Rome – India’s Ministry of Electronics and Information Technology has given Meta a peremptory order: within seven days, the company must explain how Instagram advertisements promoting child sexual abuse material were approved and published. The case, reported by The Week, comes as India tightens its regulatory framework on digital content, already a battleground with Big Tech.

The news is another wake-up call for global platforms that increasingly rely on automated systems – including Large Language Models and computer vision models – to moderate content. The scale makes human moderation alone impossible. But the Indian incident shows the fragility of relying solely on centralized AI and automatic approval pipelines that, this time, clearly failed to block manifestly illegal ads.

This case has particular resonance for anyone evaluating the deployment of artificial intelligence solutions in a fragmented regulatory landscape. It’s not just Meta’s problem: any organization that manages user-generated content – from an e-commerce platform to a specialist forum – must contend with local rules that vary by country. A cloud-only approach, where moderation is outsourced to third-party services and data flows through distributed infrastructure, makes it harder to demonstrate real-time compliance. And when something slips through, the reputational and legal damage can be huge.

That’s why some observers are looking with interest at on-premise deployment models, where the entire moderation stack – from automatic detection models to control dashboards – resides on local servers under the organization’s direct control. It’s not a simple fix: running LLMs and computer vision models in-house requires expertise, hardware (often GPUs with tens of gigabytes of VRAM), and a non-trivial total cost of ownership (TCO). But in return, it offers the ability to customize moderation policies, fine-tune filter sensitivity to local laws, and, not least, keep sensitive data out of the circuits of large providers, reducing the risk of exposure to foreign jurisdiction access requests.

Meta’s situation is obviously more complex because it operates a planetary ecosystem with billions of users. Yet even for smaller enterprises, the trade-off between cloud convenience and on-prem sovereignty becomes more pressing when the stakes involve child protection. The seven-day deadline imposed on Meta is not just an ultimatum: it’s a signal that regulators’ patience is running out and that automatic moderation mechanisms will need to be far more transparent and accountable – objectives that, by their very nature, are easier to pursue when the infrastructure is under your own roof.