Poisoning training data: Russian propaganda for AI models

A recent systematic analysis reveals that major AI chatbots, including ChatGPT and Copilot, are reproducing Russian propaganda due to a coordinated disinformation campaign known as "LLM grooming." A pro-Russian network deliberately feeds false narratives into the web content that AI models scrape for training and real-time searches. By optimizing this content for crawlers rather than human readers, the network successfully poisoned the training data, causing algorithms to prioritize and repeat Kremlin-affiliated misinformation. This highlights a critical vulnerability in how open data systems learn: models derive responses based on the statistical likelihood of data occurrences rather than inherent truth. When bad actors flood the internet with specific narratives, they manipulate the probability distributions that guide AI outputs. This demonstrates that open data ecosystems are not immune to manipulation; without robust verification mechanisms, the sheer volume of accessible but false information can distort the knowledge base of widely used AI tools. The relevance to open data is profound, as it underscores the urgent need for data provenance and quality control standards. It proves that the open availability of online information does not guarantee the reliability of synthesized insights. Developers and data stewards must implement stricter filtering and validation processes to prevent malicious actors from exploiting open datasets, ensuring that transparency in data does not inadvertently facilitate the spread of organized disinformation.

Source: heise.de
Published on 2025-03-28

Poisoning training data: Russian propaganda for AI models

Related news