AI Update Part 2: AI could still be hurting game sales, but in a less obvious way
In Part 1 of this series where I revisit the question of AI disclosure on Steam, I showed that disclosure has gone from 15% of releases to around half in under two years. Today, I’m going back to the question that got us all that attention earlier this year: does using AI hurt sales?
A quick recap. Back in December 2025 I compared 9,879 commercial Steam releases with and without an AI disclosure and found that the AI games received roughly 53% fewer reviews in their first month; given that reviews are the best proxy for sales, that means those games sold significantly less. This was even after I accounted for developer experience, publisher backing, the type of game, and the month of release. I finished with a TLDR that "AI Is Neutral For Beginners But Problematic For Experienced Devs".
A year on I have three times the data, a better model, and a new ability to see how AI is being used in each game. In today's blog, I’m going to use this to cover the following:
-
The simple view: how do AI games compare on the face of it?
-
The careful view: what happens when I compare like with like?
-
A correction, because part of last year's TLDR doesn't hold up anymore.
-
Does it matter what you use AI for?
-
What kind of games are these, is it all just low-effort shovelware, and which AI games did sell?
Spoiler
This is the internet and people quote stuff without reading the full article. So just in case you do not read the full article, I’m putting this here to make sure I'm not misquoted.
Yes, the expected loss in sales is still there on the face of it; AI games receive significantly less sales. But most of the damage is already done by launch day, before anyone has played the game, and that changes what the findings mean!
Either the penalty of disclosing AI applies before launch and people refuse to support an AI game. Or, AI is the symptom of a wider problem, with people speed running development and not putting the care and attention into marketing and community building that a successful launch requires. I urge you to read to the end of the article, but if you don’t, this is the main takeaway.
As with last year, when I say "sales" I mostly mean "reviews". Only the developer knows their true sales, and review counts are the best public proxy I have. If you would like to read about how Game Oracle uses reviews to estimate sales you can check out our dedicated blog on the subject.
On the face of it, AI still appears to have a problem
After removing free-to-play games, unreleased games, and developers who publish faster than one game every six months (the same "slop" filter I used last year, which helps us identify honest commercial intent), I have 31,003 commercial games released between January 2025 and October 2026. Of those, 9,251 (30%) have an AI disclosure; this is a big jump from last year when that figure was 17.9%!
Our first finding is that a third of all AI games have no reviews at all, implying that they received little to no sales. Compare this to titles without AI disclosure, where only around 1-in-6 have no reviews.
Now, if you remember the milestone from Part 1, I considered 1,000 sales as a healthy and realistic target for an indie release. It turns out games without an AI disclosure are about twice as likely to get to 1000 sales. The distribution of sales for the two groups are shown below.
The same pattern shows up in followers, our stand-in for wishlists: the median AI game has 17 whereas games without any disclosed have 47. So it’s not just sales that’s impacted, but also purchase interest.
But what about player reception?
Looking only at games with at least 100 reviews, so the scores mean something, the median share of positive reviews is 82.0% for AI games and 87.1% for games without an AI disclosure. Last year this was 84.6% vs 88.3%, so proportionally it hasn’t changed that much.
Enjoying this article? Sign up to our newsletter for monthly market insights and the latest indie innovations
“But AI games are newer!”
If you were about to shout this at your screen, good. You're right.
AI disclosure has grown so quickly that AI games are much younger: the median AI game in our data has been on sale for 149 days whereas the median non-AI game has been on sale for around 300 days. Younger games have had less time to collect reviews. The simplest fix is to compare games released in the same quarter:
In every one of the eight quarters I observed, AI games have a lower median review count and are less likely to reach 1,000 sales. Games from early 2025 have had well over a year to find their audience, and the AI ones still sit at half the median reviews of their non-AI counterparts.
So using AI is, at the very least, still correlated with poorer performance on Steam. I want to emphasize a word there: correlated. Correlation does not mean causation!
We can go deeper (again)
Correlation is where last year's analysis started, so I wanted to do the same here. But correlation doesn’t answer the most critical question: does AI disclosure impact sales?
There is a lot to account for if I'm to answer this question. AI games are not just newer. They are also more likely to be a first release and less likely to have a publisher. 76% of AI games are a first release vs 68% of games with no AI disclosure and 14% of games disclosing AI use have a publisher vs 22% of games without an AI disclosure.
This, along with countless other reasons, could explain the loss of reviews without AI having anything to do with it.
