AI Update Part 1: Steam’s AI flood doesn’t mean it’s game over for indies

AI slop, AI clones, AI dread… there’s plenty to feel uneasy about. But I don’t want this analysis to be another doom-and-gloom take on game development and AI’s role in it.

Fundamentally, I’m an optimist. I believe data can help us see reality more objectively, while our perspectives and lived experiences shape what that reality means to us personally. But neither data nor experience can tell us how to respond on their own. That’s where our philosophy comes in: guiding our actions, reminding us what’s within our control, and helping us decide what to do next.

I hope this analysis helps keep you informed and eases some anxieties about the world we live in right now. We’re going to deliver it across three parts:

1. This article: how much is AI being used right now? How has it changed over the last year? And does the wave of AI releases make being an indie impossible to the point where you should give up and move to the country to be a farmer? (Spoiler: no, please don’t do that)

2. Part 2: we’ll revisit the causal analysis we released last year and, again, ask the question “How does AI disclosure impact sales?”

3. Part 3: this is a new question, and one I never imagined I would need to ask (and frankly I am sad I have too), how worried should you be about your game getting cloned by AI? I’ll be using our Steam map and some carefully crafted ML models to identify what exactly is being cloned and what we can do about it.

Yes, there is more AI being used now

Let's start with the obvious. Yes, there are more games being released on Steam and yes, more of them are disclosing that they use AI.

Enjoying this article? Sign up to our newsletter for monthly market insights and the latest indie innovations

The proportion of releases with an AI disclosure back in January of 2025 was 15%, but now half of all new releases have some form of AI disclosure attached to their store page. In that same period, the number of new releases more than tripled; although that spike in October is likely a result of Next Fest. 

You’re probably already thinking, didn’t you say this wouldn’t be a doom-and-gloom article? But please, stick with me! As I said, this is just the data and an objective lay of the land. We’ll get to a more optimistic analysis later on.

How are people using AI?

Before I dive into this next section it goes without saying that we’re assuming people are being honest in their disclosures. Like I said in my original analysis last year, I think this is a fair assumption since Valve makes this a requirement for publishing your game. 

Valve has a disproportionate amount of control over whether you can access the PC gaming market at volume, and therefore lying to Valve would be stupid. However, people are irrational, so keep this in mind.

 

The chart shown above was created using, haha…AI. Well, let me be kinder to myself. I didn’t just point [insert favorite AI chatbot] at the AI disclosures and ask “how are people using AI?”. No, we used a very conservative methodology to try and avoid misleading statistics. To avoid going off on a tangent, I’ll leave the gritty details for the stats nerds in the "Methodologies" section at the end of this article. For everyone else, just know that we’ve done our best to try and not overstate AI use.

The results are pretty clear, and are not too different from our original analysis; AI is overwhelmingly used for art, marketing, and store page images. However, the proportion of disclosures that include some discussion of “human editing” has increased, as has the proportion of other use cases in disclosures; by “human editing” we mean that the disclosure includes some discussion around how AI generated the original asset and then a human edited the final result.

Let’s have a closer look at how exactly each disclosure type has changed.

 

The big takeaway from this graph is that art remains the biggest use case, but disclosure around code, localisation, and narrative text/dialogue has increased over this period. The biggest shift is the rise of “human-editing” remarks in the AI disclosure text, especially for art.

Perhaps this is driven by anxiety around how AI art is perceived. People might be hoping that showing a human intervention between the AI generation and the final product will help soften the blow? I’ll leave that up to your interpretation.

Has the wave of AI games saturated the market?

We know there is a wave of AI-generated (or at least partly generated) games flooding the Steam store, but is this making it harder for indie developers to be discovered and to reach their financial goals?

This is not a simple question to answer. First of all, what even is success? 

Without getting too deep here, I’m going to use a realistic milestone. I know that most indie developers publishing on Steam are releasing their first commercial game and in my (I will argue, informed) opinion, an objective of 1000 sales is both a healthy and realistic target for success. If you want further justification for this I point you to our previous analysis on indie success and longevity on Steam.

So with that in mind, let us reframe the quest to "As Steam releases have surged, have the number of games reaching 1,000 sales stayed fixed, or grown with the Steam releases?"

To answer this question, we took every monthly release cohort from January 2025 to September 2026 in which all games are at least 30 days old (that is 21 months and 36,228 Steam releases) and we fitted a log-log regression model on the number of games with at least 1,000 estimated sales against releases that month. 

