AI-Generated Game Assets: What Steam's Disclosure Rules Actually Require

Somewhere on nearly a third of the new games that showed up on Steam this year, there's a small, easy-to-miss disclosure box on the store page you've probably scrolled past without reading closely. It says something to the effect of "this game uses AI-generated content," sometimes with a sentence or two of detail underneath. Most players skim past it the way they skim past age ratings and language warnings. Which is a little bit of a shame, because the actual rule behind that box has changed twice in three years, tells you less than you'd assume about how much AI was actually involved in making the game, and — according to real sales data — correlates pretty strongly with a game's odds of actually selling.

Here's what the policy actually says, how Valve got to this specific version of it, and what happens to a game's fortunes once that box gets checked. It's a more interesting story than a single fixed rule, because it isn't one — it's a policy that's already been rewritten twice in three years, shaped by a lawsuit, a court ruling on who can even hold a copyright over AI output, a wave of artist protest that predates Steam's involvement entirely, and now a genuinely measurable market reaction from the millions of players deciding, purchase by purchase, what that little disclosure box is actually worth to them.

The Backlash That Set the Stage, Before Steam Ever Wrote a Policy

Valve didn't arrive at any of this in a vacuum, and the timeline matters: by the time Steam made its first move on AI content in June 2023, the wider games and art industries had already spent roughly six months in open conflict over exactly this issue. It started on the art side, in December 2022, when a working artist posted an image to her ArtStation portfolio reading "No to AI Generated Images," with the word AI crossed out — a specific, small act that snowballed within days into a mass protest as thousands of artists posted the same image to their own portfolios, flooding ArtStation's trending page. The core objection wasn't abstract: many of those artists believed, with real specificity, that the AI image generators producing work now competing with theirs commercially had been trained on their own copyrighted portfolios without permission or compensation — a genuinely pointed irony given that ArtStation is itself owned by Epic Games, the same company whose CEO would later become one of the most vocal public critics of Steam's disclosure requirement on the game-storefront side of this same fight.

The legal fight escalated in parallel. On February 7, 2023, Getty Images filed suit against Stability AI, the company behind the Stable Diffusion image generator, accusing it of copying more than twelve million Getty photographs — along with their captions and metadata — to train a commercial tool that Getty argued was now competing directly against its own licensing business. That's not a symbolic protest; it's a specific, large commercial entity alleging specific, countable infringement against a specific AI company, filed in the same window that the copyright registration fights over Zarya of the Dawn were also playing out at the US Copyright Office. Valve's decision, four months later, to start rejecting AI-touched games outright wasn't made in isolation — it landed in the middle of an industry already actively litigating, protesting, and publicly fighting over exactly the ownership question Valve's own policy was trying to avoid getting caught in the middle of.

It Started as an Outright Ban, Not a Disclosure Rule

The current disclosure system isn't where this started. In June 2023, Valve's actual practice was to reject games from Steam outright if they used AI-generated assets, on fairly narrow legal grounds: the copyright status of AI-generated content was (and largely still is) unresolved in most jurisdictions, and Valve didn't want to be in the position of distributing content where a developer couldn't clearly demonstrate they owned the rights to everything that went into training the tool that made it. That's a real, specific legal concern, not vague AI-skepticism — if an AI model was trained on copyrighted material without permission, and its output ends up in a game Valve is distributing and profiting from, that's a genuine liability question with no settled case law to point to.

That blanket-rejection approach didn't last long, mostly because it was quickly overtaken by reality: AI-assisted tools were being adopted fast enough, across enough of the developer base, that treating "any AI involvement at all" as disqualifying was becoming impossible to enforce consistently and was catching plenty of legitimate, non-infringing uses in the same net as genuinely risky ones. On January 10, 2024, Valve announced a full reversal — replacing the ban with a disclosure requirement instead, on the theory that transparency, not prohibition, was the more workable answer to a problem that wasn't going away.

The Two Categories the Current Rule Actually Splits Into

The January 2024 framework, which (with one significant refinement covered further down) is still the basic structure in place today, works by having developers fill out an expanded Content Survey — the same onboarding form every game already goes through before release — with a new section specifically asking how AI was used. That answer gets sorted into one of two categories, and the difference between them matters a lot more than most players realize.