So, like in last year’s analysis, I built a statistical model to ask a much more specific question:
For the same type of game, released in the same month, on sale for the same number of days, with the same publisher backing and the same developer experience, how many reviews does a game with an AI disclosure get compared to one without?
The ingredients come from last year's causal diagram (the one that made it look like I'd had a mental break). However, three things have changed:
-
I count all reviews to date, not just the first month, and account for how long each game has been on sale using a correction term within the model itself.
-
I count reviews from Steam purchases, not keys bought elsewhere. Although I did check, and re-running on all reviews makes almost no difference.
-
The model is different. Last year's model, and three close relatives, could not reproduce what Steam review counts actually look like: a huge pile of games with nothing, and a tiny number of hits with everything. The new model treats "does this game get any reviews?" and "how many?" as two separate questions. It is fundamentally a more robust model compared to last year.
The Results
Take two developers making the same type of game, releasing in the same month, with the same publisher status and the same level of experience. The one with the AI disclosure can expect 49% fewer reviews. If developer A got 100 reviews, developer B, who made a similar game but disclosed AI, would get about 50.
Hang on, why are there two numbers this year?
Because averages on Steam are a bit misleading. The top 1% of games hold 76% of all reviews, so an "average" is mostly a statement about the top hits.
Most of you aren't asking "what happens to the whole market?". You're asking "what happens to a game like mine?". So the model also reports the expected reduction in reviews for a typical game, the middle of the pack among games that get reviewed. To simplify things, going forward I will just discuss the expected reduction in reviews/sales for a typical game that discloses AI use compared to a typical game that does not, but I wanted to show both for transparency.
In that cohort, games with AI disclosures get 49.5% fewer reviews; so about half the reviews of a game without a disclosure. The chance of getting any reviews at all also drops, from 82.4% to 74.4%.
So far, then: compared on the things last year’s model controlled for, a typical AI game gets about half the reviews, and that barely moves across the robustness checks.
However! This is the difference between games that are comparable to the things I controlled for in last year's analysis! Later on I find something it doesn't control for and this thing accounts for most of the effect! So please don't stop reading at 49.5%!
I tried to break the model
Any good scientist knows that you should try and break things before you make your mind up. A single number from a single model deserves suspicion, so I re-ran the analysis a bunch of different ways:
Each of the results above is like running the model in a different hypothetical world with different inputs. A number of interesting scenarios here are worth discussing:
-
First-month reviews only, like last year: we would expect a 46% reduction in reviews, fitted on 26,745 games. So our result isn't an artefact of counting reviews to date.
-
Matching on store-page content (the embeddings behind our Steam Map): the loss of reviews barely shifts for the typical game. So the type of game being made accounts for little of it.
-
Leaving out early-access graduates. Games like Valheim and Palworld left early access in 2026, so Steam lists them as 2026 releases but there are years of reviews behind them; I wanted to make sure this wasn’t skewing things and the results suggest it isn’t.
None of the scenarios I ran came anywhere close to zero effect on sales. But, none of these scenarios account for how prepared a game was on launch day either. I’ll discuss that soon, and it turns out it matters a lot!
The full list of checks that I did are in the methodology at the bottom of the article.
But what about skill, talent, marketing, and just plain old luck?
I need to provide the same caveat as last year, and just like last year, it's a big one. Even though I would love to, I simply can't measure how skilled a team is, how much money they have, or how good their marketing is. If the developers who reach for AI also tend to have less of all three, then some of this reduction in reviews and sales is "under-resourced teams sell less" and not "AI games sell less".
Last year I tackled this by imagining an unmeasured "X-factor". I came up with hypothetical scenarios of how strongly this unmeasured factor is tied to using AI and how much it moves reviews, then saw what was left of the AI effect. This is called a sensitivity analysis.
Last year, an X-factor that nearly tripled the odds of using AI and cut reviews by 22% was enough to wipe out the penalty for AI disclosure. In other words, some unmeasured variable (e.g. marketing or skill) could account for the difference in sales.
This year, the same X-factor barely makes a dent. A typical AI game gets 49.5% fewer reviews than a comparable non-AI game. Allow for the X-factor and it still gets 46% fewer. That X-factor explains about 3 of those 49.5 points.
So how big would an X-factor need to be to explain the whole thing? A hidden factor can never explain a bigger gap than its own effect on reviews. So even in the most extreme case, where every AI developer has it and no other developer does, it would have to halve a game's reviews all on its own. That's as big as the AI effect itself. If it's any less tied to AI use, it would have to be bigger still. That's enormous.
