Home › AI in Game Development Statistics

AI in Game Development Statistics 2026: Adoption vs Proof

Of every hundred games professionals GDC surveyed for its 2026 report, fewer than two are pointing generative AI at a feature the player will actually interact with. That number reframes the whole topic: the industry has adopted AI widely, cites its efficiency constantly, and has published nothing that would let anyone verify it. Among the 36% of games professionals who use generative AI, 81% use it for research and brainstorming and 5% for a player-facing feature (GDC 2026 State of the Game Industry, 2,300+ respondents) — which puts player-facing use at about 1.8% of everyone GDC surveyed, on our arithmetic across GDC’s own two figures. Asset generation is commoner at 19% of AI users, so AI does reach the player: it reaches them as a texture, not as a system. Meanwhile the productivity gain executives invoke when they restructure has never been measured in a single published study of game development. This page follows the technology from the developer’s desk to the store page to the bank, and shows exactly where the evidence chain snaps. Every figure is attributed to a named, dated source, and every number we computed ourselves is labelled as ours.

Does an AI adoption percentage tell you anything about games?

Almost nothing. Every one of those percentages is a statement about workflows, not about games. “90% of game developers use AI” does not mean 90% of games contain AI. It means nine in ten people surveyed have, at some point, used a tool. The survey question and the thing readers infer from it are two different claims, and the distance between them is where this entire topic lives.

The second error follows from the first. Because adoption looks enormous, the productivity gain gets treated as settled — a premise you can build a restructuring on rather than a hypothesis somebody should test. It is not settled. It is not even measured. Our game development industry statistics page documents what happened to employment and output in 2023–2025; AI is the causal story most often offered to explain it. This page asks whether that story has evidence behind it, and the honest answer is that it has a great deal of adoption data and almost no outcome data.

So the frame here is not pro-AI or anti-AI. It is: what has actually been counted, by whom, with what question, and what has conspicuously not been counted at all.

How many game developers actually use generative AI?

36% personally use generative AI tools, per the GDC 2026 State of the Game Industry survey of more than 2,300 games professionals — and that figure has not moved since the 2025 edition. Every larger number you have seen comes from a different question asked of a different population.

GDC figures from the annual State of the Game Industry reports and their press releases; the 2026 report’s authors note that this year’s sample differs significantly from previous years and that they therefore ran fewer year-over-year comparisons than usual (as summarised by GameDev Reports), which limits strict trend reading. Google Cloud figure from the Harris Poll survey commissioned by Google Cloud, fielded 20 June – 9 July 2025 among 615 adults working in game development in the US, South Korea, Finland, Norway and Sweden; its published wording is “some form of AI,” not generative AI specifically, which is a materially wider net than GDC’s question. Unity figure from the 2025 Unity Gaming Report, drawn from a Cint survey of 300 developers (150 North America, 60 EMEA, 60 APAC, 30 South America) conducted in 2024 with a stated margin of error of ±5.66% at 95% confidence — the report is a 2025 publication of 2024 fieldwork, which matters when it is set against a survey fielded in late 2025. The 96% is the complement of the 4% who reported using no AI tool at all. Individual-level and studio-level questions are not interchangeable and are not compared as like for like anywhere on this page.

Within the GDC data the split by employer type is sharper than the headline. Use runs 30% at game studios against 58% at publishing companies, support teams and marketing or PR firms, and 58% among business professionals generally. Seniority tilts it further: 47% of upper management use the tools, 36% of studio directors, 29% of lower-level staff. Read that ordering carefully. The people furthest from the code and the art are the heaviest users, and the people running the company use it at roughly 1.6× the rate of the people they manage (16Best analysis: 47% ÷ 29%).

What do the developers who use it actually use it for?

Overwhelmingly for thinking and admin; when it does reach the player it arrives as an asset, almost never as a system. The GDC 2026 breakdown of use among AI users is the single most clarifying dataset in this topic, and it is almost never reported.

Make it concrete. Take a hundred games professionals as GDC sampled them. Thirty-six use generative AI at all. Twenty-nine of those hundred use it to brainstorm or research. Seventeen use it for email and meeting prep. Seventeen use it for code assistance. Thirteen use it to prototype. Seven make assets with it — and those assets do reach players. Four run procedural generation. And fewer than two build it into the game as a feature the player interacts with (16Best analysis: each use-case share multiplied by the 36% adoption base; shares of GDC’s sample, not of the industry). Read the gap between the seven and the two carefully, because it is the distinction most coverage collapses: AI is in plenty of shipped games as content, and in almost none of them as behaviour. The most-cited tool is ChatGPT at 74% of those using large language models, then Google Gemini at 37%, Microsoft Copilot at 22%, Midjourney at 17%, Adobe Generative Fill at 13% and Sora at 8%. That is a list of general-purpose assistants with a few media tools attached — not a games-production stack.