The slope of this model (shown in the graph below) shows whether games selling 1000 copies stay fixed or scale with the volume of releases. If the number of games selling 1,000 copies "stays fixed," then twice as many releases still produce the same number of successes, so each game's odds of being a success halve. 

However, if it "scales with volume," twice as many releases produce about twice as many successes, so each game's odds of selling 1000 copies stay the same.

 

In this graph, each dot is one month of Steam releases. It shows how many games came out that month and how many of them went on to reach 1,000 sales, and the dots get lighter for more recent months. 

If every month produced the same number of successes no matter how crowded it was, the dots would lie along the horizontal dotted line. If each game kept the same odds, they would follow the steeper dashed line.

But, the trend actually falls between the two and sits closer to the steeper line, so busier months produced more successes, though not quite in proportion to the extra releases. The most recent months had the most releases, yet still produced more games reaching 1,000 sales despite having less time to get there. Overall, this suggests that the number of games finding an audience is growing, not staying fixed.

More releases have brought more successes, but each game's odds of selling 1,000 copies have still slipped a little, from roughly 1 in 5 to 1 in 6, and some of that slip is just because the newest games haven't had as long to sell.

We also compared the first three months of 2025 with the latest three and re-checked the results several ways. 

As releases roughly doubled, the number of games reaching 1,000 sales rose from 763 to 1,311, but the proportion reaching 1,000 sales fell from about 20% to 16.5%. Part of that drop is because newer games have had less time to sell: counting only games at least six months old, the gap narrows to about 19.6% vs 18%. 

Steam isn't a fixed pool of winners split ever more thinly, but this can't tell us whether AI-made games are squeezing anyone out.

TLDR: Steam's flood of new releases hasn't left a fixed number of hits that have to be shared out amongst developers! There are more games than ever reaching 1,000 sales, even though each game's odds of doing so have slipped slightly.

In conclusion, keep your chin up!

There’s plenty here to feel uneasy about, but I’m leaving this analysis more hopeful than when I started. 

AI disclosures are becoming more common and Steam is getting busier, yet the number of games reaching 1,000 estimated sales is growing too. The opportunity for game devs isn’t a fixed pie being sliced ever thinner. 

Now, of course that doesn’t mean making a living from games is easy, or that AI has no effect on the competition, we haven’t answered those questions here, and each game’s odds of reaching that milestone have slipped a little. But it does give us a reason to resist the idea that it’s already too late. 

My takeaway is to keep our ambitions grounded, stay curious about the market, and put our energy into the things we can influence: what we make, who we make it for, and how we help those people find it. 

There’s still room to make something worth playing, and that’s a future I’m excited to be part of. See you in part 2!

Methodology (for the nerds)

A quick look under the hood of where these numbers come from, how we sorted AI disclosures into categories, and how we checked the "the market is not saturated" claim.

The Data

Everything comes from the Game Oracle Steam catalogue, frozen as a snapshot on 7th November 2026. Steam keeps changing, so the same queries run today would give slightly different answers.

From each game we used three things:

• The AI Disclosure: since 2024, Steam has asked developers to say whether and how they used generative AI. We used the free-text statement shown on the store page.

• The release date: we grouped games into monthly cohorts by their store release date.

• Estimated sales: Steam doesn’t publish sales, so these are Game Oracle’s model estimates. Each one comes with a low and a high bound. We treated "reached 1,000 units" as a rough marker of a game finding an audience. It says nothing about whether the game made money.

That gives us two groups of games:

• Everything with a disclosure: 27,894 games with non-empty AI disclosure text, released or not. This is the group behind the "what do developers use AI for" charts.

• Releases by month: 40,089 games released between January 2025 and October 2026. We left out games still marked "coming soon", and games with no release date can't be placed in a month at all. This is the group behind the release-trend and saturation charts.

One caveat applies to every time-based chart: we only see each disclosure as it reads today, at the time of writing. A game released in early 2025 might have added or changed its disclosure later. So "disclosures among March 2025 releases" really means "March 2025 releases that carry a disclosure at the time of writing".

Classifying the disclosures

Disclosures are free text, and they're messy. Some are a single word ("Marketing"), some are paragraphs, and plenty are vague. We wanted to know which parts of a game each one covers, so we used nine categories: Art, Text/dialogue, Localisation, Music, Voice, Sound effects, Code, Marketing/store page, and Live-generated content.

