Grooming the Machines: How the Pravda Network Targets the AI You Ask for the Truth

A Russian network published 3.6 million articles in a single year — for an audience of almost no one. Because the readers it wants aren't human. Inside the Pravda network's bid to "groom" the AI assistants you ask for the truth: how it works, why the famous "33%" is partly a mirage, and the slow-poisoning risk that should worry you far more. Members only.

Key Judgement

We assess with high confidence that the Pravda network (also tracked as "Portal Kombat") exists, is Kremlin-aligned, and is engineered for machines rather than people: a constellation of roughly 150 to 180 domains across dozens of countries that republished an estimated 3.6 million articles in 2024, despite drawing almost no human audience. We assess with moderate confidence that its method — termed "LLM grooming" by the American Sunlight Project — has a real but conditional effect: independent testing shows chatbots are most likely to surface Pravda content on narrow, obscure topics where little reliable information exists ("data voids"), and far less so on well-covered narratives. We judge the more serious long-term risk to be not today's retrieval citations, which are mitigable, but the slower contamination of AI training data, which is not yet confirmed and not easily reversed.

Why This Dossier Matters

Every disinformation operation until now has ultimately needed a human reader. This one largely doesn't. The Pravda network's bet is that AI assistants are becoming the layer through which millions of people learn what is true — and that poisoning the machines is more efficient than persuading the humans. Whether that bet pays off is genuinely contested, and a disciplined reading has to hold both the alarm and the skepticism at once. That tension is exactly why this warrants a Dossier: the threat is real, the mechanism is misunderstood, and the loud version of the story is partly wrong.

advertising