One counterweight, because the picture is not uniform: 21% of respondents report using an internal, company-built AI tool, rising to 30% at AAA studios. Large publishers are building bespoke systems that a general survey question about “AI tools” will capture poorly. If the concentration in tooling is going to break, that is where it breaks first.

Does any of that reach the shipped game?

Yes, and at speed — 30.8% of games released on Steam in 2026 to July carry an AI disclosure, up from 19.9% in 2025 and 10.9% in 2024, on a trajectory Haro projects will cross half of all releases in 2027 or 2028. But the disclosure measures a different thing from the surveys, and conflating the two produces most of the bad analysis in this field.

Shares from Sulka Haro’s analysis of 53,597 Steam releases, July 2023 to July 2026, published 13 July 2026 — a census of the store, not a sample. 2026 covers January to July only and will move. The 2023 figure is near zero because Valve only added the disclosure field to its content survey on 10 January 2024; Haro is explicit that contemporaneous disclosure in 2023 was structurally impossible, so the ~1% is entirely retroactive flagging, not early adoption. Percentage-point and relative changes are 16Best analysis. Our Steam statistics page tracks the total release counts these shares sit on top of.

Push one level deeper and the concentration holds inside the shipped games too. Totally Human Media examined the use cases declared by AI-disclosed titles grossing over $1M and found in-game art at 49%, voiceover and dialogue at 27%, pre-production and concept at 13%, marketing at 11%, translation and localisation at 9% — and runtime generative systems under 5%. Even where AI is in the box, it is almost never generating anything while you play. It made the picture on the wall before the disc was pressed.

Haro’s reading of the disclosure text says the same thing from the other end of the market. He compared the 138 AI-flagged titles that succeeded (defined as 1,000+ reviews) against a random 400 that flopped (under 50 reviews after at least nine months on sale). Among the flops, 72% mention AI art, images or textures; among the successes, 57%. The successes are more multimodal: three times more likely to disclose AI voice (24% vs 8%) and localisation (18% vs 6%), and their disclosure text runs to a median 22 words against 13. The median failed AI game is an art-asset play and nothing else, described in one line.

Live, runtime generative gameplay does exist, and it is countable precisely because it is rare. NVIDIA’s ACE platform — its packaging of small language models, speech and animation for game characters — launched at CES on 6 January 2025 with seven titles named in NVIDIA’s own announcement: PUBG: Battlegrounds, inZOI, NARAKA: Bladepoint Mobile PC Version, MIR5, Dead Meat, AI People and ZooPunk. Seven, itemised in a vendor launch post, against the roughly 53,600 games Steam took in the three years to July 2026. Coverage of the same launch named a couple more partner titles, so treat seven as the vendor’s own count rather than a census of runtime AI — but no version of that list runs to three figures, and that is the whole argument in one line.

Do AI-disclosed games sell?

They under-index on every measure that matters — by mid-2026 AI-flagged titles were about a third of new Steam releases but only 10–27% of estimated sales, and per game they hit Haro’s modest-success tier at about 55% of the non-AI rate. More games, proportionally less money.

All figures from Sulka Haro’s Steam census (53,597 titles, July 2023 – July 2026, published 13 July 2026); review scores computed on the ~26,000 titles with at least ten reviews. Sales are Boxleiter-method estimates derived from review counts and list prices (units ≈ reviews × 30), gross of platform fees, discounts and tax, and are order-of-magnitude accurate at best. Three different release shares appear on this page and they are not interchangeable: 30.8% is calendar 2026 to July, ~33% is the recent monthly run-rate Haro himself sets against the sales share, and 37% is the monthly release share in the specific mid-2026 months the top-100 comparison covers. The top-100 figures are likewise two scopes: ~9% pooled across the whole three-year window, ~11% in those recent months.

Totally Human Media’s 6 November 2025 count put 10,258 AI-disclosed titles — then about 8% of the Steam library — at an estimated $660 million in combined gross. Mind the scope on that figure: it is a Boxleiter estimate built from lifetime review counts, so it is what those titles have grossed to date, not what they earned in a year. At least one major outlet reported it as “$660 million this year,” which the method cannot support. The same analyst had counted 7,818 on 13 July 2025, so the disclosed population grew 31% between the two snapshots, under four months apart (16Best analysis of the two Totally Human counts). Across the 10,258, the mean is roughly $64,300 a title — and the mean describes nobody: 12 titles reached eight figures, 33 reached seven, 170 landed in six, 487 in five, and 9,556 fell under $10,000 (16Best analysis of their published counts, which sum exactly to 10,258). The distribution is the same violent power law our indie game statistics page finds across Steam generally, where the median 2025 release grossed $249 on Gamalytic’s October cut — $318 on its full-year read — against a $358,900 mean.

And that curve is not an AI artefact. On Haro’s census the top 1% of all Steam titles capture about 94% of estimated revenue, the top ten games alone about 62%, the median game with any reviews at all makes roughly $300, and about 28% of releases get zero reviews — no measurable sales whatsoever. AI did not change the shape of that curve. It added mass to the bottom of it.

