The AI Deadline Nobody In Games Is Talking About Lands On 2 August With Real Fines Attached
Article 50 does not care what Steam thinks
EU AI Act Article 50 transparency obligations take effect 2 August 2026 and were not delayed by the Digital Omnibus
Fines reach 15 million euros or 3% of worldwide turnover, and UK and US studios serving EU players are in scope
The artistic works carve-out covers video games but attenuates the disclosure duty rather than removing it
Marketing material and store pages do not get the same light touch the game itself receives
For two years the games industry has been having an argument about whether telling players you used AI is an act of honesty or an act of self-harm. Valve built a disclosure survey into the Steamworks submission flow. Tim Sweeney called mandatory disclosure irresponsible and compared it to telling players which shampoo the developers use. itch.io started tagging. GOG declined to take a position and then got caught putting AI art in its own sale banner. Everyone found a side.
While that argument ran, the European Union finished writing a law that does not care who won it. It takes effect on 2 August. If your game is playable in the EU, it reaches you wherever your studio sits, and the fines run to 15 million euros or 3% of worldwide turnover, whichever is higher.
I have not seen one piece in the games press treating this as a developer obligation. Plenty of coverage of Steam’s survey. Plenty of coverage of Sweeney calling it a scarlet letter. Almost nothing on the only AI disclosure regime in this whole story that comes with an enforcement mechanism.
If this is useful, there's more where it came from. Multiple free posts a week on game production, applied AI, leadership and how the sausage actually gets made.
The misread that is going to catch people
In June the headlines said the EU AI Act had been delayed, and a lot of teams read that and stood down. The European Parliament approved the Digital Omnibus amendments by 423 votes to 57, the Council gave final approval on 29 June, and the package was signed on 8 July. Standalone high-risk systems moved from August 2026 to December 2027. High-risk systems embedded in regulated products moved to August 2028. Sixteen months of headroom, handed out in a single vote.
None of that touched Article 50.
The delay everyone celebrated was for high-risk systems. Article 50 was never in that category.
Article 50 is a transparency layer that applies to particular categories of AI regardless of risk classification. It was never part of the high-risk regime, so postponing the high-risk regime did nothing to it. Gibson Dunn put it plainly in their read of the Omnibus: 2 August 2026 remains a live compliance date, and the Article 50 obligations are largely unaffected. The one concession is narrow. Systems already on the market before 2 August get until 2 December to implement machine-readable marking. Everything else lands on schedule.
So the position a lot of studios are in right now is that they heard “delayed”, filed it, and moved on. The bit that got delayed was the bit about credit scoring and biometric employment screening. The bit that touches games did not move at all.
What it actually asks of a game
Three parts of Article 50 matter if you ship games.
The first is deepfake disclosure. Where an AI system generates or manipulates image, audio or video content that would falsely appear authentic, the deployer has to disclose that it was artificially generated or manipulated. The definition turns on whether the content resembles real persons, objects, places or events and would pass as genuine.
The second is marking. Providers of generative systems have to ensure outputs are marked in a machine-readable format that allows detection of artificial generation. If you built the model, that one is yours. If you licensed it, it belongs to whoever provided it, though I would want that in writing rather than assumed.
The third is interaction disclosure. If a player is talking to an AI system, they have to know they are talking to an AI system. Studios shipping live language models behind NPC dialogue are in scope for this in a way that studios shipping pre-generated barks are not.
Two things are worth knowing about the edges. Ordinary editing sits outside the scope, so spellchecking, grammar correction, quality improvement and format conversion do not pull you in. And open-source systems get no exemption, which surprises people who assumed running weights locally kept them out of it.
The reach is the part I would not gamble on. The obligation follows the market rather than the company, so a studio in Guildford or Austin or Ho Chi Minh City is in scope the moment an EU player can buy the game. There is no version of this where you are too small or too far away. A self-published team of four with a Steam page that sells into Germany is captured by the same provision as a publisher with a legal department, and the smaller outfit is the one with nobody to ask.
The carve-out, and exactly where it stops
Games do get special treatment, and it is more useful than most studios expect.
Where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work, the disclosure duty is limited to disclosing that the generated content exists, in a manner that does not hamper the display or enjoyment of the work. Bird & Bird’s reading of the Commission’s draft guidelines names video games directly and describes what this looks like in practice: an unobtrusive label in the credits or a loading screen rather than an immersion-breaking in-frame disclosure. Nobody is going to make you put a watermark across your character models.
Then the limits, which is where the useful detail lives.
The guidelines clarify that clearly fantastical or physically impossible content falls outside the deepfake definition entirely. Dragons do not trigger it. Humans flying unaided do not trigger it. The test is whether a reasonable viewer could take the content for real, and a wyvern fails that test comfortably.
Dragons are fine. A photorealistic person who never existed is not.
