Document is current

Article History

Received: 4 November 2024

Accepted: 14 February 2025

First Online: 21 February 2025

Declarations

:

: LLMs are known to inherit biases from their training data [], which can manifest in their interpretations and judgements regarding the presence or absence of credibility signals in textual content. These biases may lead to inaccuracies or disparities in signal detection, potentially favouring certain types of content or perspectives over others. Moreover, the deployment of LLM-based systems in real-world applications must navigate concerns around fairness, transparency, and accountability. Researchers and developers are therefore urged to mitigate biases through rigorous testing, data preprocessing, and continuous monitoring.Also, although efforts aimed at mitigating misinformation are crucial in combating its harmful effects, it is important to acknowledge that these efforts can inadvertently empower malicious actors []. By gaining insights into which credibility signals are more easily detected by LLMs, and which correlate more strongly with veracity, malicious users could potentially exploit this knowledge to enhance their misinformation tactics and circumvent automatic detection systems. Therefore, we strongly urge researchers to apply our methodology with caution and in accordance with best practice ethics protocols.

: Not applicable.

: The authors provide their full consent for the publication of this manuscript in EPJ Data Science.

: The authors declare no competing interests.