The mechanism is visible in who is arriving. First-time developers shipping a non-AI debut have held flat — slightly declining — at roughly 950 a month for two years, while AI-debuting newcomers rose from about 140 to 440 a month. Effectively all net growth in new Steam developers is AI-driven, and 60–90% of the growth in monthly releases is AI-flagged games. The tool is manufacturing new entrants, not new revenue.

Where is the evidence that AI makes game development faster?

There isn’t any — not in game development. No study we can find measures AI-attributable output, cost or schedule change in a game studio using a design that could support the claim. This is the finding, and it deserves to be stated as flatly as the adoption figures are.

We searched for it directly: peer-reviewed work, industry studies, published post-mortems with before-and-after production metrics, publisher disclosures tying AI to a measured cost or schedule delta. What exists is a well-populated shelf of adoption surveys and an empty shelf where the outcome measurement should be. Here is what is actually cited when someone asserts the gain, with the scope of each stated plainly.

GitHub’s figure is from its own 2022 controlled experiment on a single standardised JavaScript task and was never intended as an estimate of whole-pipeline productivity. METR’s July 2025 result is the strongest available randomised evidence on AI and experienced-developer productivity, but its population is open-source maintainers, not game studios, it has 16 participants, and METR itself now labels the post “out of date” — a snapshot of early-2025 tools rather than a standing claim about current ones. Its 24 February 2026 follow-up is the honest place to look next, and it is unusually candid: with 57 developers and 800+ tasks on later tools, the point estimates are still negative (−18% for the returning cohort, −4% for new recruits) but the confidence intervals cross zero, and METR says the run “gives us an unreliable signal” because 30–50% of participants admitted withholding tasks they did not want to attempt without AI — a selection effect that biases the measurement downward. METR is redesigning the experiment rather than defending the number. Read the pair as one fact: across two experiments and a year of follow-up, the best-designed measurement in adjacent software still cannot pin down the sign of the effect — which is a strong reason not to borrow either result for game development. Lost Lore’s case was published on Game World Observer on 27 January 2023 and is a self-reported estimate with no control. It is also internally unreconciled, and worth pausing on because it is the single most-recycled dollar figure in this debate: the article’s headline claims $70,000 saved, but the only arithmetic it shows is a founder’s estimate that 100 character concepts would have cost about $50,000 the traditional way and cost about $10,000 with Midjourney. That is a $40,000 gap, self-estimated against a counterfactual that was never run, and the remaining $30,000 of the headline is never itemised (16Best analysis of the published figures). Lost Lore now trades as an AI-driven game development shop for hire, and its own site runs both that $70,000 saving and GitHub’s 55% figure as sales copy — a statistic completing the round trip from evidence to marketing collateral. None of the studies in this table measures game development output at studio scale. Where a claim cannot be scoped to game development, this page says so rather than borrowing it.

If nobody has measured it, why does everyone act as if it is proven?

Because the claim is doing organisational work that does not require it to be true. Three named, dated 2025–26 examples show what an unmeasured premise looks like when it is used as a planning assumption.

Playtika, 14 January 2026: disclosed in an SEC filing that it would cut about 500 employees, 15% of a workforce of more than 3,000, as it moved from large teams to leaner ones relying on AI and automation. Chief executive Robert Antokol put it plainly: the company is “moving away from headcount-heavy operations to streamlined teams powered by AI and automation.” Playtika expects to book $12–15 million in restructuring charges — roughly $24,000 to $30,000 per departing employee (16Best analysis, dividing the disclosed charge range by the disclosed ~500 headcount; the charge covers severance, notice-period pay, benefits and related costs, not compensation alone). It disclosed no productivity figure.

Krafton, 23 October 2025: announced a reorganisation into an “AI-first company,” committing 100 billion Korean won — about $70 million at the roughly 1,430-won-to-the-dollar rate at announcement, and reported as anything from $70m to $72m depending on the outlet’s conversion date — to a high-performance GPU cluster to underpin agentic AI, plus a separate 30 billion won a year from 2026 (about $21 million at the same rate) to give employees access to AI tools. Those are a one-off capital commitment and an annual operating one; adding them into a single headline figure, as several write-ups did, mixes two accounting bases. Weeks later, on 12 November, Krafton opened a voluntary resignation programme to all employees with severance scaling from 6 months’ salary at one year of service to 36 months above eleven years — days after reporting record quarterly results. Hiring was frozen company-wide except for original-IP and AI roles.

EA, October 2025: Business Insider reporting, covered by VGC, described management urging the company’s roughly 15,000 employees to use AI for “just about everything” — coding, concept art, even advice on sensitive management questions — with staff expected to complete AI training courses and use the tools daily. Developers described correcting flawed code the tools produced. Set that against GDC’s 2024 baseline, when just over half — 51% — said their company had instituted any policy on generative AI, only 2% said the tools were mandated in their workplace and 12% said they were not allowed to use any of them (as reported in GamesBeat’s coverage of that survey; a parallel write-up gives 32% whose company had no policy at all, so the residual is respondents who did not know). By the 2026 survey, 78% of companies had at least a basic AI policy. Formal policy arrived fast. A measurement did not arrive with it.

