Edition one
AI disclosure on Steam
What every game released on Steam since 1 July 2023 says in its AI Generated Content Disclosure, read and classified by what the AI was used for.
16,968 store pages, a seeded-random sample of the 62,612 games released on Steam since 1 July 2023. Not a census. Figures as of 26 August 2026.
Steam began asking new submissions to declare their use of generative AI in January 2024. The question everyone has asked of that data since is how many — how many games disclose, and is the number going up.
It is going up. That turns out to be the least interesting thing in the declarations, because developers write them in their own words, and the words say which part of the work the AI did. This report reads the wording and sorts each disclosure by the use it names. It is built to be run again, so the mix can be watched over time rather than guessed at.
One thing to carry into the charts: the window opens six months before the requirement did, so the first column of each one is voluntary disclosure, and thin.
The three measured figures
Share of new releases disclosing AI-written code
16,968 store pages, a seeded-random sample of the 62,612 games released on Steam since 1 July 2023. Not a census.
| Release year | Disclosing AI-written code | 95% interval | Releases | Sampled |
|---|---|---|---|---|
| Jan–Aug 2026, year to date | 4.91% | 4.33–5.56% | 234 | 4,766 |
| 2025 | 1.02% | 0.79–1.32% | 57 | 5,602 |
| 2024 | 0.30% | 0.18–0.50% | 14 | 4,741 |
| Jul–Dec 2023 | 0.11% | 0.03–0.39% | 2 | 1,859 |
Across the same span, the share carrying any AI disclosure at all rose 2.8-fold, and the share naming code rose 16.4-fold — both measured from 2024 to Jan–Aug 2026, year to date. The 2026 interval does not overlap the 2025 interval.
The easy reading of a rising line is that more developers are filling in the form. That is happening, and it is the smaller half of this chart: disclosure overall did not quite triple across the same span, while the code line went up sixteenfold. More games are filling it in, and what they are filling it in about is moving as well.
The direction has held in every year measured, and the most recent rise is the largest so far. It is still a small number — around fifty games move it a whole point — so the shape of the line is the claim here, not its exact height.
What the AI is used for
- Any disclosure
- Visual assets
- Audio
- Translation
- Code
16,968 store pages, a seeded-random sample of the 62,612 games released on Steam since 1 July 2023. Not a census. In-game writing and Use not stated are given as counts, not shares — see the accuracy table under “How this was measured”.
| Release year | Sampled | Any disclosure | Visual assets | Audio | Translation | Code | In-game writing (count) | Use not stated (count) |
|---|---|---|---|---|---|---|---|---|
| Jan–Aug 2026, year to date | 4,766 | 35.04% | 26.29% | 9.32% | 5.14% | 4.91% | 165 | 172 |
| 2025 | 5,602 | 21.58% | 17.07% | 4.68% | 1.86% | 1.02% | 70 | 114 |
| 2024 | 4,741 | 12.44% | 10.10% | 2.11% | 0.59% | 0.30% | 34 | 65 |
| Jul–Dec 2023 | 1,859 | 1.29% | 0.70% | 0.27% | 0.22% | 0.11% | 2 | 5 |
In-game writing and Use not stated are counts and not shares, because their hand-checked accuracy does not support a rate. Each is a floor — the classifier misses some, so the true number is higher, and no share can be computed from a count that is known to be short:
- In-game writing — at least 165 of the 4,766 2026 releases sampled. Story, dialogue, lore, quest and item text. Store and marketing copy is excluded.
- Use not stated — at least 172 of the 4,766 2026 releases sampled. A disclosure that names AI use but no use type.
Ask what AI in games means and the answer is art. It still is: three disclosures in four name visual assets, and art is the largest of these uses in every year on the chart.
It is also the one growing slowest — measured from 2024, the first full year under the requirement. Audio, translation and code have each multiplied faster from much smaller bases. Measure instead from the leftmost column, a voluntary half-year before the requirement existed, and that ordering reverses. Which is the catch with every growth claim on this page: it only means something once you say where you started counting.
What none of it says is how much AI went into any one game. Every column counts games that name a use, so three generated lines of dialogue and a wholly generated soundtrack land in the same place.
Self-published and externally published
The split is Steam’s own developer and publisher strings, compared after normalising case and punctuation. 16,968 store pages, a seeded-random sample of the 62,612 games released on Steam since 1 July 2023. Not a census.
| Publishing | Sampled | Any disclosure | AI-written code | 95% interval on code |
|---|---|---|---|---|
| Self-published | 12,444 | 22.51% | 2.07% | 1.84–2.34% |
| Externally published | 3,967 | 14.42% | 0.93% | 0.68–1.28% |
| Publisher not stated | 557 | 21.54% | 2.15% | 1.24–3.73% |
This is not “solo developer”, and the bias runs one way. Self-published means the developer string equals the publisher string. A studio of forty people with its own publishing arm falls in the same bucket as one person on a laptop, so the self-published bucket overcounts solo developers — which means the true gap between a solo developer and a studio is at least as large as the one above, never smaller. Steam publishes no team-size field, so nothing finer can be separated from this data.