Pre-Generated AI Content covers anything created with an AI tool during development that ships baked into the final game files — a set of textures generated in an image model, character concept art made with Midjourney or a similar tool, voice lines produced with AI speech synthesis, or even lore and dialogue text drafted with a language model. The defining feature is that it's fixed at the moment the game is packaged: what a player experiences on day one is exactly what shipped, AI-origin or not, and it doesn't change or generate anything new after that point.

Live-Generated AI Content is a genuinely different animal, and Valve treats it with real additional scrutiny for a specific reason: it's content the AI produces while the game is actually running, in response to what the player does — an NPC generating unique dialogue on the fly instead of picking from a pre-written script, a quest system that writes new objectives dynamically based on player choices, procedurally narrated events that are different every time. Because nobody, including the developer, can fully predict what a live system might generate in every possible situation a player creates, Valve's Content Survey requires developers to describe the specific guardrails they've put in place to prevent the system from generating illegal or clearly offensive material — a meaningfully higher bar than simply checking a disclosure box, because it asks the developer to show their actual safety work, not just admit AI is involved.

There's one further, quite specific line Valve drew even in the initial 2024 policy: adult sexual content produced by a live-generated AI system is handled with particular caution — Valve has said it's addressing those cases individually as they're reported rather than allowing them under the same general framework as everything else, since the unpredictability problem live-generation creates everywhere else in the policy is judged too high-risk to extend fully into that specific category.

It's worth understanding that this disclosure requirement doesn't exist in a vacuum separate from Steam's ordinary content rules — it sits on top of them. Every developer publishing on Steam already agrees to the Steamworks Distribution Agreement, which includes a standing promise that the game won't contain illegal or infringing content and that it'll be consistent with its own marketing materials. The AI disclosure doesn't replace that obligation or create a separate, looser standard for AI-touched content — Valve has been explicit that during prerelease review, AI-generated output gets evaluated against the exact same content standards as anything hand-made. The disclosure box tells a player how a piece of content was produced. It isn't a different, more lenient bar for what that content is allowed to contain.

What Actually Shows Up on the Store Page

The disclosure a developer submits through the Content Survey doesn't just sit in Valve's internal records — a version of it gets published directly on the game's public Steam store page, which is the actual box players occasionally notice and usually skim past. What's on that box, in practice, ranges enormously in how informative it actually is: some developers write a genuinely specific sentence or two — which tool was used, for what part of the game, and why — and others write the legal minimum needed to satisfy the requirement, a single generic line that technically discloses AI use without meaningfully explaining anything about it. Both satisfy Valve's rule equally. The rule requires that the box exists and isn't false; it doesn't require the box to be genuinely useful to a player trying to decide whether they care.

The January 2026 Refinement: Dev Tools Are Out, Player-Facing Content Is In

Two years into the disclosure regime, Valve made a meaningful adjustment. As of January 2026, developers are no longer required to report AI-powered tools used purely behind the scenes during development — the kind of "efficiency tooling" that includes AI code assistants, automated debugging tools, and similar production aids that never directly touch what a player actually sees or hears in the finished game. The disclosure requirement now applies specifically and only to generative AI content a player actually experiences: art, audio, dialogue, localization text, or promotional material that either ships inside the game or appears on the store page itself.

This is a real narrowing of scope, and it's worth understanding why Valve made the change rather than just noting that it happened. Under the original 2024 wording, a studio using an AI coding assistant to write boilerplate code, or an AI-powered bug-triage tool, was in a legally ambiguous position about whether that counted as reportable "AI use in development" — even though a player would have zero way of detecting or being affected by it, since it never touches anything player-facing. Narrowing the rule to player-facing content resolves that ambiguity cleanly: the disclosure now answers the specific question a player might actually care about — "did AI make something I'm going to see, hear, or read in this game" — rather than a much broader and less useful question about a studio's entire internal production pipeline.