None of that makes the efficiency claim false. It makes it unverified, which is a different and more useful thing to say. And it is worth being precise about attribution: the tracked layoff totals on our game development page — about 14,600 in 2024, about 9,053 in 2025 — predate most of these AI-first announcements and are overwhelmingly attributed to pandemic-era over-hiring, not automation. Playtika is a rare case of a company explicitly naming AI in the reasoning. Treating the whole layoff wave as an AI story would be exactly the unevidenced leap this page is documenting.

Have budgets bent the way an efficiency gain would predict?

Not on any published series — and this is the cleanest falsifiable test available, which is precisely why nobody runs it. If AI were delivering large, general productivity gains in production, the cost of making a game should decelerate.

Our cost-to-make analysis puts AAA budget inflation near 24.5% a year since the PS1 era, which doubles per-title spend roughly every 3.2 years (ln 2 ÷ ln 1.245) (16Best analysis). Apply even the most generous vendor number to that trend — GitHub’s 55% task-level speed-up — across a whole pipeline, and you would expect a visible inflection within two or three years of mass adoption. There is no published, continuous, methodologically consistent series of AAA development budgets to check it against. Budgets are disclosed sporadically, usually through litigation or leaks, and usually as unaudited estimates.

Why do developers and their employers report opposite things?

Because they are being asked by different people with different stakes — and the two most-quoted sentiment figures, 79% positive and 7% positive, come from surveys a year and one commercial interest apart, asking about two different objects.

GDC State of the Game Industry, 2024–2026 editions; the 2024 negative figure is as restated in GDC’s own 2026 release (“up from 30% last year and 18% the year prior”). GDC explicitly notes that its 2026 sampling methodology differed from prior years, so these are best read as three snapshots of a moving population rather than a clean panel. Percentage-point and relative changes are 16Best analysis computed within the GDC series only. The 2024 edition was fielded in October 2023.

The direction is unambiguous even allowing for the sampling caveat: negative sentiment roughly tripled across three editions while personal use moved five points and then stopped. And the discipline split explains why. Unfavourable views run 64% among visual and technical artists, 63% in game design and narrative, 59% in programming — against 19% favourable among executives and among business operations and services. The people who use the tools least are the people whose work the tools most plausibly replace, and the people who use them most sit furthest from the craft.

Now the conflict worth resolving. Unity’s 2025 Gaming Report, published 17 March 2025, found 79% of 300 surveyed developers feeling positive about AI tools (31% extremely, 48% somewhat, 17% neutral, 5% apprehensive) and 96% of studios using AI in select workflows. The GDC 2026 survey of 2,300+ found 7% positive. That is an 11.3× gap on the closest comparable measure (16Best analysis). Both cannot describe the same industry — but they need not, and here is the reconciliation: Unity asked about AI tools via a Cint panel of 300 developers fielded in 2024, in a report published by a company that sells AI tooling; GDC asked about generative AI’s impact on the industry to 2,300+ conference-affiliated professionals, publishing ten months after Unity and fielding later still, over a stretch in which negativity in its own series rose 22 points. Different object (tools versus industry impact), different population size by nearly 8×, different sponsor incentive, different moment. Comparing them as a single ratio is a scope error — which is exactly why we state the 11.3× as the size of the mismatch, not as a finding about developer opinion. Unity’s own data quietly concedes the point: no single AI workflow in its survey exceeded 50% usage, with automated playtesting, character animation and code assistance all just above 40%, generative artwork and level generation at about 35%, and adaptive difficulty at 30% — enthusiasm broad, application shallow, which is the same shape the GDC use-case data shows.

What rules now constrain AI use in games?

Three hard constraints landed between January 2024 and January 2026: platform disclosure, a union contract with consent and payment floors, and unresolved training-data litigation. None is a ban; all shape what studios can ship.

Valve policy dates from its 10 January 2024 announcement and the mid-January 2026 rewrite as reported by VGC, Game Developer and others. SAG-AFTRA ratification vote and terms from SAG-AFTRA’s own announcements and Deadline’s coverage; note the two dates are not interchangeable — the strike stopped on 11 June 2025 when the union suspended it after reaching a tentative agreement, and the 9 July ratification is what made the contract binding. the 7.5× scale minimum applies to Real Time Generation as defined in the agreement summary and governs union performers under that contract, not AI use generally. Litigation dates and consolidation from court records and reporting; these are filings and procedural orders, not rulings — no merits decision had issued as of July 2026, so nothing here is settled law.