Publisher not stated is reported as its own row rather than assigned to either side: those are releases where Steam gave no developer string or no publisher string.
You might expect the studios with budgets and legal departments to be the ones declaring AI-written code. It runs the other way: games published by their own developer name it more often than games with an outside publisher, and the two intervals do not overlap.
The direction is solid. The size is not — it reads at around two to one here, on an interval wide enough to admit a good deal less, and it has narrowed as the sample has grown. Nor can the split say why, or who: self-published covers a studio with its own publishing arm as readily as one person on a laptop, and Steam publishes nothing about the size of a team.
Against the published reference
- This study
- Published reference
Reference: Sulka Haro, "Three years of AI on Steam" (n=53,597 games, Jul-2023..Jul-2026).
| Release year | This study | Sampled | Reference | Difference |
|---|---|---|---|---|
| 2026 | 35.04% | 4,766 | 30.8% | +4.24 pt |
| 2025 | 21.58% | 5,602 | 19.9% | +1.68 pt |
| 2024 | 12.44% | 4,741 | 10.9% | +1.54 pt |
The two figures disagree, and the difference runs the same direction in every year. That difference is not smoothed here, and two measurement differences account for its sign.
Disclosures get added after release. The reference was measured to July 2026 and this corpus was read in August 2026, so every release cohort has had longer to acquire a disclosure by the time we looked. A later-measured census of the same cohort should read higher.
The 2026 row is not the same window. The reference’s figure for that year is year-to-July; this study’s covers Jan–Aug 2026, year to date. The rate is rising steeply within the year, so the largest difference landing on the most recent year is what a rising trend predicts.
Nothing in this pipeline is tuned against the reference. The classifier does not affect these figures at all: whether a page carries a disclosure is decided by the crawler’s extraction, and the classifier only decides what each disclosure is about.
Two implementations built independently, over different samples, land within two points of each other in both of the full years they share. That is worth showing, because there is nothing either can be checked against directly.
It is not confirmation, though. Both read the same store pages, so the agreement rules out one of them having built it wrong — not both of them reading the same thing the same wrong way.
Games that change their mind
A game can ship with an AI disclosure and later delete it, or ship without one and later add it. Nobody publishes that rate. Two methods answer it, they answer different questions, and their numbers are kept apart below rather than averaged — one is exact and short, the other is long and partial.
Both compare the disclosure text and nothing else. Comparing the pages would report every game as changed every time: Steam builds each store page per request, so the same page fetched twice seconds apart returns different bytes. Across the re-checks below, the page bytes differed in 100.00% (3,493 of 3,493) and the disclosure text was identical to the character in 99.54% (3,477 of 3,493). That gap is the whole reason this measurement is possible. A page that could not be read — an age gate, a sign-in wall, a delisting, a capture served in another language — is recorded as unreadable and excluded, never as a disclosure that disappeared, which is the one way a measurement like this manufactures a result.
Forward: what has changed since we looked. Re-fetching the live store page for a game we already hold a snapshot of gives an exact before-and-after with no archive involved. Two groups were re-checked 10.5–10.7 days later, and each is counted against its own denominator: a game can only remove a disclosure it had, and only add one it did not.
Of the 3,493 games that carried a disclosure and were readable on the re-check, 7 removed it and 9 reworded it — 0.20% (7 of 3,493) removed, 95% interval 0.10–0.41%. Read the top of the interval rather than the point: on this sample, fewer than 0.41% of disclosing games drop the disclosure in a fortnight.
Of the 4,000 that had none, not one added a disclosure over the same window. A count of zero is not a rate of zero, so the claim is the ceiling the sample supports: fewer than 0.10% of undisclosed games start disclosing in a fortnight. Adding one takes longer than that, which is what the archive below shows.
Backward: what already happened. Reading history needs a web archive, which edition one could not reach at all. It answered every request this run made — not the same claim as the archive being reliable, since roughly half of single attempts came back with a server error and were retried. 400 games were asked about and history was read for 349 — archive coverage 90.50%. A change can only be seen where two captures parsed, so the denominator is not the games we asked about but the 347 with usable history, 86.75% of them.
Of those, 8 added a disclosure after launch and 5 removed one: 3.75% (13 of 347), 95% interval 2.20–6.30%, changed their mind either way. A rewording is not a change of mind and is not counted in that figure. The interval is wide because the denominator is small, and it is the denominator that should be quoted with it.
