For years, personalization in entertainment meant pointing viewers to the best existing option.<\/p>\n
A recommendation engine could steer someone toward the closest match, but it could never build something that did not already exist. Generative artificial intelligence is starting to close that gap, creating content around one specific person instead of matching that person to something built for everyone.<\/p>\n
Fortnite<\/a> shows what that looks like in practice. The game\u2019s AI-powered non-player character feature, called conversations<\/a>, exited experimental testing and began rolling out to island creators on July 30, according to an announcement<\/a> on the company\u2019s website.<\/p>\n The conversations feature replaces scripted dialogue trees with characters who talk, react and respond to a specific player\u2019s actions, rather than repeating the same lines to every player who approaches them, according to an April announcement<\/a> on the company\u2019s website. A developer defines who a character is through simple prompts, and the character generates a new response in real time based on what the player says or does.<\/p>\n Kaon AI<\/a> applies the same logic to narrative video<\/a>. The startup generates a distinct story world for each viewer in real time, rather than recommending from a fixed library. It has raised $60 million in Series B funding on that approach.<\/p>\n \u201cIf the perfect piece of content for the user was never made, the algorithm cannot help the user,\u201d Kaon AI CEO and Co-Founder Jay Dang<\/a> told PYMNTS<\/a> last month.<\/p>\n Netflix and Spotify Are Personalizing Ads and Audio<\/p>\n Netflix<\/a> is applying similar rationality to advertising. The company has tested AI-generated ad creative dynamically matched to the visual world of the specific show a viewer is watching, rather than inserting the same generic ad break into every stream, Adweek reported<\/a> May 13.<\/p>\n \u201cBrands including DoorDash<\/a>, Target<\/a> and TurboTax<\/a> have tested the tool, and Netflix said it has significantly improved both quality and execution,\u201d the report said. \u201cThe company plans to expand it to all ad-supported regions by year\u2019s end.\u201d<\/p>\n Netflix is also testing personalized ad loads and frequency caps that adjust which ads a member sees based on their own viewing behavior, PPC Land reported<\/a> May 14.<\/p>\n Spotify<\/a> is pursuing a version of the same idea in audio. At the company\u2019s 2026 Investor Day, Co-CEO Gustav S\u00f6derstr\u00f6m<\/a> described the direction as a clear path from access to personalization and now generation, according to a May 21 recap<\/a>.<\/p>\n Spotify is building a Large Taste Model, trained on 3.4 trillion daily signals from listener activity across music, podcasts and audiobooks, to generate individualized audio and remixes rather than only recommend existing tracks, the recap said.<\/p>\n The Harder Problem Is Knowing How and What to Generate Next<\/p>\n Disney<\/a>\u2019s experience shows why that bet is harder to execute than it sounds.<\/p>\n In a December press release<\/a>, OpenAI<\/a> announced it would receive a $1 billion equity investment from Disney to let Disney+<\/a> subscribers generate fan videos using Disney, Marvel<\/a>, Pixar<\/a> and Star Wars<\/a> characters through OpenAI\u2019s Sora<\/a> tool.<\/p>\n OpenAI shut down<\/a> Sora entirely three months later, citing a pivot toward high-productivity coding and agentic tools.<\/p>\n Disney withdrew before any money changed hands, Deadline reported<\/a> March 24.<\/p>\n The deal had included specific guardrails. Characters could appear in generated videos, but talent likenesses and voices were excluded, according to the December press release.<\/p>\n Disney and OpenAI had created a joint steering committee to monitor user-generated content against a brand appendix defining what Disney would allow, Axios reported<\/a> Dec. 11.<\/p>\n Generating something new for every viewer costs more than pulling something off a shelf. Every one-off piece of content still needs compute, moderation, and a way to catch it when it goes wrong, at a scale no fixed library ever had to manage.<\/p>\n