The 7.5× scale minimum for Real Time Generation is the most underrated number in this whole topic. It puts a hard, contractual price on exactly the use case — runtime generative dialogue — that is already the rarest in shipped games. Under 5% of $1M+ AI-disclosed titles declare runtime systems, and the union contract now attaches a multiplier to the voice side of doing so with union performers. The economic gravity and the contractual gravity point the same way.

Why do the AI adoption figures disagree so wildly?

Eight definitional forks, and each one can move a headline by tens of percentage points. If you take one thing from this page to argue with, take this section.

• Question wording is the single largest driver. “Do you use generative AI?” “Has your employer implemented it?” “Is AI in your workflow?” “Does your studio use AI tools in select workflows?” produce 36%, 52%, 90% and 96% respectively across fieldwork running from 2024 to late 2025. GDC’s own 2026 edition separates 36% actively using from 52% using AI in some capacity (per GameDev Reports’ summary of the report) — two numbers, one survey, one population, and a 16-point fork created purely by the adverb. Watch for a coincidence that has caused real errors in coverage: GDC’s 2025 edition also produced a 52%, but that one is the share of respondents whose employer had implemented generative AI. Same number, different year, different question. They are not the same statistic and neither is a restatement of the other.
• Use is not output, and no instrument on this page bridges them. Every figure here answers “is the tool present?” None answers “did the work get done faster, cheaper or better?” Those are different questions requiring different designs — a controlled comparison, or at minimum a before-and-after production metric with the confounders named. Reading an adoption percentage as evidence of a productivity gain is the single most common error in this field, and it is the error this page exists to interrupt.
• “AI” conflates two unrelated technologies. Games have used AI for decades — pathfinding, finite state machines, behaviour trees, procedural generation. A question about “AI in your studio” is answered yes by anyone who has written an A* implementation. Only questions that say generative AI are measuring the thing under debate, and not all of them do.
• Individual versus studio adoption. One person at a 400-person studio using ChatGPT makes the studio an adopter and leaves individual adoption at 0.25%. Both numbers are true; they are not the same number, and they are routinely quoted interchangeably.
• Self-selected panels. GDC surveys people connected to a conference — 54% of its 2026 respondents live in the United States. Unity surveys a commercial panel of its own ecosystem. Neither is a random sample of the world’s game developers, and no such sample exists.
• Vendor sponsorship. Unity and Google Cloud both sell AI tooling to the people they surveyed, and both report adoption figures at the top of the range. That is not an accusation of dishonesty; it is a reason their instruments and framing differ from a trade body’s. Follow the funding across this whole page and the pattern is hard to unsee: the two highest adoption figures come from an engine vendor and a cloud vendor, the runtime-AI title list comes from a GPU vendor, the only in-industry cost saving comes from a studio that now sells AI services, and the most-cited speed-up comes from the company that ships the code assistant. The two lowest-enthusiasm datasets — GDC’s survey and an independent analyst’s Steam census — are the two with nothing to sell. That does not make the vendor numbers false. It does mean the range you see is not a neutral sample of reality; it is a sample of who could afford to run a survey. It generalises well beyond games: adoption statistics in any field are mostly produced by the people selling the thing being adopted.
• Disclosure captures declaration, not use. Steam’s 30.8% measures what developers chose to declare, under rules that changed twice, and that now exclude pure development-efficiency tooling. Haro dated each flag with the Wayback Machine and found 22% of significant-revenue AI games added theirs after release, clustering tightly around the January 2026 policy rewrite — a compliance sweep. Historical shares therefore get revised upward over time by construction.
• “Boxleiter” is not one method. Both Steam revenue estimates on this page carry the same label and different arithmetic: Haro uses units ≈ reviews × 30, Totally Human Media uses reviews × price × 35. Same name, roughly a sixth of difference in the units assumption before anything else. Never present two Boxleiter figures as though they were measured the same way.
• Catalogue and corpus definitions differ, and rounded shares cannot be inverted. Totally Human counted disclosed titles across the whole live library — 7,818 in July 2025, 10,258 by November 2025; Haro counts only games released between July 2023 and July 2026. Different corpora, not contradictory figures. The trap is the denominator: Totally Human’s July 2025 note gives its own base as just under 115,000 Steam games, consistent with the 117,000–125,000 range on our Steam statistics page, but its November “about 8%” is a rounded share, and any attempt to back a catalogue size out of it lands anywhere between roughly 121,000 and 137,000 (10,258 ÷ 8.5% and ÷ 7.5%). We do not publish that inferred number, because it is an artefact of rounding rather than a count. Never divide one source’s numerator by another’s denominator, and never invert a rounded percentage to recover a base.

The disambiguating rule that runs through this page: an adoption figure is a fact about a question, not a fact about games. Before quoting one, name the question, the population, the year and who paid. And keep the three measurement families strictly apart — surveys measure workflow, Steam flags measure declared content, and nothing at all measures output. The third gap is not a data problem waiting on better instruments. Nobody has built the instrument.

