San Francisco City Attorney David Chiu’s demand that Apple and Google remove 13 “nudification” apps from their stores – tools that with a few taps turn ordinary photos into explicit images – is more than a simple legal action. It exposes the structural tension between centralized control of software distribution and the decentralized nature of generative AI tools.

These apps leverage deep learning models – often generative adversarial networks or diffusion models – to alter photographs, removing clothing or changing poses and faces. Although many of them rely on cloud processing, there is no technical barrier to running them locally: a computer with a GPU featuring enough VRAM can execute similar models without ever sending data to third parties. And it is precisely this aspect that makes the attorney’s move one that goes far beyond a single app.

Chiu’s complaint against the platforms is that they are violating California laws that prohibit supporting services used to create deepfake pornography. By targeting the app stores, the legal action aims at the most visible and controllable distribution point. But this is where a crack opens: when official channels are closed, demand does not disappear; it moves elsewhere. It’s a pattern seen before with pirated content, exploit marketplaces, and, in the AI realm, with open-source language models released without safeguards. Today, a determined user can download a free model from sharing platforms like Hugging Face, install an interface such as Stable Diffusion WebUI, and generate images locally, far from any control by stores or central authorities.

For businesses and professionals evaluating generative AI deployment, the episode adds a tile to the complex puzzle of data sovereignty and compliance. Relying on cloud services means accepting an intermediary’s policies, which in turn is exposed to regulatory pressures that vary from jurisdiction to jurisdiction. An on-premise infrastructure, conversely, returns full control over the use of tools, but shifts onto the owner the responsibility of ensuring they are not employed for illicit purposes – a fine line that, in the case of deepfakes, has already begun to attract the attention of regulators worldwide.

The affair also signals a shift in institutional attitudes toward generative AI. It is no longer just a matter of content moderation on social platforms; now the intervention reaches the software distribution level, one step further upstream. This creates an indirect incentive for the development of tools that run entirely locally, perhaps optimized for consumer hardware: think of quantization efforts that allow image generation models to run on laptops with discrete GPUs, without needing cloud servers.

Ultimately, San Francisco’s demand is not just the chronicle of a measure against morally questionable apps. It is a symptom of the growing awareness that the real battleground for AI control will be the execution infrastructure. And in this scenario, every decision – cloud or local, official stores or alternative channels – will have consequences that ripple well beyond the single application.