I want you to hold onto this thought though, because further down I go looking for something that big, and I may have just found it. But first, a correction…
A correction: The Impact of using AI is not "neutral for beginners"
This is the part where I have to hold my hands up and apologise.
Last year, I read that heatmap from the sensitivity analysis as two stories: an inexperienced developer whose game was going to struggle anyway, for whom AI makes no difference, and an experienced studio that gets hammered for using it.
The problem is that the heatmap can't tell you that. Each cell answers "what would the AI effect be for everyone if this hidden factor existed?". The rows are not types of developer. Ultimately, what I did was read a chart about what I don't know as if it were a chart about who the developers are.
This year I tested the impact of prior experience directly instead. I did this by letting the model estimate a different AI effect depending on how many games a developer has released before:
The impact of using AI is not neutral for beginners. A first-time developer creating a typical game with an AI disclosure gets about 47% fewer reviews, and is 8 percentage points less likely to get any reviews at all.
Experienced developers do have more to lose, just not in the way I said. For a typical game, the loss actually grows a little with experience: 47% fewer reviews for a first release, 54% for developers with one or two previous games, and 59% for three or more (partly because a first release starts so low that it has less room to fall). And because experienced developers start from a higher base, the number of reviews lost grows much faster. A typical first release would get about 8 or 9 reviews without a disclosure and 4 or 5 with one, so it loses about 4. A typical game from a developer with three or more previous releases would get about 44 without and 18 with, so it loses about 26. The proportional hit is a bit bigger for experienced developers, but the number of reviews (and sales) lost is wildly different.
So the honest version of last year's TLDR would have been: everyone takes a big hit, beginners included; experienced developers lose a slightly bigger share, and because they start from more, it costs them far more.
Quick note: “previous releases" is not a measure of talent or budget. It is just the best proxy I have for developer experience, and it's still a rough one, so please take that into account when reading these statistics.
New this year: does it matter how you use AI?
Last year I wrote that I couldn't include how AI is used, and that it matters.
Thanks to the disclosure classifier I built for Part 1, this year I can at least look at what AI was used for, comparing games disclosing each use against games with no disclosure at all:
Every category comes with fewer reviews overall. The smallest losses in expected sales are for localisation (25%) and voice (40%), though even for localisation the loss stays clear of zero. My best guess is that translating a game with AI is the one use here that plausibly brings more players through the door, and it sits on top of a finished game rather than replacing part of it. But that is a guess.
The largest expected loss is for AI-written code (67%), which really surprised me, since my assumption was that AI generated art drives the most negative discourse around AI in games. But, before anyone uses that to win an argument online: the categories overlap heavily (a game disclosing AI code usually discloses AI art and music too), the labels come from a classifier reading free text, and plenty of developers use coding assistants without ever mentioning it.
So what kind of games are these, anyway?
Whenever I talk about this analysis, somebody says: "Well, obviously. AI games are all asset flips and incremental-game or friendslop shovelware. Of course they don't sell."
That's really two claims. One, that AI games pile into a particular corner of Steam. Two, that they're a flood of low-effort releases that were never going to sell, AI or no AI. Luckily, both are testable, at least partly.
Are AI games flooding one genre?
Maybe, but not to the extent people believe. Every one of Steam's genres sits between 25% and 35% AI, against 30% disclosing AI overall. The player-applied tags show more variation: AI disclosure runs at 40% to 45% in text-based games, card games, tabletop and management sims, and under 20% in platformers and hand-drawn games. That makes sense to be honest, the top of the list is games made largely of still images and text, which is what today's AI tools produce most easily.
The most detailed view comes from our Steam Map, which places every game on Steam according to its store page content, and groups them into a few hundred segments of similar games:
Most of the map is the same mid-red. Half of all market segments are between 23% and 34% AI, and AI games are the majority in just 8 of them, which account for 2% of all AI games on Steam. That small, bright island at the bottom left is jigsaw puzzles: 22 of its 40 games disclose AI, which also makes sense when the whole game is a set of pictures.
Is it just that those genres sell badly?
Maybe card battlers and visual novels are simply crowded, low-selling corners of Steam, right? To test this, I compared AI games with non-AI games from the same market segment, released in the same quarter. It barely changes anything: 12% of AI games reach 1,000 sales, against 24% of all non-AI games and 23% of non-AI games matched to them on market segment and release quarter. In 107 of the 111 segments with enough games to compare, AI games do worse than non-AI games in the same market segment.
