Was Any of This Made With AI? The Question Missing From Game Licence Due Diligence

Licensing a game means inheriting its production pipeline, including the parts made with AI. Here is what that does to the IP warranty you are relying on.

Every game licence agreement contains a clause where the licensor swears the work is theirs to license — that it does not infringe anyone's rights, and that if it does, they will cover you. That clause was written for a world where a human artist either drew the sprite or did not. It is now being applied to catalogues where nobody can say, asset by asset, how the art was made.

This is not a hypothetical risk creeping up on the industry. It is the current composition of the market, and the legal scaffolding around it moved twice in the last two years — once in Washington, once in Brussels, with the Brussels change taking effect eleven days ago.

📊 A Third of This Year's Releases Carry an AI Disclosure

Valve began requiring developers to disclose AI use in January 2024. That gives us three years of the only large, honest dataset on how deep generative AI has gone into commercial game production.

In an analysis of 53,597 Steam titles published in July 2026, Sulka Haro found roughly 9,400 games carrying an AI-generated content disclosure, and that about one in three of 2026's new releases now carries one. Reporting on the same trajectory puts the yearly figures at 10.9% of releases in 2024, 19.9% in 2025, and 30.8% so far in 2026.

You will also see a lower number quoted. SteamDB's count for the June 2026 Next Fest was 1,715 AI-disclosed titles out of roughly 8,700 participants — 19.71%. Both are correct and they measure different things: Next Fest is a demo showcase weighted toward unreleased projects, while the 30% figure covers full-year releases. Take the range, not either endpoint. The direction is not in dispute, and if the current slope holds, more than half of new Steam games carry the flag somewhere around 2027–2028.

Two caveats matter for a licensing buyer. First, the flag is binary in effect: a studio that ran one background texture through an upscaler and a studio that generated its entire art set both tick the box. It tells you AI was involved. It does not tell you the dose. Second, this is Steam. There is no equivalent disclosure requirement across HTML5 portals, casual game aggregators, or most APK marketplaces — so for the catalogues most of this industry actually trades in, the equivalent number is unmeasured, not zero.

⚖️ "We Own It" Is Doing Less Work Than It Used To

On 29 January 2025 the US Copyright Office published Part 2 of its report on copyright and artificial intelligence. Its conclusions, available in full at copyright.gov, are narrower than either the hype or the panic suggested:

• Human authorship remains a bedrock requirement. Material generated entirely by AI is not copyrightable.
• Prompting alone does not create authorship — even detailed, laborious prompting. The Office was explicit that effort spent writing prompts is not the same as creative control over the output.
• Where a work mixes human and AI-generated material, only the human contribution is protectable. It is assessed case by case.
• Using AI as a tool inside a human creative process — ideation, editing, cleanup — does not poison the whole work.
• Registrations must disclose more-than-de-minimis AI-generated material. The Office has already registered hundreds of works on that basis.

Read that as a licensee rather than as a lawyer. The risk is not that you get sued for using AI art. The risk is quieter and more corrosive: you may be paying for exclusivity over something that cannot be exclusive. If a meaningful share of a title's art is unprotectable, then an exclusive licence over that title does not stop a competitor from producing something visually indistinguishable, and you have no infringement claim to bring. You bought a fence with a gap in it.

That is a pricing question before it is a legal one. Exclusivity carries a premium — see how the tiers differ on our licence terms and pricing page — and the premium is only rational if the exclusivity is enforceable.

🛡️ The Indemnity Chain Breaks in the Middle

Here is the reassurance a licensor will offer: their AI tooling vendor indemnifies them, so you are covered downstream. Check that chain link by link, because it is weaker than it sounds.

The major vendors — Microsoft, OpenAI, Google, Adobe, AWS and Anthropic among them — do offer copyright indemnities. Law-firm reviews of the actual terms, including Proskauer's analysis of OpenAI's Copyright Shield and Kemp IT Law's survey of provider indemnities, consistently find the same structure: broad promise, dense carve-outs. Recurring conditions include:

• Paid, enterprise tiers only. OpenAI's Copyright Shield covers ChatGPT Enterprise and the API, not Free or Plus. Google's applies to generally available paid services, excluding previews and free tiers. A contractor working on a personal subscription is outside the perimeter.
• Guardrails must stay on. Disable, circumvent or ignore the vendor's filters and the indemnity typically evaporates. Prompting around a content filter to get a particular look is exactly the behaviour these clauses exclude.
• Known infringement is excluded. If the user had reason to believe the output infringed, cover is void.
• Copyright only. Adobe's Firefly enterprise indemnity, for instance, addresses copyright — not trademark, not rights of publicity. For anyone working near recognisable characters or real people, those are the two exposures that actually bite.

Now stack the practical reality on top: these indemnities run from the vendor to their customer. They do not automatically travel to the customer's licensee, or to that licensee's telecom partner, or to the portal operator three contracts down. Each hop needs its own warranty. In most game licence agreements, that chain is never assembled — it is simply assumed.

🇪🇺 What Changed on 2 August 2026

The EU AI Act's Article 50 transparency obligations became applicable on 2 August 2026. Per Cooley's August 2026 note on the change, the obligations break into four duties: telling users they are interacting with an AI system; embedding machine-readable markings in synthetic audio, image, video and text; notifying people subject to biometric or emotion-recognition systems; and disclosing deepfakes and AI-generated public-interest text where there has been no substantive human editorial review. Non-compliance carries fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. Generative systems already on the market before that date have until 2 December 2026 to meet the marking requirement, and content generated before 2 August 2026 does not need retroactive labelling.

Be precise about who this catches, because the scaremongering version is wrong. A studio that used Midjourney to make a menu background in 2025 is not, by that fact alone, a "provider" or "deployer" of an AI system with Article 50 duties. The obligations attach to AI systems and their operators, and the labelling duty is framed around deepfakes and public-interest text — not around every asset that passed through a model.

