Steam AI disclosure rules changed quietly in the early summer of 2026, and most of what’s written about them still describes the 2024 version. If you ship games — or commission the 3D assets that go into them — the current rule is both narrower and sharper than the coverage suggests: narrower about what you must disclose, sharper about who carries the risk when an undisclosed AI asset slips through. We sit on the asset-supplier side of this market, so this affects our clients directly; here’s the rule as it actually reads in mid-2026, the data on what disclosure does to a launch, and the provenance habits that keep a Content Survey clean.
What the Steam AI disclosure rules require in 2026
The current Steamworks Content Survey draws one line: does the AI-generated content ship and get consumed by players? Two categories survive from the 2024 policy. Pre-generated covers “any kind of content that ships with your game and is consumed by players that is created with the help of AI tools during development” — art, audio, and text alike. Live-generated covers content created while the game runs, and carries one duty on top of the same description requirement: telling Valve what guardrails you have put on the AI to keep it from generating illegal content. Live-generated adult sexual content is barred outright — the survey says the legal and customer risks are “substantial enough that we don’t want to ship Live-Generated AI Adult Only Sexual Content at this time”, a line that has been in place since Valve’s original announcement. The 2026 rewrite added the change that matters most day-to-day: efficiency tooling is no longer the section’s focus — “efficiency gains through the use of these tools is not the focus of this section,” in Valve’s own words — and the word code quietly dropped out of the list of examples the old text carried. That de-emphasis is real: if a tool helps your team work and produces nothing a player consumes, the current wording reads as putting it outside the section’s focus. What Valve has not said is that AI-generated production code compiled into a shipped game is categorically exempt — and a code assistant can write exactly that. Treat the disappearance of the word as a narrowing of focus, not a safe harbour, and re-read the section before you rely on it. So the section aims at AI-generated content that reaches players — pre-generated in the build or generated while the game runs — and the resulting disclosures surface to players on the store page. Notably, Valve announced none of this — the survey text simply changed. You can date it from the page’s own archive: the old wording (“any kind of content (art/code/sound/etc) created with the help of AI tools during development”) was still live on 13 May 2026, and the current wording was in place by 17 June. No announcement, no changelog, a five-week window. That is how this policy will keep evolving, and it is on you to re-read it.
What disclosure does at launch: the data, with its caveat
Steam AI disclosure is now common enough to measure. Totally Human Media’s tracking put AI-disclosing titles at 10,258 games — about 8% of the whole Steam library by late 2025, and roughly one in five new releases on an earlier snapshot — grossing an estimated $660 million combined, by the tracker’s own “roughest back-of-the-envelope” method. The sharper number comes from Game Oracle’s December 2025 analysis of nearly ten thousand commercial releases: games disclosing AI use collected about half the first-month reviews of comparable non-disclosing titles, after controlling for developer experience, publisher backing, type of game and month of release. Before those controls, the same games also showed lower positive ratings and about half the pre-release followers. Now the caveat the analyst himself insists on, and so do we: that’s correlation. Heavy AI reliance may simply travel with rushed craft; the study can’t prove players are punishing the label itself. What it does show is that AI-disclosing games were a materially weaker cohort even after the study’s own controls. Whether players are reacting to the label, to the AI use behind it, or to whatever else tends to travel with both, the study cannot separate — and the cohort alone doesn’t decide a launch either. Arc Raiders carries a Steam AI disclosure, became one of the year’s biggest shooters and sits at Very Positive lifetime; it also drew a 2-out-of-5 review written specifically about its AI, had its design director tell PCGamesN, discussing the game’s visuals, that it “in no way uses generative AI whatsoever” — while confirming in the same interview that its text-to-speech voice system is the reason for the Steam disclosure — and re-recorded some of its AI voice lines with real actors after launch. That is the realistic version: disclosure plus evident craft survives contact with players, and still costs you an argument. Being found out after launch adds a transparency problem on top of whatever objection players already had.
The failure mode that compounds it: discovered, not disclosed
The clearest 2025–26 controversies mostly follow one script: players found the AI use before the studio had plainly declared it. The Alters shipped with a leftover AI prompt visible in flavor text and undisclosed AI translations — players found it, and the studio’s apology conceded the point: “we should have simply let you know.” Call of Duty took months of AI-asset backlash while saying nothing, and the disclosure came after. Frontier pulled Jurassic World Evolution 3’s AI scientist portraits under fan pressure before the game shipped. In the first two, staying quiet stacked a transparency problem on top of an objection players already had to the AI itself — and it did so in an ecosystem where a curator group polices disclosures, marking non-disclosing games “Not Recommended” and honest ones “Informational”. Frontier is the harder case and worth being straight about: it disclosed up front, on the store page, exactly as the rule asks — and the disclosure itself is what players revolted against. Declaring removes the specific accusation that you hid it. It does not protect you from an audience that objects to the tool. And here’s the part that concerns anyone buying outsourced assets: Steam’s rule attaches to the shipped content regardless of who made it. Your vendor’s undisclosed generative shortcut becomes your Content Survey problem, your store label, and your apology thread. A leftover asset and some unannounced translations were enough to force a public apology out of 11 bit studios.
Buying assets with clean provenance
The practical consequence of the 2026 rule for commissioning teams is a short checklist, all of it sourced from how the platforms now work:
- Put AI use in writing with every asset vendor. A one-paragraph statement per delivery — what was hand-modeled, what used generative tools and where — is what lets you fill the Content Survey accurately instead of hopefully. Any studio doing genuinely manual work will provide this without flinching; we consider it a fair ask of ourselves and of any competitor.
- Use the marketplace labels where they exist. itch.io has required AI tagging on asset pages since late 2024, with untagged work dropped from indexing and, in its own guidelines, liable to delisting; Epic’s Fab launched “requiring sellers to clearly state if their products were created with the help of generative AI programs”. So the declarations are visible — check them, but read a label as the seller’s own representation rather than an audit. One Fab creator published “over 41,000+ AI-generated assets in a short timeframe” before Epic investigated and pulled the work, which is roughly how much protection a self-report buys you.
- Scope the exemption correctly. A vendor using AI purely to work faster — reference searches, scheduling, internal notes — is not what the section is aimed at. What matters is whether generative output ships in the files you receive. Ask the question in exactly that form, and don’t let “we only use it internally” close the conversation until you know what “internally” produced.
- Extend the same care to marketing assets. Whether store-page art falls inside the rule’s shipped-content wording is genuinely less clear-cut — but in practice it hardly matters: marketing and promotion accounts for about one in nine disclosed AI uses among Totally Human’s $1M+ earners, and players have forced AI marketing art down without any rule involved. Treat the capsule-art brief with the same provenance care as the hero character’s.
Our own corner of this: hand-modeled work — the kind we deliver as game-ready low-poly assets and character design — keeps that survey answer simple by construction. That’s not an argument that AI tooling is illegitimate; it’s an argument that in 2026, knowing exactly what’s in your files is a commercial asset in itself.