IGF). On November 9, Austin Walker posted a statement on Bluesky from 19 IGF jurors who say they will no longer vote for games that use generative AI "in any capacity." Big Walk isn't named in it, but juror Bruno Dias publicly confirmed that this was the game in question. House House then added the details to the FAQ on its website, and on November 10 and 11 the Steam pile-on followed.

Our thesis: a test for "no AI in any capacity" can't be passed, and it can't be verified either. So it doesn't measure who uses AI. It only measures who admits it.

What exactly happened, and what it cost

The jurors write that they refuse "to vote for any game that uses generative AI in any capacity." They continue:

"AI-generated code is no exception. Programming is a creative discipline whose work deserves the same respect as any other part of game development."

According to Aftermath, IGF leadership approved the statement's publication. The signatories stress, however, that they don't speak for the organization. Walker says there are about 50 jurors this year, so the 19 are not a majority. They're calling on the IGF to introduce a rule for "future years" that excludes games containing "any AI-generated work." No IGF decision on this has been announced so far. According to Game Informer, the awards ceremony is on March 3, 2027.

Dexerto tallied Big Walk's Steam reviews over several days. We calculated the percentages from its numbers:

Gigazine counts more than 400 negative reviews in two days. Steam filters this kind of "review bombing" (a flood of negative reviews in a short window) out of the overall score, which is why Big Walk still sits at "Very Positive" overall. We could not independently verify the daily figures on SteamDB.

House House apologized, called the lack of an internal AI policy "an oversight," and announced that it would create one. One worry the studio itself raised, according to GameSpot, stands out: voluntary disclosures could normalize AI. The studio hadn't declared anything on Steam, for a simple reason: Steam doesn't ask about code or chatbots.

Layer by layer: where AI code already lives

Only a small part of a typical indie game is the studio's own code. Underneath sit an engine, libraries, drivers and the operating system, and the code gets written in development tools that often have AI built in. Here's what's publicly documented for each layer:

A pull request (PR) is a proposed change submitted to a project. Agentic workflows are processes in which AI agents largely write, run and fix code on their own.

According to Wassmer, Claude and Codex eliminated "non-deterministic cooking bugs" at Epic that the team had failed to fix for a decade. Cooking is the step where the engine converts game content into the format of the target platform; non-deterministic means the bug shows up sometimes and sometimes doesn't. So if you ship an Unreal game, you ship those fixes too. With Big Walk there's an extra irony: the ten lines aren't in the game code at all, but in Unity's own engine. How much AI is in the rest of the URP code, House House has no way of knowing.

What the rulebooks actually require

The Indie Game Awards have already shown how this plays out in practice. In December 2025, they stripped Clair Obscur: Expedition 33 of its Game of the Year and Best Debut awards. Sandfall Interactive had stated at submission that it used no generative AI, but later admitted to AI placeholder textures that had already been patched out. Again, the case only came to light because the studio owned up. ARC Raiders is the flip side: on Steam, Embark vaguely disclosed "procedural and AI-based tools" and uses AI text-to-speech for voices, while also saying it doesn't use generative AI. In practice, every studio decides for itself what counts as "generative."

Why the test measures honesty

Self-reporting is the only data source there is. Ichiro Lambe (Totally Human) used the Steam API to count 7,818 games with an AI disclosure, about 7 percent of the catalog. He stresses that this figure is based solely on voluntary disclosures and the real share is likely higher. Ross Burton of Game Oracle analyzed nearly 10,000 Steam releases and, according to games.gg, found that an AI disclosure goes along with weaker sales, regardless of quality. That's a correlation, not proof. But it fits the picture: disclosing costs you something, staying quiet costs nothing.

Experts make the same point. According to Gigazine, developer and lecturer Brendan Keogh considers a blanket ban counterproductive, because it gives the studios using AI most heavily a reason to hide it. Juror Liz Ryerson criticizes that the difference in scale is ignored, and that of all studios, the one getting hammered is the one that honestly reported something minor. Critics of the boycott point out, per Inven Global, that development environments like Visual Studio have shipped AI autocomplete for a long time, and that asking a chatbot is a lot like searching Stack Overflow. One user on the Defold forum argues that an "AI code" label is simply unenforceable.

The other side has serious arguments. Boycott supporters, according to Inven, note that commercial language models were trained on copyrighted works without consent, and that small exceptions erode boundaries later on. Dias argues that the distinction between "advisory" use and actual code generation makes no difference. That's a moral position, and a defensible one. But as a competition rule, it only works if it can be verified.

Our take

Our assessment: taken literally, "in any capacity" would rule out every future Unreal game, probably every game running on Linux, and likely every game written in Rider with default settings. No studio can rule that out for its entire software stack, not even the engine maker Defold. So the rule can only be enforced against those who disclose something, which means it rewards silence. House House didn't have to report anything to comply with Steam's rules, and if it hadn't, it would probably have been left alone.

There are alternatives that start from reality. Steam asks about what players see and hear. Linux and Mesa require labeling without punishing it. A tiered approach based on what was made with AI (art, voices, text, code), where (in your own work or in third-party software) and how much would still let juries apply strict standards without penalizing honesty. It remains to be seen whether the IGF issues a rule at all, and how it defines "AI-generated." That will show whether it ends up judging tools, or just confessions.