Key takeaways

• Adoption is real, shallow and pointed at the pipeline rather than the play. 36% of games professionals use generative AI; among them 81% for research and brainstorming, 19% for asset generation, 5% for player-facing features — roughly 6.8% and 1.8% of all GDC respondents respectively (16Best analysis of GDC 2026). AI reaches players as content far more often than as behaviour.
• What ships is assets, not gameplay. AI disclosures hit 30.8% of 2026 Steam releases and are projected to cross half in 2027–2028, but among AI-disclosed games grossing $1M+, art (49%) and voice (27%) dominate and runtime generative systems sit under 5%. NVIDIA’s ACE launched with seven titles named in its own announcement.
• Lowering the barrier to shipping did not lower the barrier to selling. AI-flagged games were about a third of mid-2026 Steam releases and 10–27% of estimated sales — a sales-to-release index of 0.30 to 0.82 — and 93.2% of the 10,258 disclosed titles are estimated under $10,000 gross (16Best analysis).
• The productivity claim is unmeasured. No published study measures AI-attributable output in game development. The nearest randomised evidence, in open-source software, found a 19% slowdown against a 20% perceived speed-up — a 39-point perception gap that indicts every self-reported adoption figure in the field. The same team’s February 2026 follow-up, on later tools, still cannot resolve the sign of the effect. If three years of careful measurement in software cannot settle it there, nobody is entitled to assume it in games.
• The falsifiable test cannot currently be run. AAA budget inflation near 24.5% a year doubles per-title spend every ~3.2 years, and no continuous published budget or schedule series exists in which to look for an AI-shaped break.
• Sentiment moved; use barely did. Negative views went 18% → 30% → 52% across three GDC editions while personal use went 31% → 36% → 36%. Artists (64% negative) and designers (63%) are the most opposed; executives and business operations the most favourable (19% each).
• So the spine, proven: the industry has adopted AI widely, cites its efficiency constantly, and has published nothing that would let anyone verify it. Every headline percentage in this field measures a question about workflows. Not one measures a game getting made faster.

Frequently asked questions

How many game developers use AI in 2026?

About 36% of games professionals personally use generative AI tools, according to the GDC 2026 State of the Game Industry survey of more than 2,300 respondents, published January 2026. That splits to roughly 30% inside game studios and 58% at publishing, support and marketing firms. Higher figures circulate because they answer different questions: 52% of GDC 2025 respondents worked at a company that had implemented generative AI, a Google Cloud and Harris Poll survey of 615 developers fielded in mid-2025 found 90% already using some form of AI in their workflows — a wider category than generative AI — and Unity’s 2025 report — based on a Cint survey of 300 developers conducted in 2024 — found 96% say their studio uses AI tools in select workflows. The spread is 2.7 times, driven by question wording, population, field date and sponsorship.

What do game developers actually use generative AI for?

Mostly for thinking and admin. Among GDC 2026 respondents who use generative AI, 81% use it for research and brainstorming, 47% for daily tasks such as emails and meeting planning, 47% for coding assistance, 35% for prototyping, 19% for asset generation, 10% for procedural generation and 5% for player-facing features. Applied to the 36% adoption base, that puts asset generation at roughly 6.8% of all GDC respondents and player-facing features at roughly 1.8% — so AI-made content does reach players, while AI running as a game feature almost never does. The most-used tools are ChatGPT (74% of language-model users), Google Gemini (37%), Microsoft Copilot (22%), Midjourney (17%), Adobe Generative Fill (13%) and Sora (8%).

What percentage of Steam games use AI?

AI disclosures appeared on 10.9% of Steam releases in 2024, 19.9% in 2025 and 30.8% of 2026 releases through July, based on analyst Sulka Haro’s census of 53,597 titles released between July 2023 and July 2026 and published on 13 July 2026. Separately, Totally Human Media counted 10,258 titles carrying disclosures in November 2025, about 8% of the Steam library at the time. Read these carefully: Valve’s disclosure covers AI-generated content in the shipped game, store page or marketing, and since the January 2026 policy rewrite explicitly excludes AI coding assistants and other efficiency tools. It is a measure of declared content, not of AI use in development.

Is there proof that AI makes game development faster or cheaper?

No published study measures AI-attributable output, cost or schedule change in game development. What gets cited instead is a GitHub-run 2022 experiment in which 95 recruited developers built an HTTP server in JavaScript, where the AI group finished 55% faster; a METR randomised trial published in July 2025 in which 16 experienced open-source developers were 19% slower with AI while estimating they had been 20% faster; a self-reported 2023 case study from mobile studio Lost Lore headlined at about $70,000 saved on art, of which only about $40,000 is itemised anywhere in the article; and executive statements carrying no figures at all. METR has since marked that 2025 write-up out of date: its follow-up experiment, posted 24 February 2026 with 57 developers and more than 800 tasks on later tools, reports point estimates of −18% and −4% with confidence intervals that cross zero, and METR itself calls the signal unreliable because 30 to 50% of participants withheld tasks they did not want to attempt without AI. None of these measures a game studio at scale, and the pair shows both that self-reported productivity can invert the measured result and that the measured result is still unsettled in software, let alone in games.

