House House's AI Code Disclosure Exposes Three Different Rules for Game Studios
On this page 6 sections

AI-generated editorial illustration.
In a public answer added to the Big Walk FAQ, developer House House described a problem in Unity’s rendering pipeline that made its game run about 50% slower for roughly 20 seconds around sunrise and sunset. The studio says its own research and a conversation with a Unity engineer had not produced a fix. A chatbot eventually suggested a route that the team modified extensively. House House estimates the result involved about ten lines of code inside Unity’s Universal Render Pipeline.
The debugging account became a dispute about what a studio owes its players, workers and award jurors. House House says well under 0.1% of the game’s code had anything to do with its use of large language models, and says it used generative AI in no other part of making the game. It also says the game was built without an internal AI-use policy, a gap the studio now calls a failure. The percentages and boundaries come from the studio; no independent code audit is public.
Ten lines cannot settle the argument. Valve’s published Steam survey focuses on AI-created material shipped with a game and consumed by players; it excludes efficiency tools from that section’s focus. Nineteen Independent Games Festival jurors said they would not vote for any game made with generative AI in any capacity, including code. A studio can answer Valve’s question accurately and still lose the confidence of people who ask the jurors’ question.
This distinction arrives alongside a fresh labor signal. In the GDC Festival of Gaming’s November 2026 survey preview, 40% of respondents said they use generative AI at work, up from 36% in the comparison cited by GDC. Yet 55% described its impact on the industry as negative. Many people use a tool they distrust. House House’s account shows why a studio needs to know which question it is answering before it calls its disclosure complete.
A small fix became a public test
House House’s FAQ explanation is unusually specific for a public account of development-stage AI use. The studio describes shadow casters that Unity’s Universal Render Pipeline rendered at distant positions behind the player. It says that behavior slowed the game at two times of day. It asked a chatbot about a third-party code problem after other approaches failed. Some response snippets served as a starting point for code that the developers changed. The studio’s estimate of roughly ten lines and well under 0.1% applies to its own account of the final codebase; it does not measure the time spent researching, testing or reviewing the fix.
House House says it supplied a version of the explanation as a suggested part of its Independent Games Festival submission. It did not make an AI disclosure on Steam, saying the platform did not request information about code or chatbot use. When the issue drew public attention, the studio put the account on its FAQ. The answer describes its reading of a platform form. It does not say Valve reviewed the fix or endorsed that reading.
The statement describes a practical debugging use, then declines to endorse wider AI adoption. House House apologizes to people for whom even limited use crosses a line and says it will develop a more comprehensive internal policy. A tool can help solve a technical problem and still carry costs to workers or their craft. The ten-line estimate tells readers about the claimed size of one code change; it says little about those other costs.
The FAQ does not publish the code diff, model, prompts, dependency terms, internal review notes or a before-and-after benchmark. The stated 50% slowdown, roughly 20-second window and ten-line fix remain the studio’s account. There is no evidence that an AI system authored the game, that the fix escaped human review, or that the suggestion went into the build unchanged. What is visible is the studio’s use of a chatbot during development and its later decision to explain that use outside the Steam form.
Steam asks about what reaches players
Valve’s Steamworks Content Survey documentation says developers complete a survey before submitting a store page and build for review. Its generative-AI section recognizes that many development environments contain AI-powered tools. The stated focus is AI used to create content that ships with the game and is consumed by players. The examples include artwork, sound, narrative and localization. Valve says efficiency gains from tools are not the focus of that section.
The survey then separates two cases. Pre-generated AI covers player-consumed material created with AI assistance during development. Live-generated AI covers material created with AI services while the game runs. For live generation, Valve asks for a description of guardrails intended to prevent illegal content. The page also says that disclosure does not override the platform’s other content rules. Those categories answer a distribution question: what is in the product players receive, and what might be generated for them while playing?
House House’s account sits on a different side of that line. Its reported chatbot use concerned a programming problem and a small amount of modified code; the studio says no generative AI was used elsewhere in making the game. The code runs in a game, but Valve’s documentation expressly distinguishes efficiency-tool use from AI-created content that players consume. A code-assistance history and player-facing generated art are therefore not interchangeable categories. That is an interpretation of the published survey, not a claim about a private Valve review of Big Walk.