A photorealistic invented human sits squarely inside the deepfake definition. Bird & Bird call this the fictional-but-plausible category, and it is the one that catches modern character art. The carve-out then softens the obligation for that character because the game around it is evidently fictional, but the obligation still exists. Covington’s read of the guidelines is direct on this: the treatment is attenuated but not eliminated.
The word doing the heaviest lifting is “evidently”. Content that is ambiguous to the audience is excluded from the carve-out. A game is an obvious fiction. A piece of marketing material that looks like a photograph is not.
Which brings me to the part I think will catch studios hardest. The Commission’s guidelines exclude from the artistic allowance any content that serves primarily an informative or commercial purpose and is recognisable as such. Your game is a creative work. Your trailer, your store page screenshots, your key art and your ad creative are commercial material, and they do not inherit the game’s carve-out.
Your game may get the light touch. Your trailer will not.
A studio could therefore end up in the position of having a compliant game and a non-compliant marketing campaign, built from the same asset library, by the same team, in the same week. I have not seen anyone in games raise that yet and it seems like the most likely way a competent studio gets caught.
This is a production problem wearing legal clothes
Here is the part that makes this my business rather than a lawyer’s. I have taken issue with legal advice on AI in games before, not because the law was wrong but because the advice addressed a fraction of how studios actually use the tools. The same gap is open here.
Every obligation above is retrospective. Article 50 lands on projects that have been in development for three or four years, built by teams that rotated, using tools that changed twice, with outsourcers who came and went. The law asks a simple question: which content in this build was AI-generated. For a large number of studios that question has no answer, because nobody was ever asked to record it.
I have argued before that “was AI used” is the wrong question, because it collapses several different questions into one and gives you an answer you cannot act on. Article 50 makes that concrete. The regulation cares where the output lands and whether a player could take it for real. A binary yes on a survey form tells you nothing about either.
You cannot mark what you cannot identify. You cannot decide whether the carve-out applies to an asset if you do not know the asset is in scope. You cannot tell a publisher which parts of the build need disclosure if the only record is somebody’s memory of a Slack thread from 2024.
You cannot disclose what you never tracked.
This is the same failure I keep running into on the licensing side, and it has the same shape. The underlying obligation is to be able to say what happened. Disclosure is only the visible end of it. A studio that kept a record has a compliance task in front of it. A studio that kept nothing has an archaeology project, and eleven days.
The ordinary editing exclusion makes this sharper rather than softer. Somebody has to decide where the line sits between a quality improvement and a generation. That decision needs to be made by a person, recorded, and defensible later. Which means it needs a process, and process is production work.
Thirty-plus years of production work went into this. Most of it didn’t make the post. New posts every week. Free to read, free to subscribe.
There is an ownership problem underneath this that studios will discover the hard way. Compliance obligations tend to arrive addressed to nobody. I wrote a few months ago about how long Godot went without anyone deciding it had an AI question to answer, and the interesting part was never the policy they landed on. It was the months in which the question sat there belonging to no one. Legal reads the regulation and assumes the content question belongs to production. Production assumes anything with a fine attached belongs to legal. Art knows which assets were generated and has never been asked. The outsourcer knows what they delivered and was never told it mattered. Every one of those groups holds a fragment, and the fragment is worthless on its own.
The studios that handle this well will be the ones where somebody is named. Not a committee, a person, with the authority to ask the awkward question of the art lead and the outsourcing manager and get a straight answer. That is a producer’s job by any reasonable reading, and it is the kind of work that gets done badly when it is treated as an administrative chore rather than a delivery risk with a date on it.
What I would actually do this week
None of this is legal advice and I am not the person to give it. It is production advice, which is a different thing.
Establish scope first, because a lot of studios will find they are barely in it. If your AI use was tooling, code assistance and internal workflow, and no generated content reaches the player, your exposure is small. If you shipped generated character art, generated voice, or a live model behind dialogue, you are in.
Then find out what is actually in the build. Ask the leads, ask the outsourcers, and write it down in one place. This is the discovery sweep I bang on about, and the same record answers your publisher, your licensor and your storefront, so the work is not single-purpose.
Decide where the disclosure lives before somebody decides it for you. Credits and loading screens are the shape the guidelines point at, and choosing deliberately is better than having marketing bolt something onto the store page in a panic on 1 August.
Split the game from the marketing in your own head, and check the marketing separately.
Then start recording from today, even where the past is unrecoverable. Partial evidence with a date on it beats a complete absence, and the next project inherits whatever habit you set now.
Platform relationships have always been more contractual than developers expect, and this is a reminder that the storefront is only one of the parties you answer to. The storefront argument was never the important one. Sweeney can be entirely right that a disclosure tag on a store page tells players nothing useful, and it does not matter, because Article 50 is not a storefront policy you get to argue with. Valve can keep narrowing its rules to player-facing content only, and that narrowing has no bearing on what a regulator asks you for. One in three new Steam releases now carries an AI disclosure, by one analysis of roughly 53,600 releases since the survey launched, and none of those disclosures were written to satisfy Article 50.
If this changed how you think about even one thing, the next post might too.