The BBC investigation has ignited a fire that now reaches Meta's top ranks. India's Ministry of Electronics and Information Technology has announced it will summon the company's executives to demand a formal explanation: according to the reconstruction, Instagram ran paid advertisements containing child sexual abuse material (CSAM) to users in the country. Minister Ashwini Vaishnaw directed officials to obtain an official response. This is not the first time a platform finds itself under scrutiny for content moderation, but the fact that these were paid ads — sold and distributed by the platform's own advertising system — makes the affair more sensitive.

Behind the case lies something of keen interest to those dealing with artificial intelligence deployment in companies and governments: who controls the models deciding what gets through and what gets blocked? Automatic moderation systems, often based on a mix of computer vision and Large Language Models (LLMs), run on centralized cloud infrastructure, far from local judges and regulators. When a government like India's demands accountability for a failure, the audit process collides with technical black boxes and the logistical and legal distance of data centers.

The illusion of remote control

For a global platform, offloading moderation to a single cloud provider is efficient in terms of operating costs, but it creates a dependency that is difficult to untangle when an investigation starts. Logs of model decisions, neural network weights, inference pipelines: everything resides outside national jurisdiction. If an authority wanted to inspect in real time the sensitivity thresholds of the classifier that should have caught CSAM, it would find it cannot do so without the active cooperation of the company — which, legitimately, also guards its trade secrets.

It is therefore unsurprising that India's demand is not merely a formal summons, but a political signal: digital sovereignty also means being able to touch the models shaping citizens' experience. And here a theme dear to AI-RADAR comes into play: self-hosted. Moving part of the moderation inference onto local infrastructure, perhaps with quantized models to run on less exotic hardware, is no longer science fiction. The same vendors offering proprietary LLMs are making reduced versions available, optimized for on-premise deployment, precisely to meet the needs of government clients and enterprises with data residency constraints.

Of course, trade-offs abound. An on-premise moderation system requires investment in compute capacity — GPUs with adequate VRAM, fast storage, continuous model update pipelines — and a team capable of managing the AI lifecycle without depending on the vendor. Total Cost of Ownership (TCO) must be weighed against the cost of fines and reputational damage when centralized moderation fails. Not to mention that, in markets like India, network latency and the ability to operate in air-gapped environments can become decisive arguments.

The Meta-India affair might look like just a political-bureaucratic episode. In reality, it accelerates a discussion many IT managers are already having: how far can we delegate automatic content surveillance to a foreign cloud, and when is it time to bring part of that capability inside national borders, with models that can be updated, inspected, and placed under the control of those subject to local law? For now, the answer is suspended between engineering and diplomacy.