So genre isn't the explanation. Whatever is going on with the impact of using AI happens within all genres.
A flood of games nobody was waiting for?
That leaves the second claim, which I find much harder to dismiss: that AI has made it cheap to finish a game, so we’re seeing a wave of games shipped quickly, without the months of wishlist-building, community work, and polish that a successful launch normally needs.
I can't measure effort, but I can measure some things that tend to come along with effort. AI games are a bit more likely to be a developer’s first game, a bit less likely to have a publisher or a demo, and support more languages. Most of those differences are modest, but according to the data two are not:
| Games Using AI | Games Not Using AI | |
| Days the store page was up before release; store-page runway (median) | 52 | 167 |
| Released within 60 days of the store page going up | 53% | 30% |
| Followers on launch day (median) | 6 | 20 |
| Fewer than 10 followers on launch day | 61% | 39% |
The typical AI game had its store page up for 52 days before release whereas the typical non-AI game had theirs up for 167 days. More than half of AI games launched within two months of their page going live. This also shows up in the audience they built as well: the typical AI game goes into launch day with 6 followers, versus a median of 20 followers for non-AI games.
If you know anything about wishlists you know where this is going. When I compare AI games with non-AI games that were in the same position on launch day this is what I see:
Now matching it up like this is not very robust, so I also went back to the model and gave it the launch-day measures (how long has the store page been up prior to launch and how many followers a game had on launch day), alongside everything it already controls for:
| What the model also controls for | Expected loss in reviews/sales for a typical game | Drop in the chance of any reviews |
| Nothing extra (the main result) | 49% | 8 points |
| How long the store page was up | 29% | 4 points |
| Followers on launch day | 9% | 2 points |
| How long the store page was up for and followers on launch day | 16% | 2 points |
Accounting for the store-page runway alone reduces the expected loss in reviews from 49% to 29%. If I instead account for followers on launch day, the reduction in reviews falls to just 9%. Finally, I asked the model to control for both together, which is the fairest approach. This model reported the reduction in reviews as 16% for a typical game. So the impact of AI disclosure is smaller, but not zero.
But why does adding the store-page runway on top of followers make the reduction in expected reviews a little bigger?
Among games with the same following on launch day, the ones that built their following quickly go on to sell better (a fast-growing audience is a good sign), and AI games are much more likely to have gathered their followers quickly. When I looked at followers alone, it appears that AI games were quietly getting credit for this rapid accumulation effect.
So here's where that leaves the “X-factor” I spoke about earlier: I said its effect would need to be enormous to explain the reduction in reviews. This right here is what an enormous effect looks like! The size of the audience a game has built by launch day is one of the strongest predictors of sales anyone can have, and according to the data AI games arrive with a fraction of the audience size that is needed!
I must admit though, this isn't quite the whole story. Among games that launched with a real following, 100 followers or more, the typical AI game still ends up with about 12% fewer reviews than a non-AI game that started with the same audience. But that is a very different number from the 49% reduction in reviews we saw at the beginning of this analysis!
Two ways to read this
I want to be careful here, because this result can be made to say two very different things.
-
It was never really about AI. AI tools have made it possible to get a game onto Steam faster and cheaper than ever. A lot of the games doing that are first projects, released with a few weeks' notice, no demo, no publisher, and no audience. Games like that have always struggled on Steam. They did not sink because they use (and disclose) AI, it's just what a lot of them have in common.
-
The penalty of disclosing AI arrives before launch. Followers result from players looking at a store page and deciding they care about the game being made. The AI disclosure is on that page from day one, and so, usually, is the AI art. If players are scrolling past it because they’re put off by AI use, a low follower count on launch day isn't an alternative explanation for the AI penalty, it is because of the use of AI that a game has a low follower count to begin with. This is essentially the same AI penalty I warned about in our original 2025 analysis, but it is just showing up before launch rather than after.
Honestly, I can't tell you which explanation is right. Personally, I suspect both are right, and they’re mixed in a way I can't measure. The short runway really has nothing to do with AI sentiment, it is a choice made by the developer. However, the low follower count could be caused by a short runway to launch and player sentiment about AI.
What I can say is narrower, and I think it’s still useful: whatever is happening to AI games has mostly happened by launch day. An AI game that does build an audience goes on to sell only a little worse than any other game with the same run-up and the same audience.