• Used AI: the disclosure says AI was used here.

• Used AI, but human-edited: it says a person reworked the AI output. Saying someone “reviewed”, “selected” or “directed” the output wasn’t enough for this label.

• Not disclosed: the disclosure doesn’t say AI was used here. This doesn’t mean the game didn’t use AI for that category. It only means the statement doesn’t say so.

A single disclosure can cover several categories, so the category totals overlap and don’t add up to 100%.

How the labels were made

We labelled a sample of 1,000 disclosures, with an AI model doing the first pass. I then checked 100 of them by hand, and then we audited the whole set against a few house rules so that similar cases were labelled the same way. For example:

• Art made only for the store page counts as marketing, not game art.

• “We used AI for audio” with no detail counts as music, voice and sound effects.

• A vague “we used AI” doesn’t get marked as every category.

The model

We fine-tuned a small open-source text classifier, GLiClass (https://huggingface.co/knowledgator/gliclass-base-v3.0), on those labels and ran it over all 27,894 disclosures. The model only sees the disclosure text, nothing else about the game.

How well does it work?

We kept the 100 hand-checked disclosures out of training and tested on them:

• It picked the right label for a category 96% of the time, against about 79% for a lazy baseline that always guesses the most common label.

• It got all nine categories right on 73% of disclosures.

• It’s weaker on some things. It caught only 4 of the 9 text/dialogue uses in the test set, and the test set had no live-generated examples at all, so treat the live-generation numbers as rough.

To be honest about the limits: 100 test examples is a small check, and those same 100 helped shape the labelling rules, so the test isn’t fully independent. The model also flags about 4,300 disclosures (15%) where it was unsure between two labels. Leaving those out moved no category by more than about 3 percentage points. Every category figure in this post is a model estimate, not a hand count.

Is the market saturated?

The worry goes like this: AI makes games cheaper to make, more games come out, and a fixed number of players gets split across more and more titles. If that’s true, the number of games that do well each month should stay roughly flat however many come out, and each game’s chances should shrink.

We tested that with monthly release cohorts:

1. Count releases and hits each month. A “hit” is a game with at least 1,000 estimated Steam sales by the snapshot date.

2. Drop months that are too recent. Sales add up over time, so a game released last week hasn’t had a fair chance. We kept a month only if every game in it was at least 30 days old. That cut October 2026 and left 21 months (January 2025 to September 2026) and 36,228 releases.

3. Measure how hits scale with releases. We fitted a simple curve: when releases go up by 1%, how much do hits go up? We call that number β.

• β = 0 would mean a fixed number of hits each month, whatever the number of releases. That’s the “fixed pie” world and the worst case scenario.

• β = 1 would mean hits grow in step with releases, so each game’s odds stay the same.

• Anywhere in between means more releases bring more hits, but each game’s odds go down.

4. Adjust for volume. Using β, we scaled each month’s hit rate to what it would look like at January 2025’s release volume. Overall, this shows whether the hit rate fell by more than the growth in releases alone would suggest.

What we found

β came out at about 0.72, with a range of roughly 0.47–0.96 when we resampled the months. That’s well away from zero, so there’s no fixed pie. More releases came with more hits; there were 763 in the first three months against 1,311 in the last three months. The share of games that became hits did fall, from about 20% to 16.5%, but after the volume adjustment it’s roughly flat, at about 20% at both ends.

Checks we ran:

• Defining hits with the low and high sales estimates instead of the central one barely moved β (0.72 and 0.74).

• Requiring 90 or 180 days of age instead of 30 left fewer months and gave wider ranges, but told the same broad story.

• Adding a time trend to the fit pushed β up to about 0.94. Release volume and calendar time rise together, so it’s hard to tell them apart. That’s a good reason not to read too much into the exact number.

What this does and doesn’t show

It does show that “more releases means a fixed number of winners” doesn’t fit the data. It doesn’t show that AI games have no effect on the competition. We didn’t compare AI and non-AI games here, and this isn’t a causal analysis. It also can’t tell you whether developers are making more money, because 1,000 units is a milestone, not revenue. The volume adjustment uses a relationship we fitted to the same data, so a flat adjusted line is a description, not proof. And 21 months is a short series to lean on.