Where it does land squarely on this industry is live generation inside the product. Ship a game with an AI companion character, an AI-generated level, a chat NPC, or user-facing image generation, and you are running an AI system in front of EU users. Then the disclosure and marking duties are yours, and they are yours regardless of where your studio sits — the Act reaches anyone placing AI on the EU market or whose outputs are used there. If your catalogue includes titles with live generative features and you distribute into Europe, that is a compliance item with a date on it, not a roadmap item.

📱 The Store Layer Is a Separate Obligation Again

Platform rules are not the law, and they do not track it. They are a third set of duties.

Google Play's AI-Generated Content policy, documented in Play Console Help, requires that apps generating content with AI include in-app reporting or flagging so users can report offensive output without leaving the app, and holds the developer responsible for what their app generates — covering text, voice and image prompt inputs. Google tightened the labelling and moderation language in January 2025 and added specific enforcement for apps aimed at younger audiences.

So a single APK title can simultaneously owe: a Steam-style disclosure it does not owe at all (no such rule on Play), a Play-specific in-app reporting mechanism it does owe, and an EU marking duty that depends on whether generation happens at runtime or happened in the studio two years ago. These are not one obligation with three names. Anyone assembling a multi-territory catalogue needs to answer them separately per title.

🧸 Branded IP Is Where the Cost Lands

The commercial pressure point is not indie portals. It is brand work.

Disney, Universal and Warner Bros. sued Midjourney in 2025 over the generation of their characters, and the case moved into discovery through 2026 — with Midjourney pushing to compel the studios to disclose their own internal AI use, a magistrate limiting that to consumer-facing content, and Midjourney appealing, as reported by Variety and TechCrunch in July 2026. Midjourney's defence rests on fair use. Nothing is decided.

Separately, the UK High Court ruled on 4 November 2025 in Getty Images v Stability AI that model weights are not a "copy" of the training images under UK copyright law, rejecting Getty's secondary infringement claim and leaving only a narrow trademark finding. Getty's US case is on its own track.

Two live cases, two jurisdictions, no settled answer. Which is exactly why brand approval pipelines have hardened. A rights holder currently in litigation over AI-generated depictions of its characters is not going to approve a licensed game whose asset provenance is undocumented. Expect the AI-use question to appear in the approval questionnaire, expect it to be asked per asset rather than per project, and expect "we're not sure" to be treated as a no. If you build character-led and cartoon-style games for brand clients, the ability to answer that question with paperwork is now a qualification to bid.

🚫 Five Ways Buyers Get This Wrong

• Accepting the boilerplate warranty and stopping there. "Licensor warrants it owns all rights" was drafted before any of this. A warranty is only worth the licensor's ability to pay a claim — and for a small studio, that is often nothing.
• Treating an AI vendor's indemnity as transitive. It runs to the vendor's own customer. Get an express warranty and indemnity in your agreement, or you have nothing to point at.
• Buying reskins without asking what was reskinned. Reskin and source-code deals are where provenance disappears fastest, because the asset trail runs through two or three parties before it reaches you. If you are taking source code, you can actually audit the asset folder — so do.
• Assuming disclosure is the developer's problem. On Play, the publisher of record carries policy compliance. Under the AI Act, deployer duties attach to whoever runs the system in front of users. That is frequently you, not the studio you licensed from.
• Paying an exclusivity premium on unprotectable art. If you cannot enforce it, you are buying a promise that the licensor will not resell — which is a contractual right against one counterparty, not exclusivity in the market. Price it as the former.

📋 Six Questions to Add to Licence Due Diligence

1. Was generative AI used in producing any asset in this title — art, audio, voice, code, localisation? Ask it as a yes/no/don't-know, and treat don't-know as a yes.
2. If yes, which tools, on which account tier? This determines whether any vendor indemnity exists at all.
3. Does the game generate content at runtime? This is the question that decides your EU AI Act position, and it is entirely separate from question 1.
4. What is warranted, and by whom, with what cap? An uncapped indemnity from a two-person studio and a capped one from a company with a balance sheet are different products.
5. Can the licensor produce provenance documentation on request? If a brand client asks in nine months, you need an answer in days, not a forwarded email chain.
6. Who handles store-level disclosure, and in which territories? Write it into the agreement rather than discovering it at submission.

None of this requires a clean-room, no-AI-anywhere policy. Plenty of AI use is unremarkable — upscaling, background removal, localisation drafts reviewed by humans, placeholder art replaced before ship. The point is to know which category a title falls into before you wire the money, not after a brand's legal team asks.

🎮 Where a Licensed Catalogue Helps

Forestry Games has licensed games since 2017 and maintains a catalogue of 1,049 titles across HTML5 and Android, develops HTML5 games in-house, and works with branded IP. The practical relevance to everything above is narrow but real: a licensor that developed titles internally can answer provenance questions about those titles directly, and a single agreement covering many titles means one warranty negotiation rather than forty.

🧭 What to Do This Quarter

Start with what you already shipped. Take your live catalogue, and sort it into three buckets: titles you built, titles you licensed with a written warranty, and titles you licensed on a purchase order and a handshake. The third bucket is your exposure, and it is usually larger than operators expect.

Then do one thing that costs nothing: add the runtime-generation question to your intake form. Titles with live generative features are a different compliance object from titles with AI-assisted art, they carry a hard EU date, and most catalogue operators currently cannot tell you which of their games are which. Being able to answer that in an afternoon — rather than in a fire drill when a carrier or a brand asks — is the entire return on this work.

The warranty clause is not going to update itself. Someone on your side has to notice that the sentence has quietly stopped meaning what it used to.