The AI games that made it
Everything above is about averages and typical games. There are of course plenty of games that are neither average nor typical. In our data, 1,065 games with an AI disclosure have passed 1,000 estimated sales, 347 have passed 10,000, and fifty have passed 100,000. Here are some of the biggest, with what their disclosure actually says:
|
Game |
What it is |
Est. sales* |
Positive reviews |
What the disclosure says |
|
ARC Raiders |
Extraction shooter from Embark Studios |
7.3M |
82% |
"...we may use procedural- and AI-based tools to assist with content creation. In all such cases, the final product reflects the creativity and expression of our own development team." |
|
Crimson Desert Enhanced |
Open-world action adventure from Pearl Abyss |
2.7M |
84% |
Generative AI used "in a supplementary capacity during the creation of some 2D prop assets", which are then "replaced through our production pipeline by our art and development teams". |
|
Revenge on gold diggers |
Interactive live-action story game, a first release |
2.0M |
80% |
"Due to budget constraints", AI generated "a minimal portion of pre-rendered visuals", "no more than 1% of total runtime". |
|
EA SPORTS FC 26 |
Football |
2.0M |
51% |
"Generative AI may have been used in creating pre-rendered or live-generated content... the result of a human-led artistic and creative process." (F1 25 carries the same wording.) |
|
Librarian: Tidy Up the Arcane Library! |
Cosy organising game, a first release, $5.99 |
1.5M |
94% |
Four assets "produced using additional AI for refinement", and AI "helped with grammatical corrections for in-game text, which were then carefully reviewed and adjusted by hand". |
|
inZOI |
Life sim |
1.1M |
78% |
AI is a feature: players "generate unique textures for character outfits" from text, and "create 3D objects from 2D image input". |
|
My Winter Car |
Car-building life sim |
710k |
94% |
"Some base textures", "some of the imagery and speech audio in the in-game TV programs", and "some in-game radio music". |
|
BidKing |
Multiplayer strategy game, $2.99 |
740k |
32% |
"Certain artistic assets in this work were produced using Stable Diffusion", including scene images, loading images and item icons. |
|
Big Ambitions |
Business life sim |
570k |
92% |
"We only used AI to translate the game into more languages. Nothing else is AI." |
|
Jump Space |
Co-op sci-fi shooter |
550k |
85% |
Text-to-speech voices as "placeholders [that] will be replaced by real voice acting as development continues". |
|
Sephiria |
Pixel-art action roguelite |
500k |
95% |
"AI tools were used for parts of the translation." |
*Sales are our estimates, built from review counts
A few things are striking about this list:
-
They aren't one kind of game. We’ve got a blockbuster shooter, a football game, a life sim, a live-action story game, a cosy tidying game, and a pixel-art roguelite. There is a big mix of genres here.
-
In most of them, AI is used for small, specific things: some paintings on a wall, a translation, placeholder voices, four assets, or one percent of the runtime, and the developers go out of their way to tell you exactly how small.
-
Some of them are just simply big games. Embark, Pearl Abyss and EA were going to sell games whatever their store page said.
-
They took their time. The median store-page runway for an AI game that passed 10,000 sales is 359 days. For those that sold under 1,000 the runway is 43 days
Another thing: BidKing is in that table with three quarters of a million copies sold but it was priced at $2.99 with AI-generated art and has a 32% positive rating. A game can sell and still not be liked.
None of this proves that careful, limited AI use is why these games sold; I picked them because they sold. But it proves the narrower point: an AI disclosure doesn't rule out a hit, and the AI games that do well look a lot like any other game that does well.
What about a first-time dev who leaned on AI?
A fair criticism of the table above is that it is made up of mostly big studios and games where AI was a footnote. So I went looking for the opposite: a developer's first game on Steam, no publisher, a disclosure that says AI did real work, and sales well above normal for its release quarter.
"Well above normal" here means at least three times the median estimated sales of every game released in the same quarter. That sounds higher than it is, since the median game sells somewhere between 15 and 110 copies.
Among first releases from self-published developers, 21% of AI games clear that bar, compared to 30% of non-AI games. Narrow it to AI games whose disclosure covers at least two kinds of in-game content, and 19% still clear it. These games are harder to find, but not exactly rare. Furthermore, it turns out 38 of those AI games have passed 10,000 estimated sales.
