Article: AI researchers have uncovered a coordinated effort by a network of about 150 low-traffic news sites to flood large language model (LLM) training data with pro-Kremlin narratives. The sites, collectively known as the Pravda network, push out roughly three million articles a year, and their content now appears in the Common Crawl dataset that powers many commercial AI systems.

How the disinformation pipeline works

The Pravda network does not aim to attract human readers. Instead, each site publishes repetitive articles that repeat a narrow set of talking points. By stuffing the text with keywords that search engines and web crawlers prioritize, the sites increase the chance that an AI model will encounter them during data collection.

Two poisoning mechanisms are at play:

  • Retrieval-time poisoning – When an AI assistant pulls live information from the web, it can surface a Pravda article alongside legitimate sources. Blocking known malicious domains can curb this exposure.
  • Training-time poisoning – If the articles slip into the bulk corpora used to pre-train a model, the false claims become part of the model’s internal knowledge. Removing such content after the fact is far harder.

Researchers have already traced Pravda articles inside Common Crawl, a public archive used to train many AI models. That means the poisoning is not hypothetical; it has already altered the knowledge base of models that many users rely on today.

Why it matters

Do not trust AI just because it says "many sources agree." If you build AI tools, use blocklists for known disinformation sites. Do not use "number of mentions" as a way to measure truth. Quantity does not equal quality. Verify everything an AI tells you, especially if it seems biased.

The internet is the training set for our future technology. Anyone with a website can try to bias the machines we rely on.

What developers can do now

  • Maintain blocklists – Add known disinformation domains to crawling filters; a blocklist stops both retrieval-time and training-time exposure.
  • Rethink frequency heuristics – Stop equating the number of mentions with veracity. A claim repeated across dozens of low-quality sites can outvote a single reputable source in a naïve frequency-based model.
  • Apply provenance checks – Trace information back to its original source. If the chain ends at a site with negligible traffic and a propaganda history, flag the output for review.
  • Implement post-training sanitization – Run curated audits on model outputs that touch politically sensitive topics. Human reviewers can catch systematic bias that automated filters miss.

Counter-points and challenges

Balancing security and inclusivity remains an open problem.

Looking ahead

The Pravda network shows how state actors can weaponize the open web to shape the next generation of AI.

Takeaway: The integrity of AI hinges on the cleanliness of the data it learns from. A coordinated flood of low-traffic, propaganda-laden sites can already embed falsehoods into the models we trust. Developers must treat source reputation as a first-class concern, not an afterthought, to keep the machines we build from becoming unwitting mouthpieces for disinformation.