How Fast This Has Actually Grown — The Real Numbers

The scale of what this policy now covers is worth putting in actual numbers rather than a vague "AI is everywhere now" gesture, and there's genuinely solid data to draw on here. Game designer and data analyst Sulka Haro — formerly a lead designer on Habbo Hotel, now CEO of the studio Mainframe Industries — published a full census of every game released on Steam between July 2023 and July 2026, cross-referencing each one against Valve's AI-disclosure field. The resulting picture is a clean, steep growth curve: AI-flagged releases made up about 7% of new games when the disclosure field first went live in early 2024, climbing to 10.9% across all of 2024, 19.9% across 2025, and 30.8% for 2026 so far as of Haro's analysis. On a monthly basis, AI-flagged launches grew from roughly 13 a month at the very start to around 530 a month by mid-2026 — and Haro's analysis found that AI-flagged titles account for somewhere between 60% and 90% of the entire month-over-month growth in Steam's total release count over that period. Non-AI releases have grown too, from roughly 1,030 to about 1,320 a month across the same window, but nowhere near as fast — nearly all of the platform's recent acceleration in raw release volume is coming specifically from the AI-flagged side of the ledger. If the current trajectory holds, Haro's analysis projects AI-disclosed games crossing half of all new Steam releases sometime around 2027 or 2028.

A single-week snapshot makes the trend concrete in a way the yearly percentages can undersell: in the week immediately before Steam Next Fest in June 2026, more than 300 new games shipped on the platform, and 120 of them — a 40% share of that week's releases — carried an AI disclosure.

The Harder Story the Growth Numbers Don't Tell on Their Own

Here's where it gets genuinely more interesting than a simple "AI adoption is skyrocketing" headline, and it's the part that matters most if you're actually trying to understand what this trend means rather than just how big it is. Haro's same analysis also tracked estimated sales, not just release counts, and the two numbers pull in noticeably different directions. In the first quarter of 2026, AI-flagged games accounted for 28% of new releases but only about 17% of estimated sales across those releases. In the second quarter, the gap widened further: AI-flagged titles made up 33% of new releases but captured only around 10% of estimated sales. In plain terms, AI-flagged games are showing up in ever-larger numbers, but a shrinking share of the money is actually flowing toward them relative to their share of the shelf space.

It's worth being fair to what that gap does and doesn't prove. Steam's overall economics are brutally top-heavy for every category of game, AI-involved or not — the top 1% of all titles on the platform account for something like 94% of total estimated revenue, meaning the overwhelming majority of any given release, AI-flagged or not, was never going to make meaningful money regardless of how it was made. So "AI games underperform proportionally" isn't the same claim as "AI games are uniquely doomed" — most games of any kind are commercial long shots on Steam's current release volume. What the gap does suggest is that AI-assisted development is lowering the barrier to releasing a game considerably faster than it's lowering the barrier to making a game people actually want to buy — more supply, without a proportional increase in games that clear the bar players are actually willing to pay for. The indie game boom's real drivers were mostly about distribution and tooling lowering the barrier to reaching an audience at all, not about the volume of releases — this is a related but distinct story, where a new tool category is mostly increasing raw release volume without moving the harder, separate needle of actually making something people want.

One more genuinely interesting wrinkle from the same dataset: Haro's analysis looked at the actual wording developers used in their disclosure text, comparing games that went on to succeed against ones that flopped. Disclosures on flopped games skewed shorter and more generic — a median of about 13 words, with 35% of them being a single line of ten words or fewer. Disclosures on successful games ran longer on average, around 22 words, and were considerably more likely to include hedging or reassuring language about how limited or careful the AI use actually was — 59% of successful games' disclosures used that kind of minimizing framing, compared to 41% of the flops. That's a correlation, not a controlled experiment, and it's entirely possible both patterns are downstream of the same underlying cause (a more careful, better-resourced team writes both a better game and a more thoughtful disclosure) rather than the disclosure wording itself changing anyone's buying decision. But it's a real, measured pattern in the actual data, not a guess.