Every one of those is an interval, not a date. The archive never gives the moment of change. It gives “true at this capture, different at that one”, and the change happened somewhere between them — a median span of 186.4 days and a widest of 430.5. There is no midpoint anywhere on this page, because a midpoint is a number we could not defend if someone checked it.
How this was measured
Steam publishes a game’s AI Generated Content Disclosure only in the store page HTML — not in the store API, and not in any structured field — so the corpus is store pages, fetched one at a time at a rate well below the one that earns an IP block, with the language pinned so the heading is stable. Pages that are age-gated or are not really an app page are recorded as unreadable rather than as “no disclosure”, and every rate below is over readable pages only.
The catalogue was enumerated in full: 181,621 games, of which 62,612 were released on or after 1 July 2023. Store pages are then fetched in a seeded random order, so any prefix of the crawl is an unbiased sample of the window rather than a front-loaded slice by release date. This edition reads 16,968 of them — 27.10% of the window. Every figure on this page is a sample estimate, which is why each one carries its denominator and its 95% Wilson interval.
The classifier is a pure function over the disclosure text. It never reads the game’s name, tags or other store metadata, so a label always traces back to the words that produced it. It is multi-label: a disclosure naming two uses counts in both.
Accuracy was measured by hand, twice. Two hundred disclosures were read and labelled by hand and the classifier was tuned against them; a further 100, disjoint by app, were drawn afterwards and labelled once with no tuning. The held-out set is the one quoted below. All five labels correct on the same record: 92% of the held-out set.
A label is quoted as a rate only where both its held-out precision and its held-out recall reach 0.85. Below that it is reported as a count. That rule is applied to the measured numbers, not to a list of approved labels — which labels appear as rates on this page is decided by the table below and by nothing else.
Held-out accuracy, measured on 100 hand-labelled disclosures drawn after the classifier was finished.
| Label | What it counts | Precision | Recall | Hand-labelled positives | Reported as |
|---|---|---|---|---|---|
| Visual assets | Images, textures, 3D models, icons, UI, animation or video — in the game or on the store page. | 1.00 | 0.96 | 80 | Rate |
| Audio | Music, sound effects, voice, text-to-speech or narration. | 0.96 | 0.96 | 23 | Rate |
| In-game writing | Story, dialogue, lore, quest and item text. Store and marketing copy is excluded. | 0.86 | 0.75 | 8 | Count only |
| Code | Source code, scripting, programming or debugging. | 0.89 | 0.89 | 9 | Rate |
| Translation | Localisation into other languages, including machine translation. | 1.00 | 1.00 | 10 | Rate |
| Use not stated | A disclosure that names AI use but no use type. | 0.57 | 0.80 | 5 | Count only |
What this does not measure
- Disclosure, not use. This measures what developers say. A game that used AI and said nothing is invisible from here, so every figure is a floor on real use.
- In-game or marketing. A large share of visual assets disclosures are store page and capsule art only, and many say so outright. The correct reading is “discloses AI-generated visual assets, including marketing”, never “contains AI art”.
- Development-time or runtime. A game whose feature is a generative AI companion and a game whose developer used a model to fix a bug both count once under code.
- How much. Three lines of dialogue and every image in the game each count once. There is no intensity in this data.
- A sample, not a census. 16,968 store pages, a seeded-random sample of the 62,612 games released on Steam since 1 July 2023. Not a census. Rates are estimates with the intervals printed beside them, and they will move as the crawl continues.
- Engine. Steam publishes no engine field and the data collected carries no reliable engine signal, so no cut by engine is computed or published here.
- In-game writing and Use not stated, as rates. Reported as floors and never as shares, because their hand-checked accuracy is too low to support one. Non-English disclosures are the largest known miss.
- When a disclosure changed. The archive bounds every change between two captures and never dates it, so the churn figures above are rates over a period, not a timeline. See above.
Two of these have closed since this page first went up. Backward churn — how many games quietly add or remove a disclosure after they ship — needed an archive that was unreachable for the whole first run, and it is measured above now, on a small enough sample that the interval is the honest part of it. What is still open: narrative needs a classifier good enough to report as a rate instead of a floor, and coverage should be the whole window rather than the sample this edition rests on.
Disclosures not written in English are the largest known reason one ends up unclassified, and the most likely source of a correction. If you have one this reads wrongly, send it to ed@flockbay.com. A correction is worth more to the next edition than agreement is.
Reuse and citation
The charts and figures on this page may be republished with attribution to flockbay.com/reports/ai-in-games. If a figure is quoted, quote the denominator and the interval with it — a rate from this page without its n is a rate this page did not publish.
Flockbay, “AI disclosure on Steam”, edition one, 26 August 2026. https://flockbay.com/reports/ai-in-games
Classification as of 26 August 2026; catalogue coverage as of 26 August 2026; churn as of 5 September 2026.