The story comes from Reddit and is picked up by The Verge: a tool capable of generating pornographic deepfakes of women and children was allegedly hosted on Hugging Face, the go-to platform for open models. The immediate outcry is predictable: “think of the children” once again becomes the rallying cry to demand restrictions on open-source artificial intelligence.

The case is thorny and the severity of the abuse is beyond question. But if we stop at indignation, we risk missing the real strategic knot. This story is the umpteenth episode of a recurring script: every accessible AI tool is exploited for illicit purposes, and the hunt for open models begins instantly. The target is not just the technology, but the distribution paradigm that underpins it.

Image generation models like Stable Diffusion, often used as a base for fine-tuning toward deepfake applications, are open weight: anyone can download them, modify them, and run them on their own hardware. This means a company, a public administration, or a hospital can perform inference entirely on-premise, without sending sensitive data to external servers. That is the cornerstone of digital sovereignty: full control over data, latency, costs, and compliance (including GDPR).

The gut reaction to clamp down on open source in the name of safety threatens to throw the baby out with the bathwater. If regulators were to force platforms like Hugging Face to restrict model access, or if public opinion pushed toward centralizing distribution, the losers would not only be the malicious actors. The biggest casualties would be the legitimate entities that need self-hosted AI for privacy, reliability, or simple TCO reasons. And the winners would be the large cloud vendors, which already offer AI-as-a-service on a subscription basis with little transparency about data handling.

From a technical standpoint, nudification does not require esoteric infrastructure: a consumer GPU with enough VRAM to run the quantized model and a few preprocessing scripts are all it takes. This workload could run on an on-premise company server, with all data staying local. The fundamental difference between an abuser and a researcher studying generative models lies in intent, not in technology. Blocking the latter to stop the former is a mistake we already witnessed with encryption and peer-to-peer.

The question is not whether it is right to condemn misuse (it is), but whether the proposed solution – restricting open source – is proportionate or even counterproductive. Alternatives exist: moderation systems integrated into data loading pipelines, forensic watermarking applied at inference time, and synthetic content detection models that can also run locally. These are tools that respect data sovereignty and allow those deploying on-premise to maintain control without yielding to centralized censorship imposed by third parties.

This incident is a structural signal: the battle for the future of AI will not be fought only over benchmarks or GPUs, but over perceived safety. Proponents of centralized cloud will use every news story to paint open models as uncontrollable, while those who need technological autonomy will have to prove that use can be governed without breaking the open backbone on which innovation rests. For anyone evaluating an on-premise deployment, the lesson is clear: investing today in moderation tooling and internal auditing is not a luxury, but the only way to avoid having rules dictated by those who would like to close everything down. And perhaps, to rob the usual alarmists of their “think of the children” argument.