Do AI-generated games sell well on Steam?

They under-index. By mid-2026 AI-flagged titles were about a third of new Steam releases but only 10 to 27% of estimated sales quarter to quarter, up from 3 to 6% in 2024 — a sales-to-release index of 0.30 to 0.82. Controlling for launch month, an AI-flagged game reaches 100 or more reviews, roughly 3,000 units on the same estimator, at about 55% of the non-AI rate, and that ratio has not improved in two years. They score lower with players too: a mean review score of 77% positive against 81%, and 71% reaching Mostly Positive against 80%. Of 10,258 AI-disclosed titles, an estimated 9,556 grossed under $10,000 and 45 cleared $1 million, against a combined estimated gross to date of $660 million. Sales figures are Boxleiter-method estimates from review counts, not audited revenue.

Are game industry layoffs caused by AI?

Mostly not, on the published evidence. The bulk of tracked cuts — roughly 14,600 in 2024 and 9,053 in 2025 — are attributed to a correction of pandemic-era over-hiring, and predate most AI-first restructurings. A small number of companies have named AI explicitly: Playtika disclosed on 14 January 2026 that it would cut about 500 employees, 15% of its workforce, moving to leaner teams relying on AI and automation, expecting $12 to 15 million in restructuring charges — about $24,000 to $30,000 per departing employee. Krafton declared an AI-first reorganisation on 23 October 2025, committing 100 billion won (about $70 million at roughly 1,430 won to the dollar) to a GPU cluster, and opened a voluntary resignation programme weeks later. Neither published a productivity figure supporting the change.

What rules govern AI use in video games?

Three main constraints. Valve added an AI disclosure requirement to Steam’s content survey on 10 January 2024, splitting it into pre-generated and live-generated content, and significantly rewrote it in mid-January 2026, dropping code assistants and other development-efficiency tools from mandatory disclosure while keeping and tightening the live-generated requirement with guardrail and player-reporting obligations. SAG-AFTRA members ratified the 2025 Interactive Media Agreement on 9 July 2025 by 95.04% to 4.96%; the strike it settled began on 26 July 2024 and had already been suspended on 11 June 2025 once a tentative deal was reached. The contract requires separate, written, specific consent before creating or using a performer’s digital replica and sets a minimum of 7.5 times scale for Real Time Generation. And training-data litigation is unresolved: Disney and Universal sued Midjourney on 11 June 2025 in the Central District of California, Warner Bros. sued on 4 September 2025, and the cases were consolidated on 4 November 2025 under the Disney docket. These are filings and a procedural consolidation order, not rulings; no merits decision had issued as of July 2026.