Here are three games where the disclosure leaves no doubt that AI did a lot of the work:
|
Game |
Released |
Est. sales* |
Positive reviews |
What the disclosure says |
|
Tasty Chef |
Sep 2026, $12.99 |
26k |
91% |
"Generative AI was used during development to create all visual art assets and music featured in the game." |
|
Settlements Rising |
Jan 2025, $18.99 |
25k |
81% |
"Out of a total of 12 music tracks, 9 were produced using AI. A total of 25 villager profile pictures were generated with AI." Most of the translations too. |
|
Corner Kitchen Fast Food Simulator |
Mar 2025, $13.99 |
24k |
85% |
"Ai generated 2D art and voice/sound." |
*Sales are our estimates, built from review counts
These games sold roughly 200 to 700 times more copies than a typical game in their release quarter. None of them were a “mega-hit”, but every one is a first game from a self-published developer, and every one is well-reviewed.
They don't all fit the launch-day story either. Settlements Rising had its Steam store page up for well over a year and launched with about 4,500 followers: the textbook run-up with strong consistent marketing. However, Tasty Chef put their store page up only five weeks before launch and only had around 250 followers.
One more for the exceptions pile that didn’t make the table: Dub Together, a $4.99 party game with AI-generated character art, launched two weeks after its page went up, with 12 followers, and has sold an estimated 50,000 copies.
I found these by looking for them, and a handful of exceptions doesn't change an average. But "a first-time, self-published developer who used AI for most of their art and still found an audience" is a thing that happens, and it happens more often than the loudest corners of the internet would have you believe.
So, what does this actually tell us?
What the data supports:
-
Games with an AI disclosure get far fewer reviews than games that are comparable on experience, publisher, genre and timing: about half as many reviews for a typical AI game and this has been the case for nearly two years, across 31,003 games, and holds up to every way I could think of to break it.
-
The effect of AI disclosure on reviews (and therefore sales) isn't only a beginner problem, it shows up for every kind of AI use I could measure, and it isn't a genre problem either.
-
Most of the effect that results in the reduction in reviews is already in place on launch day. AI games go live after a much shorter run-up and with a much smaller following. When I account for the run-up, the expected reduction in reviews for a typical game falls from 49% to 29%, and when I account for the run-up and the audience on launch day together it falls to about 16%.
-
An AI disclosure doesn't rule out success, including for first-time developers.
What it can't tell you:
-
That AI causes the drop. This is a comparison of games that exist, not an experiment. The amount of time a store page has been up before launch and the number of followers accumulated prior to launch show how much an unmeasured factor can move the expected reduction in reviews, but skill, budget and marketing effort are still mostly unmeasured.
-
Why the reduction in reviews happens. Last year I wrote that the "AI Stigma" is real and severely punishes developers. I was far too quick to make this claim. A wave of quickly made and lightly marketed games would leave almost the same fingerprints in this data. Personally, I think both things are happening at the same time.
-
Who actually used AI. I can only see who filled in the disclosure box. Games that used AI and kept quiet are sitting in the "no AI" group.
-
The actual sales a game has accumulated. I can only measure reviews and then estimate sales. If people who buy AI games simply review less often, that would also explain the reduction in reviews seen for AI games.
One thing I'm deliberately not making a fuss about is that the reduction in reviews looks more severe for games released from May 2026 onwards. It's tempting to write "the backlash is growing", but those games are young and their review counts are still settling. Ask me again in six months.
TLDR: the reduction in sales is real, but it is not a verdict on AI
In reality, the dust has not settled on the debate of what AI means for the games industry. If you're sceptical of AI in games, this data tells you that games disclosing AI are reaching a smaller audience in every genre and for every kind of use, and about half for a typical game, two years running. You might also conclude that players are quietly declining to follow games with an AI disclosure.
If you're enthusiastic about AI in games, then you might read this and think the reduction in reviews/sales we’ve seen is not simply a judgement handed down after launch. Rather, the reduction is explained away by how long a developer spent gathering an audience towards release, and AI games that put time into marketing and community building see different outcomes.
And if, like most developers I talk to, you're somewhere in the middle, then I don't think this analysis tells you whether to use AI or not. I think it tells you that the label on your store page matters a great deal less than what you've built by the day you press the release button. The games that struggle here are overwhelmingly games that arrive quickly and quietly. That was a losing strategy on Steam long before generative AI existed.
Approach AI with caution, use it gracefully, and give your game the run-up to launch it deserves.
In Part 3 I'll be looking at a question I never thought I would have to ask: how worried should you be about your game being cloned by AI? See you in the next one.
Methodology (for the nerds)
If you want to know exactly which games I analysed, what the model does, which checks passed and failed, and which rules I changed, it's all here.
