Two 2026 Surveys Measured AI in Game Studios and Landed 54 Points Apart

AI cut the cost of building a casual game. It did not cut the cost of testing, clearing or distributing one — and that is where the money in a catalogue goes. The two big 2026 adoption surveys disagree by more than fifty points, and the gap between them is more useful than either number on its own.

If you buy games rather than make them, the AI question arrives at your desk in a strange shape. Nobody is asking you whether to adopt anything. You are being asked to price, licence and publish titles that somebody else built, using a pipeline you cannot see, in a year when the industry cannot agree on how common that pipeline even is.

So start with the disagreement. It is the most instructive data point available.

📊 Two Surveys, One Industry, a 54-Point Gap

GDC published its 2026 State of the Game Industry report on 29 January 2026, its fourteenth annual survey, drawing on more than 2,300 game industry professionals. It found that 36% use generative AI tools in their own work.

Google Cloud published a survey run by The Harris Poll, fielded between 20 June and 9 July 2025 among 615 game developers in the United States, South Korea, Finland, Norway and Sweden. It reported that 90% were already integrating AI into their workflows, and 97% believed generative AI was reshaping the industry.

Thirty-six against ninety. Both surveys are real, both name their methodology, and neither is lying. Four things explain most of the gap:

• The question is not the same. "Do you use these tools in your work" and "is AI integrated into your workflow" are different questions. The second is satisfied by a colleague, a middleware vendor, or an engine feature you did not choose.
• One survey has a seller attached. Google Cloud sells the infrastructure that AI workloads run on. That does not invalidate the fieldwork, but sponsored research reliably lands on the optimistic side of a range.
• The samples are different populations. 615 developers in five countries, three of them Nordic, against 2,300+ self-selected respondents from GDC's audience, which skews toward studio production roles.
• Role mix moves the number hard. GDC found 30% usage among people at game studios and 58% among those at publishing companies, support teams and marketing or PR firms. Where you draw the industry boundary decides your headline.

The honest reading is a range, not a figure: somewhere between a third and the large majority of the companies that sell you games have generative AI somewhere in their process, and no survey can tell you where it sits in the specific title you are about to licence. That makes it a due-diligence question, not a market statistic.

One more number from the GDC report matters for anyone negotiating with studios: 52% of respondents think generative AI is having a negative impact on the industry, up from 30% the year before and 18% the year before that. Positive sentiment fell to 7%, from 13%. Whatever adoption looks like, enthusiasm is going the other way — sharply.

🧪 Adoption Clusters at the Cheap End of the Pipeline

Both surveys agree about something more useful than the headline: where the tools are actually being used.

GDC's breakdown of what its 36% do with the tools: 81% research or brainstorming, 47% daily tasks and code assistance, 35% prototyping. Google Cloud's task list, from the optimistic end of the range, puts playtesting and balancing at 47%, localisation and translation at 45%, code generation and scripting at 44%, content optimisation at 44%, procedural world generation at 37%, and dynamic level design, animation and dialogue writing at 36%.

Read both lists and the same shape appears. The heaviest use is at the front of a project — ideas, drafts, scaffolding, a first playable. The lists thin out as you move toward the things that decide whether a title survives contact with players: final art passes, audio, performance work on real devices, live balancing against real retention curves.

That matters because the front of the pipeline was never the expensive part of a casual title. Getting to a prototype was always the fastest week of the project. Getting from a prototype to something a portal will accept, a store will approve and a player will return to on day seven was always the slow part, and it still is.

💸 Production Was Never the Line That Decided Whether a Catalogue Earned

Here is the part that gets skipped in the "games are cheaper to make now" conversation. If you operate a portal, a telecom games service or a branded campaign, build cost sits in someone else's P&L. Your costs are licence fees, integration, hosting and egress, moderation, ad operations, compliance work, staff, and whatever you spend to get a human being in front of a game. Generative AI moves almost none of those.

It does not renegotiate your revenue share with an ad network. It does not reduce the cost of a support rota. It does not lower a carrier's integration timeline, or make a store's review queue shorter, or buy traffic. A catalogue that costs 30% less to produce and still has no distribution plan is a catalogue that loses money 30% more efficiently.

The same logic applies inside a studio. A cheaper first playable does not shrink store review, device QA, age rating submissions, localisation QA, or the cost of persuading anyone to play the thing. It shrinks the week before all of that.

📈 The Supply Side Answered Immediately. The Demand Side Didn't.

