How easily can Russian propaganda fool AI models? A new benchmark finds out

Problem — The paper addresses the lack of standardized metrics for assessing AI language models’ susceptibility to misinformation, specifically Russian propaganda. This gap is particularly relevant given the increasing use of AI in content generation and information dissemination. The work is presented as a preprint, indicating it has not yet undergone peer review.

Method — The authors propose a benchmark that consists of a curated dataset of Russian propaganda texts, designed to evaluate the performance of various AI language models in identifying and resisting such content. The benchmark includes a set of tasks that require models to discern between propaganda and non-propaganda texts. The evaluation metrics focus on accuracy, precision, recall, and F1-score, providing a comprehensive assessment of model performance. While specific architectures and training compute details are not disclosed, the benchmark is tested on several well-known models, including GPT-3 and BERT variants, to establish baseline performance.

Results — The benchmark reveals that state-of-the-art models exhibit significant vulnerabilities when exposed to Russian propaganda. For instance, GPT-3 achieved an accuracy of only 65% in identifying propaganda, while BERT variants performed slightly better at around 70%. These results are contrasted against a baseline accuracy of 85% on non-propaganda texts, indicating a substantial drop in performance when faced with propaganda. The effect size is notable, suggesting that the models struggle to generalize their understanding of nuanced misinformation tactics employed in the propaganda.

Limitations — The authors acknowledge several limitations, including the potential bias in the dataset, which may not encompass the full spectrum of propaganda techniques. Additionally, the benchmark’s focus on Russian propaganda may limit its applicability to other forms of misinformation. The lack of peer review raises questions about the robustness of the findings, and the authors do not address the potential for adversarial attacks that could further exploit model weaknesses.

Why it matters — This work is crucial for understanding the vulnerabilities of AI language models in the context of misinformation, particularly as these models are increasingly deployed in sensitive applications such as news generation and social media moderation. The benchmark provides a foundation for future research aimed at enhancing model robustness against propaganda, which is essential for maintaining the integrity of information systems. As highlighted in the paper, addressing these vulnerabilities is vital for the ethical deployment of AI technologies in society, especially in politically charged environments. This research is available on The Decoder.