Games

Games and AI: Who is training whom?

Video games have always been a proving ground for demanding new technology. One of the world’s most valuable AI companies spent its first two decades making graphics chips for gamers, and in a full-circle moment the relationship between the sectors now runs both ways. Gameplay is prized training data for a new class of “world models,” while generative AI is redefining game production, storefront policy, talent negotiations, and even the price of hardware. For games businesses, the challenge remains in managing the rights, obligations, and risks this brings across the value chain.

Games as a training ground for world models

Game engines were always expected to produce high-quality synthetic training data, given their capabilities in generating graphical content. The more significant development is that AI labs increasingly treat games themselves as the training ground. Microsoft’s Muse model was trained on years of human gameplay from a single title, and Google’s Genie generates interactive environments from a prompt. AI-generated assets and voice are increasingly common in games, and conversational characters are beginning to ship in commercial releases as AI-driven non-playable characters. For rights holders, this brings familiar questions into an unfamiliar context. Who owns gameplay footage and player telemetry? On what terms can it be licensed for model training? Do player-facing terms and privacy notices cover that use? Gameplay data is a genuine asset, but capturing its value means securing and reserving the underlying rights, auditing the consents on which it was collected, and setting licensing terms before it trains third-party models. 

Performers and digital replicas 

One of the most consequential games-specific developments for talent is the resolution of the SAG-AFTRA interactive media dispute. After an 11-month strike, members ratified a new Interactive Media Agreement in July 2025. It requires clear consent and disclosure before a performer’s voice or likeness is replicated, compensation linked to use, and the right to suspend consent for new AI-generated material during future strikes. Even without union exposure, these terms are fast becoming a market baseline for talent contracts. The risks were illustrated in 2025 when a major online title shipped a conversational character built on an AI recreation of a late actor’s voice, with his estate’s consent. Players manipulated it into producing offensive content, and the union filed an unfair labor practice charge over the replacement of performer work with AI. Its primary claim was that the title’s production company, a signatory to the union’s agreement, had replaced work previously performed by union members with an AI voice without notice or bargaining, and the union supports the right of members and their estates to control their replicas. Studios deploying replicas therefore need contractual controls over use, safeguards against player misuse, and clear liability allocation when a replica behaves in ways the rights holder never intended. Even where consent has been obtained, backlash and scrutiny remain likely, unless and until player and consumer sentiment toward generative AI changes.

Storefront disclosure rules

Distribution platforms have moved faster than legislators. Steam has required disclosure of AI-generated content since January 2024 and rewrote its rules in January 2026 to focus on player-facing content, distinguishing pre-generated from live-generated material, carving out development tools, and adding an in-game mechanism to report unlawful live-generated content. The number of games disclosing AI-generated content of either kind grew from around 1,000 in 2024 to several thousand by mid-2025. Google Play has required in-app reporting of offensive AI output since January 2024, and Apple’s guidelines have required disclosure and consent before apps share personal data with third-party AI since November 2025. Other storefronts have declined to impose labels at all, meaning disclosure duties vary by distribution channel. Statutory obligations are now also layering on top, with China’s AI labeling rules taking effect in September 2025 and most of the EU AI Act’s Article 50 transparency duties now in force. The European Commission’s final guidelines and the voluntary Code of Practice arrived only weeks before that start date, so much of this detail is new and untested.

For games businesses, the biggest challenge is how to comply in practice. Often, two separate duties are engaged. Users must be told when they are interacting with an AI system, so a conversational AI character must tell players it is artificial unless that is obvious from context, and supporting guidelines interpret that exception narrowly. Separately, businesses deploying AI must disclose any “deepfake,” meaning an AI-generated or altered image, audio, or video that resembles a real person, place, or event and could pass as authentic. An intent to deceive is irrelevant. Synthetic voices, a de-aged or resurrected performer, or a fully AI-generated actor can therefore constitute a “deepfake” even where all underlying consent has been obtained, and failure to disclose can result in fines of up to €15 million or 3% of worldwide turnover.

Consumer and developer sentiment

Consumer and developer sentiment toward AI in games has hardened even as adoption has grown. In the 2026 GDC State of the Game Industry survey, 52% of developers said generative AI is having a negative effect on the industry, up from 30% and 18% in the two years before, though over a third use the tools in their work. Players have review-bombed titles over AI-generated assets, and undisclosed AI art in games and marketing has repeatedly become a reputational and commercial issue. Where and whether to use generative assets is a commercial as much as a regulatory consideration, engaging brand trust, quality control and, in an increasing number of contexts, advertising and consumer law.

AI and the price of hardware 

AI demand is repricing game hardware itself. Memory makers have reallocated capacity to high-bandwidth memory for AI data centers, and memory prices rose by around 80–90% in the first quarter of 2026 alone. Consoles and PC components have seen unusual mid-generation price rises, and analysts do not expect meaningful relief before late 2027, with knock-on effects on install-base growth, hardware margins, peripheral supply chains and, potentially, software pricing. Buyer leverage is limited while demand outstrips supply, so the realistic aim for games hardware manufacturers is to mitigate exposure rather than shift risk entirely. Procurement teams should aim to scrutinize price adjustment clauses, cap or index permitted increases where they can, lock in volumes and firm lead times with delivery remedies, dual-source critical components, and review force majeure and change-of-law provisions so that shortage risk does not sit wholly with the buyer.

IP: Training data, ownership, and clones

Publishers’ concerns about third-party models training on their intellectual property have moved from principle to action, and trade bodies are increasingly willing to act on their members’ behalf. One console platform holder said publicly in October 2025 that it would act against infringement of its intellectual property whether or not generative AI was involved. Just weeks later, a Japanese rights-holder body representing several major publishers wrote to a leading model provider asserting that the reproduction of protected works during training may constitute infringement under Japanese copyright law.

Publishers are also strengthening contractual protections. Reservations of rights and prohibitions on using game content to train models are increasingly common in player terms, and the Unreal Engine license terms have barred use of the licensed technology for training or prompt input since 2024. At the same time, studio adoption is accelerating. Where concerns that AI-generated assets could attract no copyright protection once kept generated outputs out of shipped games, commercial practice has markedly moved on. A census of 53,597 Steam releases found that the share displaying an AI disclosure rose from 10.9% in 2024 to 19.9% in 2025 and 30.8% for 2026 to July, with around a fifth of disclosures on higher-earning titles added after release. Large publishers have since announced AI-first strategies and quality assurance automation targets publicly, and studios are increasingly willing to disclose what they used and where. 

Cloning has also become cheaper and faster. AI coding assistants can turn a short gameplay clip into a reasonably playable copy in hours, and clones of indie prototypes have appeared on storefronts within days of a designer posting footage – in some cases before the original has even launched. Analysts warn that good games will be cloned and buried. Holding prototypes and public displays back until the store page is ready is therefore now a commercial decision as much as a marketing one.

What to do now

These developments make AI a value-chain issue for games businesses, spanning the data and performances that go into a game and the storefronts and hardware through which it is sold. In practice:

• Audit where generative AI sits in the development pipeline and map the outputs against storefront disclosure rules and labeling laws in each territory.
• Decide whether and how AI characters and any deepfake content will be disclosed to players, channel by channel, before launch.
• Reserve training rights expressly, in player terms and in machine-readable form on the studio’s own sites.
• Keep records of human contribution for every AI-assisted asset the studio intends to protect. In many jurisdictions, sufficient human authorship remains the price of copyright protection.
• Bring replica consent, AI-use disclosure, and human authorship terms into talent, development, and publishing agreements, and expect unions to bargain over replica use even where an estate has consented.