The Data
This analysis uses the Game Oracle Steam catalogue, frozen on 7 November 2026: 31,003 paid games released between January 2025 and October 2026. Of these, 9,251 (30%) have an AI disclosure.
I excluded free-to-play games, games priced above $100, and developers whose previous releases averaged less than six months apart. This is the same rapid-release filter used last year. I checked, and restoring the last two groups makes little difference to the result.
The main outcome is review count, primarily from Steam purchases, although some records use total reviews. Games with zero reviews stay in the analysis. Estimated sales are derived from reviews, so they provide an alternative measure, not independent evidence.
“AI” means the game had an AI disclosure at the snapshot date, not necessarily at launch. Likewise, followers in the opening section are current followers.
Log-scale charts omit zero values and report their share separately. Table medians include all games.
Historical data and last year’s comparison
Rechecking last year’s release window produces broadly similar patterns, though catalogue coverage and some measurements have changed. Last year’s 53% headline and this year’s 63% are not directly comparable: the model has changed, and this year’s main outcome is reviews to date rather than first-month reviews.
What the model measures
The model adjusts for:
-
Developer experience, measured by previous Steam releases.
-
Publisher backing.
-
Steam genres.
-
Release month and time on sale.
-
Differences between developers with multiple games in the dataset.
Price, ratings and followers are deliberately excluded from the main adjustment. They may themselves be affected by AI use or players’ reactions to it; controlling for them could remove part of the relationship I want to measure.
The model has two parts: whether a game gets any reviews, and how many it gets if it does. A log-normal distribution accommodates the large differences between ordinary releases and Steam’s biggest hits.
For the headline comparison, the model predicts every game twice, with and without an AI disclosure, while holding all other measured characteristics constant.
|
Measure |
Model estimate |
|
Average reviews without an AI disclosure |
544 |
|
Average reviews with an AI disclosure |
201 |
|
Difference in reviews for an average-game |
−63% (95% credible interval: −68% to −56%) |
|
Difference in reviews for a typical-game, among games receiving reviews |
−49.5% (−53% to −46%) |
|
Chance of receiving any reviews |
82.4% without a disclosure; 74.4% with one |
The top 1% of games hold 76% of all reviews, so the average is heavily influenced by hits. The typical-game estimate describes the middle of the reviewed-game distribution and is more stable across the checks below.
These are adjusted comparisons, but I must stress they are not proof that AI causes the effect.
Model selection and fit
The four originally planned models struggled to reproduce Steam’s combination of many zero-review games and a few enormous hits. The final model was added after my tests and performed substantially better, although it still slightly underpredicts the largest non-AI hits.
The acceptance threshold and one developer-level convergence criterion were relaxed after inspecting model performance; since it is the sales performance for a typical game that is most interesting I was less concerned about poorer performance for predicting the sales of non-AI hits.
These changes concerned model assessment, not the AI-effect specification. Across all five models, the estimated average reduction ranged from 56% to 63%.
Robustness checks
Twenty of the 22 planned variations passed their quality checks. Across them, the average reduction ranged from 53% to 72%, and the typical-game review reduction from 46% to 50%.
|
Alternative analysis |
Average reduction |
|
First-month reviews |
−60% |
|
Games on sale for at least 90 days |
−60% |
|
Excluding the 31 biggest games |
−62% |
|
Restoring rapid-release developers and $100+ games |
−65% |
|
Estimated sales instead of reviews |
−72% |
|
All reviews rather than Steam-purchase reviews |
−62% |
|
Adjusting for store-page content similarity |
−53% |
The first-month result was also similar using only our own snapshots. Content matching narrowed the observed reduction in reviews, suggesting that differences in the kinds of games being made account for some of the association. Because it uses current store-page content, it may also absorb consequences of AI use.
The two remaining planned checks were not relied on: the earlier-release-only model failed its quality checks, and the independently verified review-source subset was too small.
An additional, unplanned check excluded 2,013 games with reviews predating their listed release date, consistent with earlier early-access launches. These games account for 27% of reviews. Removing them narrowed the average reduction in reviews to 59%, while the difference for a typical-game remained about 49%.
Unmeasured differences
The “X-factor” sensitivity tests show how a hypothetical unmeasured factor could change the result. Under its assumptions, removing the reduction in reviews entirely would require a factor that both sharply reduces reviews and is almost confined to AI-disclosing developers. This is a sensitivity exercise, not evidence that unmeasured differences are absent.