The scope matters to developers and customers. A player who reads a store badge as a complete inventory of development tools assumes more than the survey promises. A studio that treats the absence of a badge as a complete account of its process may miss a separate promise to employees, collaborators, award programs or fans. Valve defines the information it needs for store review. Other audiences ask for different information.
Even within the store survey, the timing of use changes the question. A developer might use an image model to make an early mock-up and replace it before shipping. A different studio might keep generated artwork in the finished game. A third might connect a model that creates dialogue while a player is online. The published Steam categories focus on material that reaches the player, with a further guardrail question for live generation. They do not resolve whether the first studio should tell its workers or fans about its discarded mock-up. They do tell the second and third studios where to begin their platform disclosure.
The distinction also affects a claim like “AI-free.” Valve’s form cannot certify that claim across a production history it does not ask developers to report. A creator who wants to make the broader promise would need its own definition, a way to check work from contractors and a correction path if the definition is later breached. A player who cares only about generated content visible or audible in the game may accept Valve’s narrower scope. Both positions are coherent, provided the label identifies which one it means.
The product boundary is particularly difficult for software. A rendered image, spoken line or paragraph is easy to identify as player-facing content. A code change is invisible to the player but can shape the experience, including performance. House House says the final fix came from substantial modification after the chatbot’s suggestion. Under Valve’s stated survey scope, that kind of debugging assistance is treated differently from a generated asset. Under the jurors’ stated test, invisible code still counts. The same fix can therefore be outside one disclosure form and inside another person’s judgment of the work.
Jurors drew a wider line
On November 9, a group of 19 named jurors for the 2027 Independent Games Festival issued a statement saying they would decline to vote for any game that used generative AI during development. Aftermath reproduced the full statement, and one signer, Austin Walker, posted its introduction. The signers said their objection includes AI-generated code as well as art, writing, voice and marketing material. They framed programming as creative labor and linked their position to concerns about workers, training data and pressure to adopt tools.
The group’s own limits are essential to the story. The statement says IGF leadership approved its release, but the signers do not speak for the festival or every juror. It says there was no universal consensus among judges and jurors. The public 2027 competition rules describe entry requirements and judging authority; the jurors’ statement does not create a blanket published ban or establish that Big Walk was formally disqualified. A refusal by some jurors to cast a vote is a consequential expression of professional judgment. It is not the same thing as a festival-wide eligibility decision.
That difference changes how a studio should read the dispute. Steam asks what AI-created content players encounter. These jurors ask whether the production process used generative AI at all. House House asks how to explain a narrow use in a way that acknowledges worker and audience concerns. Each question has a different owner and consequence: storefront review, an individual juror’s vote, or the studio’s relationship with its team and audience.
The jurors make a labor argument as well as an aesthetic one. Their statement says code belongs with art and writing in the category of creative work; they object to training on past work without consent and to employers compelling adoption. A studio may reject that position and still need to understand what it asks. It is not solved by proving that an output was useful or that a model only touched a fraction of a build. The concern is about the conditions under which the tool and its training material entered the workplace.
There is a countervailing practical problem. A team close to release can spend days on a defect in software it did not write, as House House says it did with Unity URP. If an AI-assisted clue helps engineers find a fix they test and edit themselves, the benefit is specific. A rule against all generative AI use would exclude that clue along with generated character art and dialogue. The jurors knowingly draw that wider line. Whether a studio accepts it is an artistic and employment decision with consequences for schedule, cost and reputation. The public evidence does not show how much House House saved or whether another method would have worked.

AI-generated editorial illustration. The three paths represent different audiences for a studio’s account of AI use.
The IGF submission page says released games can enter the competition and lists its 2027 schedule. It also notes that judges may see entries after the submission deadline and that later builds are permitted without a guarantee every juror will see an update. These mechanics make a submission disclosure more than a store-page checkbox: a studio is presenting a version of the work to a particular review process. They do not prove how any individual juror handled Big Walk’s entry.
