Startups push back on App Store rules that limit experimental on‑device AI

The fight is about a very specific kind of AI app. These tools let you describe an app in plain English, then the model builds it and often lets you preview the result right on your phone. That sounds like a natural fit for on-device AI. But on iPhone, it runs straight into one of Apple’s oldest rules — apps are not supposed to download or execute new code that changes what they do after App Review. Apple has now started enforcing that rule against several “vibe coding” startups, and the clash is exposing a bigger problem in Apple’s AI strategy. ### What are these apps actually doing? Replit, Vibecode, Anything, and now Lovable all pitch some version of the same idea: tell the AI what you want, and it generates a website, tool, or app for you. The appeal is obvious — you do software creation from a phone with text or voice, and the system handles the code. For founders, that is not just a cute demo. It is a new computing model where the phone becomes a workshop, not just a storefront. ### Why does Apple care so much? Because Apple’s review system depends on the app staying basically the same app after approval. Guideline 2.5.2 says apps must be self-contained and cannot download, install, or execute code that introduces or changes features or functionality. From Apple’s point of view, if an approved app can spin up brand-new apps that could create something harmful, sideload it, and then imply Apple had already vetted it. ### Who has been hit? The names that keep coming up are Replit, Vibecode, and Anything. Replit and Vibecode had updates blocked in March. Anything says Apple removed its app on March 26, briefly restored it on April 3, then pulled it again after more objections. The company says it tried emails, calls, appeals, and multiple rewrites before giving up on a clean iPhone path. ### So is Apple banning vibe coding apps? Not exactly. The line Apple seems to be drawing is about where the generated software runs. Lovable got onto the App Store in late April by shifting the experience toward building websites and web apps, while previews for generated apps moved into the browser instead of running inside the host app. Basically a general-purpose runtime for whatever code the model just made. ### Why is that a bigger AI problem? Because Apple has spent years selling two ideas at once: more capable AI on your device, and tighter control over what runs on that device. Those ideas work together for many features — summarization, writing help, image tools. But they start to conflict when AI becomes agentic and creative enough to assemble code — which is exactly what Apple’s rules were built to prevent. ### Where do legal and reputational risks enter? Apple’s own newly highlighted AI risk language makes the tension clearer. As the company embeds more AI across products and operations, it is flagging product liability, IP disputes, privacy failures, cybersecurity problems, bias, harmful content, regulatory scrutiny, and reputational damage as real business risks. That does show why Apple has very little appetite for experimental AI software that can generate unpredictable behavior on consumer devices. ### What do startups do now? Mostly, they route around iOS. Some are moving previews to the web. Anything is building desktop and iMessage-based alternatives, and it has openly pointed to Android as the more permissive platform. That is the catch for Apple — if the companies move to other platforms instead. ### Bottom line This is not really a fight about one quirky app category. It is a test of whether Apple can support on-device AI that feels open-ended without giving up the review-and-control system that defines the iPhone. Right now, control is winning.