This year’s analysis is a fresh calculation and not a directly comparable update of last year’s. A separate scenario assuming some undisclosed AI use did widen the estimated reduction in reviews, but that result depends on the assumed pattern of non-disclosure and I just can’t reliably justify that analysis.
Experience and types of AI use
For the average game, there is no clear evidence that developer experience changes the AI association with review counts. For a typical game, the reduction grows with experience: 47% for first releases, 54% for one or two previous releases and 59% for three or more (the 12-point gap between first releases and three or more has a 95% credible interval of 1.5 to 21 points). The reduction in the chance of receiving any reviews is similar across groups: about eight percentage points for first releases and seven for developers with three or more previous releases.
Previous releases measure experience on Steam, not skill, resources or budget. Seventy per cent of games in the dataset are first releases.
AI-use categories come from a classifier reading store-page disclosures. Each category is compared with games without a disclosure, using the same model. Three limitations in this classification model matter:
-
Categories overlap: a game may disclose art, code and music together, so category estimates cannot isolate or rank their individual effects.
-
Labels describe disclosed use, not an inspection of development files. Removing uncertain classifications changed no category’s average estimate by more than about ten percentage points.
-
Undisclosed use remains invisible, particularly for coding assistants.
Live-generated content was excluded from category comparisons because only 85 games disclosed it.
Game types and the Steam Map
AI disclosure appears across the market: between 25% and 35% of games in each Steam genre. It is more common in text-based games, card games, tabletop games and management-related tags, and less common in platformers and hand-drawn games.
The Steam Map groups games into market segments using their store text and images. Among segments with at least 100 games, AI disclosure ranges from 12% to 60%, although most sit much closer to the overall 30% share. AI games are a majority in only eight segments, containing under 2% of all AI games.
The outcome also appears within market segments: in 107 of 111 sufficiently large, mixed segments, a smaller share of AI-disclosing games reaches 1,000 estimated sales.
These comparisons describe where AI games sit in the market but they do not establish why their outcomes differ.
What games bring to launch
The groups differ before release as well as afterwards.
|
At launch |
AI disclosure |
No AI disclosure |
|
Developer’s first release |
76% |
68% |
|
Has a publisher |
14% |
22% |
|
Priced under $5 |
53% |
47% |
|
Has a demo |
27% |
33% |
|
Median store-page runway |
52 days |
167 days |
|
Median launch-day followers |
6 |
20 |
|
Fewer than 10 launch-day followers |
61% |
39% |
None of the data here is a direct measure of effort. They may reflect the studios resources, production choices, development speed or audience response, including reactions to an AI disclosure.
Launch-day followers use the last snapshot within 30 days before release. Follower counts were available for 27,471 games (89%).
Comparing games with similar launch positions
Simple comparisons match the groups within release quarters on Steam Map market segment, runway or launch-day following. Among games launching with at least 100 followers, the median AI game has 68 reviews against 77 for matched non-AI games; 63% versus 65% reach 1,000 estimated sales.
My exploratory model runs tell a similar story:
|
Adjustment |
Difference in reviews for an average-game |
Difference in reviews for a typical-game |
|
Main model |
−63.0% |
−49.5% |
|
Add store-page runway |
−36.2% |
−28.8% |
|
Add launch-day followers |
−2.4% |
−9.4% |
|
Add both |
−7.5% |
−16.0% |
Follower-adjusted runs use only games with a launch-day count. On that same subset, the reduction in reviews for the main-model remains 63%, so the narrowing is not simply a change in sample.
With followers alone, the average-game interval includes zero (−7.9% to +3.8%). With both measures, it is −12.8% to −1.9%. Adding the store-page runway widens the follower-adjusted reduction because, at the same follower count at launch, games that built it over a shorter runway tend to receive more reviews. It just so happens that AI games are more concentrated among this group.
These runs locate much of the difference in games’ position at launch. They do not explain its cause. If AI changes development time or discourages players from following, adjusting for runway or followers removes part of that effect.
What my study does, and doesn’t, show
Games that disclose AI use receive substantially fewer reviews than games without a disclosure, even after adjusting for several measured differences. The pattern persists across alternative analyses, but narrows sharply when comparing games with similar launch-day followings.
The study does not establish that AI causes the reduction in reviews, distinguish production choices from audience reactions, or measure skill, budget and marketing directly. It measures disclosed AI use, not all AI use, and review-based outcomes, not independently observed sales.