The survey does not turn use into consent
GDC’s November preview offers a broader view of the disagreement without settling it. Its published figures say 40% of surveyed game-industry professionals use generative AI at work, compared with 36% in the prior period cited on that page. Among AI adopters, 57% say they use it daily. The rate among respondents in business operations and services is 71%. These are descriptions of surveyed people, not an audited count of studios or tools in shipped games.
On sentiment, 55% of respondents say generative AI has a negative impact on the industry, 35% see a mixed impact and 8% a positive one. GDC reports particularly negative views among quality assurance and community-support respondents, at 76%. The survey says people may find value in repetitive tasks, research, brainstorming and early prototyping while worrying about environmental cost, legal questions, jobs and skill loss. There is no need to force all those responses onto one axis. A worker may use a tool because it is expected in the job and still oppose the terms on which it was introduced.
Nor can the survey be used to infer a verdict on House House. Respondents were not polled on its specific Unity fix in the public preview. GDC does not show that generative AI caused a given layoff or reduced the hours needed to ship a game. It reports that 34% of respondents had been laid off in the past two years, and 54% of those laid-off respondents remained unemployed. GDC itself warns that its survey appears to attract more people affected by layoffs than aggregate employment models would imply. Those figures reveal the pressure felt in this respondent pool. They are not an industry-wide unemployment rate or a causal estimate for AI.
The labor context still matters. GDC reports that 57% of nondeveloper game workers in its survey usually worked more than 40 hours a week, compared with 44% of developers. It reports 8% experienced mandatory overtime; among that group, 69% received no compensation. A late-stage technical problem can sit inside a difficult production schedule. That does not prove House House faced those conditions, or that its chatbot use displaced a worker. It explains why a dispute over a debugging shortcut can connect to questions about who bears the cost of speed and who gets a say in the tools used to achieve it.
For a founder, a worker and a customer, the same disclosure can answer different questions. A founder may want to know whether a tool improved a deadline without adding unacceptable rights or quality risk. A programmer may want to know whether using it was optional and whether the team’s own work will train a vendor model. A player may care about generated art or writing in the product, but may also care about whether the studio publicly committed to a human-made process. The GDC percentages cannot decide those preferences for them.
An early-career signal in the same GDC preview makes the workforce question harder to postpone. GDC says 62% of educators in its survey allow, encourage or require students to use generative AI in game-development coursework, while only 20% of students say they actually use it. Those are two respondent groups, not a measured conversion funnel. Still, the gap warns a studio against assuming the next hiring cohort shares one settled practice. A junior programmer may arrive having been encouraged to use a tool in class, taught to avoid it, or given no clear standard at all. A team policy written only after a public controversy leaves the newcomer to guess.
Another GDC finding locates the budget pressure outside the headline AI debate. Forty-five percent of surveyed studio workers said their companies hired co-development services for their most recent projects. Half said self-funding was their primary source of funding. GDC’s categories describe respondents, and the public preview does not cross-tab those answers with AI use. They do show that a studio’s “team” may include outside contributors and that money is often tight. A practical policy needs to cover a contractor’s assets or code and say who checks the work before submission. Otherwise the studio may know what its core employees did but lack an answer about the whole game.
No survey result measures the price of a stricter AI rule. More paid review and production time may be required. A late dispute can also consume time and trust, including attention from jurors and customers. The balance depends on the project and audience. The Big Walk record supports a narrower point: House House calls its missing internal policy a failure even while describing a specific, limited technical use. The cost of waiting to decide appears at the moment a studio must explain itself.
A studio needs one record and several answers
A studio facing these different questions needs a record it can consult before answering any of them. That record may stay internal. It should separate chatbot research from code that entered the build, and both from artwork or dialogue shipped to players. It should identify who reviewed the output and which public claim the evidence can support. The table below is an editorial decision aid drawn from the published sources, not a legal requirement or a rule from Valve or IGF.
The first choice is who decides. A studio head can publish a rule quickly, but a rule that programmers, artists or contractors encounter only after work has been done will be hard to apply fairly. The relevant questions are concrete: can a worker use a model to research a bug; can output be pasted into a repository; must generated assets be replaced; who approves an exception; and what is owed to a collaborator who objects? These are proposed questions for a team conversation, not details House House disclosed about its own management.