What "Guardrails" Actually Means in Practice

Valve's requirement that developers describe their guardrails for live-generated content is easy to read past as a vague compliance phrase, so it's worth being concrete about what that description typically actually covers, in terms any developer building one of these systems would recognize. Content filtering is usually the first layer — a separate model or rule-based system that screens an AI's output before it ever reaches the player, checking for specific categories of clearly prohibited material rather than trying to judge quality or appropriateness in general. Word- and phrase-level deny-lists handle the crudest, most obvious violations cheaply, while a second AI model specifically trained to classify harmful content handles the subtler cases a simple list would miss. Rate-limiting and scope-limiting are a second common layer — deliberately constraining what topics or actions a live-generation system is even allowed to attempt, rather than trying to filter an unlimited range of possible output after the fact; a system that can only generate combat dialogue variations, say, has a much smaller space of possible failure than one that can generate anything at all in response to open text input. Human review queues, where genuinely unusual or high-confidence-of-risk outputs get flagged for a person to check before or shortly after they reach players, are a third layer many live-service systems already use for other purposes and can extend to AI-generated content specifically. None of this makes a live-generation system provably safe in some absolute sense — it can't, given how the category is defined — but it's the actual, specific kind of engineering work Valve's Content Survey is asking a developer to describe honestly, not a box that can be satisfied with a sentence claiming good intentions.

Steam Is Genuinely an Outlier Here — Other Platforms Handle This Very Differently

It's easy to read a policy this detailed and assume every major PC storefront has converged on something similar. They haven't, and the gap is wide enough that it's become a genuine point of public disagreement between platform holders rather than a settled industry norm. The Epic Games Store has no AI disclosure requirement at all as of this writing, and that's a deliberate choice, not an oversight — Epic's CEO, Tim Sweeney, has publicly and repeatedly criticized Steam's approach, describing the disclosure box as effectively a "scarlet letter" that he argues does more to invite review-bombing and organized "cancel" campaigns against small developers than it does to genuinely inform buyers. GOG, the DRM-free storefront, likewise has no clear equivalent disclosure policy — and found itself in the somewhat awkward position of facing its own backlash after using AI-generated artwork in its storefront's own promotional material, without disclosing it. Itch.io, the indie-focused platform, sits somewhere in between: it offers a searchable "AI Generated" content tag that developers can voluntarily apply, and has required disclosure of generative AI use in some form since November 2024, but doesn't enforce a Steam-style mandatory field on every game's page the way Valve does.

The Sweeney "scarlet letter" framing is a real position with real industry backing, but it's not the only view developers themselves hold. A separate industry survey on the subject found the opposite instinct is actually more common among working developers than Sweeney's public position might suggest: roughly nine in ten of the developers surveyed said they thought AI use in games should be disclosed more fully than current norms require, not less — a genuinely inconvenient data point for the argument that disclosure requirements are primarily something imposed on an unwilling developer community from outside, rather than something a meaningful share of that community actively wants.

What Happens When Disclosure Fails: Two Real Cases

The abstract policy is easier to understand once you see what actually happens when a studio gets the disclosure question wrong — and there are two real, public examples from the last couple of years worth knowing, because they illustrate a pattern that's more nuanced than "using AI gets you in trouble."

The first involves Clair Obscur: Expedition 33, which lost recognition at an Indie Game Awards ceremony not because it used AI-assisted content in its production, but specifically because it hadn't been transparent about that use beforehand — the industry reaction made clear that the penalty was for the non-disclosure itself, not for the underlying use of the tool. The second involves Crimson Desert, where the developers ended up publicly apologizing after failing to disclose generative AI involvement at launch — the AI disclosure was only added to the game's Steam listing after its release date had already passed, a sequencing that on its own generated real player frustration, including at least one documented case of a player being denied a Steam refund over a purchase made before that disclosure existed.

Notice what Valve itself didn't do in either case: it didn't delist either game, algorithmically suppress it, or apply any platform-level penalty beyond publishing the (eventually corrected) disclosure. Valve's own enforcement posture here is closer to how it treats a mature-content or violence flag than how it treats, say, a confirmed case of stolen assets — the rule requires the information to be accurate and present, and then leaves the consequences, whatever they turn out to be, to how players actually react to that information once they have it. In practice, per the same sales data discussed above, those consequences have been real and measurable — researchers analyzing the broader dataset have described a genuine "AI stigma" effect, where games disclosing generative AI use see meaningfully fewer reviews, weaker wishlist activity, and lower sales on average than comparable non-AI titles. Valve built the disclosure mechanism. The market, not Valve, is the part actually doing the punishing.

