Quick summary: On 17 January 2026, Valve rewrote its AI disclosure policy to separate AI tools used during development — which need no disclosure — from AI-generated content that ships to players, which does. By March 2026 more than 7,300 Steam games had disclosed AI content. Epic Games Store has no disclosure requirement at all, and Epic CEO Tim Sweeney has called Steam’s rules irresponsible. The underlying fight is not about ethics; it is about discovery.
For two years, "does this game use AI?" was a question storefronts answered inconsistently or not at all. In January 2026 Valve made its position specific, and in doing so created the closest thing the industry has to a standard. It also created a labelling system that a large part of the development community now argues is actively harmful. Both things are true at once, which is what makes this worth understanding properly.
What the January 2026 rewrite actually changed
The original policy was ambiguous about a distinction developers cared about enormously: the difference between using an AI tool to help make a game and shipping AI-generated material inside it. Under the rewrite, that distinction is explicit.
| Use of AI | Disclosure required? |
|---|---|
| Code assistants, debugging tools, refactoring help | No |
| Playtesting analysis, internal prototyping, concept exploration | No |
| Art or textures that ship in the game | Yes |
| Audio, music or voice lines players hear | Yes |
| Dialogue or written content players read | Yes |
| Localisation shipped to players | Yes |
| Marketing assets on the Steam store page | Yes |
The rule of thumb Valve landed on is clean: if a player sees or hears it, disclose it. If it only ever existed in the studio, it is a tool like any other. Digital Watch Observatory’s summary of the rule change is a good primer if you want the policy language itself.
For developers: the safest reading is that disclosure attaches to shipped artefacts, not to workflow. A game written entirely with an AI pair-programmer but shipping hand-drawn art and human-written dialogue does not require a disclosure under the 2026 policy.
The scale: how many games are tagged
By March 2026, more than 7,300 games on Steam had disclosed AI content. Reported figures put AI disclosure at around one in five new releases — a share that would have been unthinkable when the original policy was drafted.
That number is doing two different jobs at once, and conflating them is where most of the argument comes from. Some of those 7,300 are studios using generative tools responsibly for placeholder art, localisation passes or texture variation. Others are the phenomenon the community has bluntly named "slop" — high-volume, low-effort releases produced quickly enough to flood a storefront and make genuine discovery harder for everyone, including other small developers.
Why Epic disagrees
Epic Games CEO Tim Sweeney has been the loudest critic of the approach, arguing that AI disclosures are becoming meaningless because AI will soon touch nearly all game production. The Epic Games Store carries no AI disclosure requirement.
The argument is not unreasonable. If a label eventually applies to almost every game, it stops distinguishing anything — the same way a "made with a computer" tag would have been informative in 1985 and useless by 1995. And there is evidence the label carries a real commercial cost: reporting suggests AI-tagged games attract substantially fewer reviews than untagged ones, which on Steam is close to a proxy for visibility.
The counter-argument is equally straightforward. A label that costs something is a label that means something. If disclosure had no commercial consequence, it would not function as disclosure — and the players choosing to avoid tagged games are exercising exactly the informed preference the policy exists to enable.
Worth flagging honestly: the "53% fewer reviews" figure circulating in coverage comes from third-party analyses rather than from Valve, and the causation is unclear — tagged games may under-perform because of the tag, or the tag may correlate with lower-effort releases that would have under-performed anyway. Treat it as directional, not settled.
The real issue is discovery, not AI
Strip out the AI framing and the underlying problem is one storefronts have had for a decade: when publishing is free and unlimited, the constraint moves from production to attention. Generative tools did not create that problem. They accelerated it by removing the last practical brake on release volume.
This is why the disclosure fight matters more than it looks. A tag is a crude instrument, but it is one of the very few signals a player gets before spending money, on a storefront where the alternative signals — review counts, wishlists, algorithmic recommendations — are all downstream of visibility the game has to earn first.
What this means if you just want to play something
For players, the practical takeaway is smaller than the discourse suggests. Check the disclosure box on a store page if it matters to you; it is genuinely informative now in a way it was not in 2024. Beyond that, the slop problem is a search problem, and the answer to a search problem is curation.
That is one honest advantage of a small, hand-built catalogue over an infinite one: every game in it was chosen. Our own games are hand-coded rather than generated — if you want something with real mechanical depth to sit against the argument above, Vault Heist is a stealth puzzle where you plan a route and then watch it execute, Hashi Bridges is a full logic-puzzle implementation with six difficulty packs, and Ink Rally is a draw-the-road physics game across 24 hand-designed levels. You can browse the whole puzzle category without a store page in sight.
For the broader shift in what browser games can now do technically, WebGPU Shipped Everywhere covers the rendering side of the same story.
What to watch next
- Whether Epic maintains its no-disclosure position as regulatory attention on AI labelling increases.
- Whether Valve refines the tag into categories — "AI-assisted" versus "AI-generated" would resolve much of the current objection.
- Whether disclosure rates keep climbing toward the universality Sweeney predicts, at which point the label really does stop discriminating.
- Whether storefront discovery changes structurally, which is the only thing that addresses the actual problem.