The PC market gives the clearest public picture of what happens when production gets cheaper faster than distribution gets wider. Inven Global reported on 23 July 2026, using SteamDB release counts and Alinea Analytics revenue estimates, that roughly 21,000 games launched on Steam in 2025, against about 18,000 in 2024 and about 14,000 in 2023. Steam's estimated revenue for the first half of 2026 hit $11.1 billion, up 14.5% year on year — and games released that year accounted for 21% of it. The other 79% went to back catalogue.

Two caveats before anyone builds a strategy on that. Steam is premium PC, not web or Android casual, and the economics differ in almost every respect. And the release surge began well before current AI tooling was widely available, so this is not a clean causal chart of "AI made more games."

The mechanism still generalises. When the cost of producing content falls faster than the cost of putting it in front of players, the scarce asset stops being content and becomes distribution. Anyone who has tried to get a new title placed on a portal's front page, or into an operator's bundle, already knew that. Cheaper production makes it more true, not less.

🧾 The Obligations That Do Not Compress

Four costs sit downstream of the build, and none of them respond to a faster asset pipeline.

Device testing

A generated sprite sheet does not know what a four-year-old Android phone does with it. Frame budgets, memory ceilings, texture formats and first-playable-frame timings are measured on hardware, by a person, after the build exists. That work is the same length it was in 2023.

Store responsibility, which lands on the publishing account

Google Play's AI-Generated Content policy makes the developer responsible for ensuring their generative AI apps do not produce offensive content, on top of every other Developer Program policy. If you licence an Android title and publish it under your own account, that responsibility is yours — not the licensor's — regardless of who wrote the code. Worth reading closely before you take on a catalogue of Android games whose provenance you have not asked about.

Provenance and the warranty chain

"We own it" is a shorter sentence than it used to be. If generated assets are in a title, the question is whether the licensor's warranty and indemnity actually name them, or go quiet at exactly that point. This is sharpest in source-code deals, where you inherit the asset tree and the obligation to keep the thing shippable.

Localisation, which is not translation

Google Cloud's own respondents put localisation and translation among the top AI use cases at 45%. Machine output has genuinely improved. It still does not choose your markets, handle right-to-left layout, localise a payment flow, or survive a search engine's view of bulk machine-translated pages. The cheap part got cheaper; the part that decides whether a market works did not.

📋 Six Questions to Put to a Licensor

1. Which assets in this title were generated, and with which tools?
2. Do you hold commercial output rights under those tools' terms, and can you evidence it?
3. Does your warranty and indemnity name generated assets specifically, or is it silent on them?
4. Has this title ever been live anywhere? What were the real session and retention numbers?
5. What device matrix was it tested on, and what is the oldest device it passes on?
6. If a store pulls it for a generated-content reason, who fixes it, on whose clock, and at whose cost?

Question four is the one that separates a catalogue from a folder of files. A title with a play history has been tested by the only instrument that matters. A title generated last month has not, however good the screenshots look.

🚫 Five Ways Operators Misread the AI Cost Story

• Treating a cheaper build as a cheaper licence. Licence price tracks what a title earns and what rights you are buying, not what it cost to make. Arguing production cost in a negotiation signals that you are pricing the wrong thing.
• Buying volume because volume got cheap. Play concentrates in a small share of any catalogue. Doubling the shelf does not double the sessions; it doubles the maintenance.
• Applying a survey number to a specific supplier. Neither 36% nor 90% tells you anything about the studio in front of you. Ask them.
• Expecting AI localisation to substitute for market entry. Translated strings are the cheapest item on a market-entry list and never the one that fails.
• Skipping provenance because the price was small. The cost of an IP problem has no relationship to what you paid for the title that caused it.

🎮 Where a Licensed Catalogue Fits

Forestry Games has licensed games since 2017 and maintains a catalogue of 1,049 titles across HTML5 and Android, with titles also published on Google Play and the Apple App Store. It develops HTML5 games in-house, works with branded IP, and holds brand partnerships including Disney, Nickelodeon, Cartoon Network and Warner Bros.

The relevance to this topic is narrow and worth stating plainly: a catalogue that has been operated and distributed over years carries evidence a freshly generated one cannot. Live history, device coverage, and a single counterparty to hold the warranty. If you are sizing a portal or a bundle, the catalogue is the place to start comparing on those terms rather than on unit price.

🧭 What to Do This Quarter

Add the six questions above to your licence due diligence template — not as a separate AI review, but as standard fields alongside territory, term and platform rights. Then go back through the titles you licensed in the last eighteen months and check whether your existing agreements say anything at all about generated assets. Most contracts signed before 2025 do not, and silence in a contract is not the same as protection.

And when a supplier tells you their production got 40% cheaper, ask what happened to their testing budget. The answer tells you which half of the pipeline they optimised, and which half you are about to inherit.