The Legal Backdrop This Whole Policy Is Actually Responding To

None of Valve's back-and-forth on this makes much sense without understanding the specific legal uncertainty sitting underneath it, which predates Steam's policy entirely and remains genuinely unresolved. The clearest illustration is a real, specific case: in September 2022, an author named Kris Kashtanova received a US copyright registration for a graphic novel called Zarya of the Dawn, part of which was illustrated using the AI image tool Midjourney. When the Copyright Office learned how the images had actually been made, it revisited the registration — and in a decision issued February 21, 2023, it drew a sharp and specific line: Kashtanova retained copyright over the book's written text and over the creative choices involved in selecting, arranging, and combining those elements into a finished work, but the individual AI-generated images themselves were found to lack the human authorship the Copyright Office considers a legal prerequisite for copyright protection, and were excluded from the registration.

A second case reinforced the same principle from a different angle. Stephen Thaler had sued the Copyright Office after it denied registration for a visual artwork he claimed was produced entirely autonomously by an AI system, with no human author at all. In August 2023, a federal district court sided with the Copyright Office, holding plainly that human authorship is an essential element of a valid copyright claim under US law — a work with no human author, by design, doesn't qualify.

Put together, these two rulings are exactly the uncertainty Valve's original June 2023 stance was trying to avoid distributing into: if the images, audio, or text an AI tool generates for a game may not carry enforceable copyright protection at all — or may carry it only in some partial, legally untested way depending on exactly how much human creative control shaped the specific output — then a storefront distributing that content is operating in real legal grey area, regardless of how good the tool or the resulting asset actually is. Valve's shift from banning to disclosing didn't resolve that underlying uncertainty. It's a business decision about how to keep operating a functioning storefront while the legal system, working through cases like these two, gradually sorts out an answer that doesn't exist cleanly yet.

The most direct test of the other side of this question — not whether AI output is protectable, but whether training an AI model on someone else's copyrighted material in the first place is itself infringement — reached a judgment more recently, and it cuts in a genuinely different direction. On November 4, 2025, the UK's High Court issued its ruling in Getty Images's case against Stability AI, the first UK judgment to squarely address copyright infringement in AI model training. The court rejected Getty's core copyright infringement claim, finding only narrow, limited liability on a separate trademark issue — a result that reads, on its face, as more favorable to AI developers than the earlier US decisions might suggest. But the win was considerably narrower than the headline implies: Getty had conceded during the proceedings that Stable Diffusion's actual training took place outside the UK, which meant the court's jurisdiction over the core training-copyright question was sharply limited from the start, and the substantive question of whether training on scraped copyrighted images constitutes infringement was left largely unresolved rather than settled in either direction. Two and a half years after Valve's first move on this issue, in two different countries, on two different legal questions, the pattern holds: real court decisions keep landing, and none of them yet add up to a clear, generally applicable answer either side of this debate can point to as final.

What a Genuinely Specific Disclosure Looks Like, Compared to a Minimal One

It helps to see the actual gap between "technically compliant" and "genuinely informative" side by side, since the earlier discussion of disclosure length and hedging language stays fairly abstract without a concrete comparison. A minimal disclosure, satisfying the letter of Valve's requirement and nothing more, might read something close to: "This game contains AI-generated content." That's a true, complete, rule-satisfying sentence. It's also close to useless to a player trying to decide whether they care — it doesn't say what was AI-generated, how much of the game it touches, or which of the two categories it falls under.

A genuinely specific disclosure, covering the same underlying reality, might instead read something like: "Background environment textures for secondary areas were generated using an AI image tool and then hand-edited by our art team; all character art, animation, and voice acting were created entirely by our human artists and cast. This is pre-generated content — nothing in the shipped game changes or generates new material during play." Same underlying legal category, same disclosure requirement satisfied, but a completely different amount of actual information handed to a player deciding whether to buy. Given the correlation Haro's analysis found between disclosure specificity and sales performance, the second version isn't just more honest — it's also, on the available evidence, the more commercially sensible choice, even though nothing in Valve's rule requires that level of detail.

What This Actually Means If You're a Player

Practically, if you're deciding whether a specific game's AI disclosure should factor into your purchase, there are a few things worth actually checking rather than treating the box as a single up-or-down signal, and it's worth knowing where to actually look: the disclosure appears in a dedicated section of the store page, usually below the main description and system requirements, distinct from the age-rating and content-warning information that appears elsewhere on the same page — it's easy to miss if you're used to skimming past the standard boilerplate sections every store page carries.

