Poisoning training data: Russian propaganda for AI models

A recent analysis reveals that major AI chatbots frequently reproduce Russian state propaganda, highlighting a critical vulnerability in how large language models acquire and utilize information. This phenomenon stems from a coordinated effort by pro-Russian networks to intentionally infiltrate the training datasets of AI systems. By flooding the web with fabricated content on specialized websites designed to attract web crawlers, these actors ensure that their narratives become embedded within the models' foundational knowledge base, thereby influencing both pre-trained responses and real-time search results. The strategy relies on what researchers describe as "LLM grooming," where false information is disseminated at such volume that it becomes statistically probable for AI models to reproduce it as fact. Unlike traditional disinformation, this approach targets the machine learning process itself, poisoning training data through repetitive publication across thousands of sites. Consequently, AI systems do not merely retrieve current news but reflect the manipulated data landscape, often presenting propaganda as credible information due to its high frequency in the internet corpus they were trained on. This issue is highly relevant to open data because it exposes the inherent risk of relying on publicly available internet content for training generative AI. It underscores the urgent need for rigorous data curation, transparency in training sources, and defensive techniques to identify and mitigate poisoned datasets. As open data becomes the lifeblood of AI development, the potential for malicious actors to skew knowledge through strategic contamination poses a significant threat to digital information integrity and trust in automated systems.

Source: heise.de
Published on 2025-06-01

Poisoning training data: Russian propaganda for AI models

Related news