Sources

• Game Developers Conference — GDC 2026 State of the Game Industry, January 2026 (2,300+ respondents; 36% AI use; 52% negative, 7% positive; 30% studios vs 58% publishing; 64/63/59% negative by discipline; 19% favourable among executives) — primary survey
• Game Developer — One third of game workers use generative AI, but half think it is bad for the industry (use-case split: 81% research, 47% admin, 47% coding, 35% prototyping, 19% assets, 10% procedural, 5% player-facing; tool shares; 21% internal tools, 30% at AAA) — trade coverage of the primary survey
• Game Developers Conference — GDC 2025 State of the Game Industry (3,000+ respondents; 52% employer adoption, 36% personal use, 30% negative, 13% positive) — primary survey
• Game Developer — GDC 2024 State of the Game Industry (fielded 11–29 October 2023, 3,000+ respondents; 49% workplace use, 31% personal use; 84% ethically concerned) — primary survey
• GamesBeat — GDC survey exposes game dev views of genAI and layoffs (GDC 2024 policy detail: 51% of respondents at companies with some generative AI policy, 2% mandated, 12% not allowed to use any; 21% of AAA developers at companies banning the tools against 9% of indies) — trade coverage of the primary survey
• GameDev Reports — GDC: The State of the Game Industry 2026 (36% actively using vs 52% in some capacity; 78% of companies with an AI policy; 54% US respondents; sampling-methodology change noted) — secondary summary of the GDC report
• Sulka Haro — Three years of AI on Steam, 13 July 2026 (full census of 53,597 titles, July 2023 – July 2026; 10.9% / 19.9% / 30.8% disclosure shares; ~530 AI launches a month; 60–90% of release growth; sales share 3–6% in 2024 rising to 10–27% against a ~33% launch share; ~55% relative success rate at 100+ reviews; top two AI games 45–79% of quarterly AI revenue; top 1% take ~94% of revenue; 22% retroactive flags; 138 successes vs 400 flops disclosure-text analysis; ~50% of releases projected for 2027–2028) — independent census, revenue estimated not reported
• Sulka Haro — AI flagged games on Steam, Part 2, 13 July 2026 (review scores 77% vs 81% mean, 80% vs 85% median; 71% vs 80% reaching Mostly Positive; median price $4.99 vs $5.99; ~9% pooled and ~11% recent share of top-100 monthly releases against 37% of releases in those months; ~950 non-AI first-time devs a month vs 140→440 AI debuts) — independent census
• Totally Human Media — Games with AI disclosures have grossed an estimated $660M on Steam, November 2025 (10,258 titles, about 8% of library; 12/33/170/487/9,556 revenue-bracket counts; use cases among $1M+ earners; runtime generative systems under 5%; Boxleiter as reviews × price × 35) — independent analysis, Boxleiter estimates
• Tom’s Hardware — 1 in 5 Steam games released in 2025 use generative AI, 16 July 2025 (reporting Totally Human Media’s earlier snapshot: 7,818 disclosed titles, 7% of a stated base of just under 115,000 Steam games) — trade coverage
• METR — Measuring the impact of early-2025 AI on experienced open-source developer productivity, 10 July 2025 (16 developers, 246 tasks, Feb–Jun 2025; 19% slower, 24% expected speed-up, 20% perceived) — randomised controlled trial, since marked out of date by its authors
• METR — We are changing our developer productivity experiment design, 24 February 2026 (57 developers, 800+ tasks, 143 repositories from August 2025; −18% and −4% point estimates with intervals crossing zero; 30–50% of developers withholding tasks; described by METR as an unreliable signal and a likely lower bound) — follow-up experiment, preliminary
• GitHub — Quantifying GitHub Copilot’s impact on developer productivity and happiness (95 developers, single JavaScript task, May–June 2022; 55% faster, 95% CI 21–89%) — vendor-run experiment
• Unity — 2025 Unity Gaming Report, published 17 March 2025 (Cint survey of 300 developers conducted in 2024, ±5.66% at 95% confidence; 96% studio use in select workflows, being the complement of the 4% using none; 79% positive; no single workflow above 50% usage) — vendor survey
• Game Developer — Unity claims 79% of developers are feeling positive about generative AI, 17 March 2025 (sample breakdown, workflow usage, conflict with GDC data) — trade coverage
• Google Cloud / Harris Poll — 90% of games developers already using AI in workflows, 18 August 2025 (615 respondents across five countries, fielded 20 June – 9 July 2025; the question covers “some form of AI,” not generative AI specifically) — vendor-commissioned survey
• NVIDIA — NVIDIA ACE autonomous game characters, CES, 6 January 2025 (seven titles named in the announcement itself: PUBG, inZOI, NARAKA: Bladepoint Mobile PC, MIR5, Dead Meat, AI People, ZooPunk; some trade coverage of the same launch lists additional partner titles) — vendor announcement
• GeekWire — Valve reveals new rules for AI-powered game development on Steam, 10 January 2024
• VGC — Valve has significantly rewritten Steam’s AI disclosure rules, January 2026 (efficiency tools and code assistants no longer trigger disclosure)
• GamesRadar+ — Epic boss Tim Sweeney calls Steam AI disclosures “really irresponsible of Valve” (reporting a PC Gamer interview published 25 June 2026)
• SAG-AFTRA — Members approve the 2025 Interactive Media Agreement, 9 July 2025 (95.04% to 4.96%; AI consent terms; 7.5x scale for Real Time Generation) — primary contract announcement
• Deadline — SAG-AFTRA overwhelmingly approves video game agreement, 9 July 2025 (marking the official end of the strike begun 26 July 2024)
• SAG-AFTRA — Video game strike suspended at noon PT (11 June 2025, following the tentative agreement announced 9 June) — primary union announcement
• Times of Israel — Playtika cuts 15% of its workforce amid shift to leaner teams that rely on AI, January 2026 (about 500 roles, 15% of the workforce; $12–15M in severance, notice pay, benefits and related charges; Form 8-K filed 14 January 2026)
• PC Gamer — Krafton is now an AI-first company, spending about $70 million on a GPU cluster (23 October 2025)
• Korea Bizwire — Krafton pledges 100 billion won to become an AI-first company (October 2025; plus about 30 billion won a year from 2026 for employee AI tooling)
• PC Gamer — Krafton launches voluntary resignation programme (from 12 November 2025; 6–36 months’ severance; company-wide hiring freeze)
• VGC — EA staff struggling with management demands to use AI for just about everything, October 2025 (Business Insider reporting; ~15,000 employees; flawed code being corrected)
• Georgetown Law Tech Institute — Disney, NBCUniversal and DreamWorks file copyright suit against Midjourney (June 2025)
• Graphic Policy — Disney, Universal and Warner Bros. consolidate their cases against Midjourney (order dated 4 November 2025)
• Game World Observer — AI use case: how a mobile game development studio saved $70K in expenses (Lost Lore / Bearverse, January 2023; self-reported, no control) — single-studio case study