First, check which category it falls under — pre-generated content that shipped fixed on day one is a fundamentally different situation from live-generated content still producing new material every time you play, and the disclosure text should tell you which one you're looking at, even if it doesn't always use Valve's exact category names. Second, check how specific the disclosure text actually is, using the comparison above as a rough benchmark: a developer who names the exact tool and the exact asset type it touched is telling you something concrete; a single generic sentence is telling you only the legal minimum, which is worth knowing but shouldn't be read as reassurance either way. Third, and this follows directly from the sales data above: an AI disclosure by itself predicts essentially nothing about whether a specific game is good, since both excellent and poor games exist on both sides of that box being checked — what the aggregate data shows is a market-wide pattern across thousands of titles, not a verdict on any individual game you're actually looking at. A short, well-reviewed game from a small team using AI-generated background art responsibly and a bloated, poorly-made game using entirely hand-crafted assets can both exist, and frequently do — the disclosure tells you about method, not about quality, and conflating the two is the most common way this specific piece of information gets misread.

What This Means If You're the One Making the Game

For developers, the practical compliance picture as of 2026 is narrower and clearer than it was under the original 2024 wording, which is worth knowing if you last checked this rule more than a year or two ago and assumed it still worked the old way. Internal efficiency tooling — AI-assisted code completion, automated testing or debugging aids, project-management or asset-pipeline tooling that never produces something a player directly sees or hears — no longer needs to be reported at all under the January 2026 update. What still needs disclosure is narrower and more specific: any art, audio, voice, dialogue, localization, or promotional material generated with AI that ends up in front of a player, whether that's baked into the shipped build or generated live during play. If your live-generated system produces player-facing content of any kind, Valve's Content Survey still requires you to describe the actual guardrails preventing it from generating illegal or clearly offensive output — a real design requirement, not just paperwork, since you need those guardrails to actually exist and actually work before you can honestly describe them.

Given the sales data above, there's also a genuine strategic argument, separate from the legal compliance question, for writing a specific and substantive disclosure rather than the shortest sentence that technically satisfies the rule — not because Valve rewards longer disclosures in any algorithmic sense, but because the correlation between vague, minimal disclosures and weaker sales suggests players (and whatever word-of-mouth or review dynamics follow from their reaction) respond better to a team that seems to have actually thought about how it's using the tool, rather than one that appears to be checking a box as fast as possible and moving on.

A couple of edge cases are worth flagging specifically, because they're exactly the kind of boundary case that trips developers up more than the obvious ones. Localization and translation text generated with an AI tool counts as player-facing content requiring disclosure under the current rule, even though it's easy to mentally file translation work under "internal production tooling" the way code assistance is — the distinction that actually matters isn't where in the pipeline the tool sits, it's whether a player ultimately reads, hears, or sees the output, and translated in-game text is squarely something a player reads. The same logic extends to store page and marketing material: AI-generated screenshots, key art, or trailer narration used to market the game is covered by the same disclosure requirement as content inside the game itself, since a player deciding whether to buy is just as much "player-facing" as a player already in the middle of the game. The practical rule of thumb that resolves most edge cases cleanly: ask whether a player, at any point in the purchase or play experience, is the one actually consuming the specific output in question — if yes, it needs disclosure regardless of which part of the production pipeline it came from; if a tool's output is consumed only by the development team internally and never reaches a player in any form, it falls under the January 2026 exemption instead.

The One Thing Worth Remembering Through All of This

Go back to the Zarya of the Dawn and Thaler rulings for a moment, because they're the piece easiest to lose track of once a topic turns into a policy-and-numbers discussion. Those decisions didn't get overturned or superseded by anything Valve, Epic, or itch.io has done since — the underlying legal uncertainty about who, if anyone, actually owns AI-generated output is exactly as open today as it was when Valve first reacted to it in June 2023. Every disclosure box, every percentage-point growth statistic, every platform-comparison argument about scarlet letters and stigma is downstream of that one still-unresolved question. "Steam allows it now, with disclosure" is a real and meaningful policy fact. It is not the same claim as "the ownership question got answered," and treating the two as equivalent is the single easiest mistake to make when this topic comes up. Only one of those two things has actually happened.