Home / AI Game News / Takes
A Game Can Disclose AI and Still Hold 91% Positive
Tim Sweeney's twenty posts call Steam's AI label a scarlet letter that crushes studios; the 2,424 new releases we counted, and the reviews of three AI-disclosing games, put the damage in break rates, save files and open microphones instead.

This piece responds to 20 posts on AI and games that Epic Games CEO Tim Sweeney made on X between November 2025 and November 2026.
Tim Sweeney's twenty posts, written between November 2025 and November 2026, treat Steam's AI disclosure as a weapon aimed at developers. The sharpest one, from November 3, says a developer can 'Lie on the question and thrive, or disclose and have your studio's livelihood crushed by a cancel mob.' We agree with two of the posts and not with that one. Sweeney is right that AI 'will be involved in nearly all future production', and right that content piled up without creative or architectural direction ships a bad product. Treating the label as 'a cancerous beacon to attract haters' does not survive the review sections we read.
The strongest version of the argument is the one about craft. On October 18 Sweeney writes that developers who use AI to 'blindly build up bodies of code or content without providing creative or architectural insight' end up with 'either an unshippable mess or shipping a bad product.' Our count points the same way. Of 2,424 Steam games released since August 1, 2026 with at least 20 reviews and a checked store page, 720 disclose AI-generated content, or 29.7%. The share that reaches 500 reviews is 9.7% among them and 17.2% among the rest. That settles the factual half of the argument. AI is already ordinary in shipped games, and the group that discloses it reaches fewer players.
The blind spot sits in Sweeney's June 26 argument that food lists ingredients because they directly affect the buyer, whereas AI development tools are not comparable. None of the posts brings any Steam market or review data, and the only named production test is converting Fortnite's blueprints to a new engine approach. Our angle is to split game AI into two books that affect buyers differently. One book is development-time assets, made once before launch, where the store label tests whether low-cost output got enough design polish. The other book is live runtime generation, which Sweeney praises for dynamic NPC dialogue and which in practice brings cloud token bills, API outages and microphone uploads into a buy-to-play game. Sweeney writes on December 27 that a videogame NPC emulating human emotion should be clearly AI to the player, yet the disclosure posts treat the store label as if every AI game only used offline development tools.
Disclosed AI games rate six points lower, and players still show up
The median positive rate is 81% for the 720 AI-disclosing games and 87% for the 1,704 without a disclosure. Six points is a real gap and a long way from a livelihood crushed by a mob.
| Discloses AI-generated content | No disclosure | |
|---|---|---|
| Games with 20+ reviews | 720 | 1,704 |
| Median positive rate | 81% | 87% |
| Median review count | 57 | 84 |
| Share with 500+ reviews | 9.7% | 17.2% |
| Games past 1,000 reviews | 33 | 182 |
| Median positive rate, 1,000+ reviews | 76% | 85% |
Tasty Chef is the clean test of the disclosure line. Its store page says generative AI was used for all visual art assets and all music, and that the game generates no AI content during play. It holds 91% positive. A player who recommends it writes 'Dont let the use of ai for assets stop you from purchasing this game - the devs are nailing the game play and they are active on discord.' Another opens with 'First of all, I appreciate the warning of AI generated content.' The complaints that do land are about the quality of the art. One negative review wrote that the visuals show not a trace of human work. A positive review asked why, if AI made the pictures, they could not at least look better. Those are complaints about plain art, and the game is at 91%. Its reviews spend far more words on recipe variety, staff behaviour and the gacha-style hiring than on the label.
High-profile AI games lose their ratings to break rates and save files
ForgeTax discloses AI-generated content for store images, some audio, a limited number of 3D assets and localization, and states that no live AI runs during play. It sits at 54% positive across 2,046 reviews. The negative reviews we sampled focus on the break percentage. One opens with 'The probability in this game is complete garbage.' Others say the break chance fires far more often than the number on screen. Patches drew complaints too. An October review objected to an update changing item attributes hours after purchase, and three November reviews say an update wiped their save files. None of the complaints we read is about the disclosure line. The defence is not about AI either: 'People who complain about RNG are just bad at math', and 'Fun gameplay loop, active dev, lots of potential.' Both sides argue about numbers and content, which is the failure Sweeney describes when AI is pointed at output rather than at a design. The label on the store page neither caused that nor hid it.
Real-time AI turns the label into a service description
Scam With Your Friends discloses AI-generated content for real-time dialogue and responses from fictional characters. It has 5,341 reviews at 76% positive. Several sampled reviews describe the game holding the microphone open, recording in the background and running after the window closes. One writes the developers are 'begging for donations via discord and inside the game because they can't afford the ai.' Another traced the audio path and found speech recognition running on the local machine while transcripts go to online AI services for the reply. Sweeney's October 18 post makes dynamic conversation the creative opening, with designers writing a character's bio and mannerisms instead of a dialogue tree. The players here enjoy the calls and still want to know where the audio goes: 'I'm excited about this game's AI, but I want clear answers about whether my voice is recorded.' A player who opened the game files wrote that the browser history it reads belongs to the fake desktop inside the game, not to Chrome or Edge. Another review gets both sides into one line: 'very awesome very cool idea very good use of ai LLM, however the loop goes stale really quickly.' The model is thin enough that a player noticed: 'you soon realise the AI is very basic.' A game that keeps a model on the other end of a microphone is selling a service, and the store page is where the buyer can learn the terms.
Where this could be wrong
Disclosures are written by developers, and the 1,704 games that say nothing may include plenty that use AI without mentioning it, so the path Sweeney calls 'Lie on the question and thrive' is invisible in these numbers. The count stops at games with 20 or more reviews. A player who reads the label and never buys leaves no review, so a loss at the point of purchase would not show up in any positive rate here. Rating and reach gaps are correlations. Studios that lean on AI tend to be smaller and poorer, and small studios reach 500 reviews less often whatever tools they use. We read samples of 48, 63 and 75 reviews for the three games, not the full review sets. The data covers August 1 to November 19, 2026, and AI tools change faster than a quarter.
For anyone shipping a game with AI in it
Decide which book you are in before writing the store page. If AI only touches development-time art, audio or code, the disclosure is one line and the loop is what gets judged; Tasty Chef's reviewers spend their words on recipes, staff and hiring. If a model runs while someone plays, put three things on the page. State who pays for the tokens, what the game does when the API is down, and whether voice is stored or sent to a third party. Sweeney's own test for AI work applies to both books, and it is the clearest sentence in the thread. When nobody on the team can say what the tool is directed at, the reviews will say it instead.