The next choice is when to revisit the answer. A studio may change tools halfway through development or patch a game after release. The IGF submission page allows updated builds but does not guarantee every juror will see them. Steam says some content-survey answers cannot be edited after approval without contacting support. The two processes have different update mechanics. A studio that keeps a dated record of tool use and the version it submitted can correct the right statement at the right destination instead of posting a vague blanket correction.
| Decision surface | Question supported by the public record | Responsible answer |
|---|---|---|
| Steam store survey | Did AI help create content players consume, or generate content during play? | Apply Valve’s stated pre-generated and live-generated categories; do not treat a code-assistance log as a player-facing asset. |
| Festival submission | What does a particular competition ask, and how might jurors judge production methods? | Check current IGF rules and submission questions; distinguish a juror’s personal voting standard from festival policy. |
| Studio and team | Which tools were used, by whom, on what material, with what review and approval? | Keep an internal account that can support a precise public statement and a worker conversation. House House says its missing policy was a failure. |
| Audience statement | What can the studio substantiate about the final game and its development? | State the category and scope of use, label estimates as estimates, and avoid implying a platform’s silence is an endorsement. |
The matrix is deliberately not a yes-or-no AI badge. A studio could discover, for example, that a tool helped a programmer investigate a third-party engine bug but no generated assets reached players. It might have no Steam disclosure under the documented player-content scope, yet still need to answer a festival question and its own team’s policy commitments. Conversely, an AI-generated voice line in a shipped game would raise a player-facing survey question even if no award submission existed. The same word, “AI,” covers different production events.
For the internal account, five fields make the later explanation testable: the tool and date, the material supplied to it, what came back, what entered the final product, and who checked the result. A studio should also keep the policy version in force at the time and the applicable external forms. That is a practical record design inferred from the House House episode and the published platform categories. It is not evidence that House House did or did not keep such records.
The distinction between input and final output is especially important in a creative team. An engineer may use a chatbot to search for a possible approach, then write and test the fix. An artist may ask a model for a mood board that never appears in the game. Another team may ship a generated texture. A single declaration that “AI was used” hides these differences; a declaration that “no AI content was disclosed on Steam” may hide them in the other direction. A useful statement says what happened in the relevant category and acknowledges what it cannot establish.
Percentages deserve the same care. “Well under 0.1%” is House House’s estimate of code associated with AI use, with the final codebase as its apparent denominator. It is not a measure of development time saved, creative decisions affected, or risk to an artist’s livelihood. Those would need different records and denominators. Equally, a claim about zero player-facing generated art would not establish that no developer consulted a model. The more a studio relies on a small number to reassure people, the more it should state what was counted.
It should also distinguish permission from disclosure. A team can allow a particular use while recording it for a festival submission or internal review. Another team can prohibit use even if a platform never asks about it. These choices can change as tools or staff change, but the public account of a shipped game should reflect what the team actually did. A policy that defines use precisely gives employees a way to make that distinction before they have to reconstruct it under pressure.
Trust has more than one audience
The House House dispute is likely to remain uncomfortable because the three tests are not merely administrative variants. Valve’s survey is designed around shipped content and customer expectations. The jurors’ stance concerns the origin of creative work and the labor system behind it. A studio policy concerns who can use tools, when, and under what review. None can be substituted for the others by a low percentage of code.
House House’s public answer provides enough detail to locate the disagreement. A rendering bug, a chatbot suggestion, a heavily modified fix, about ten lines, no Steam code-use disclosure requirement as the studio understood it, and an internal policy it says should have existed. The remaining questions are not answered by a better slogan. They concern whether a team had a chance to set its own boundaries, whether the studio can support the scope of its claim, and whether the audience was promised more than a store survey actually asks.
For the next small studio approaching a deadline, the decision comes before the tool is opened. It can decide which kinds of assistance may enter the build, what goes into the internal record, and when a public explanation is owed. House House gave a detailed account after the question became public. The next team may face the same three audiences, each asking for a different answer about the same few lines of work.