When Algorithms ‘Think’: A Computational Analysis of Anthropomorphism in AI News Coverage

Abstract

Over the past few years, researchers have observed an increase in anthropomorphic framing of artificial intelligence (AI) in news media. If anthropomorphic framings of AI remain uncritical about the true capabilities of the technology and become normalised in society, then this can have consequences for how people perceive, assess, and engage with the technology. Further research into the ways that anthropomorphic language is being used in relation to AI entities is needed to better to understand it. While previous research has focused mainly on identifying the mere presence of anthropomorphism, the present study looks deeper into the different forms of the phenomenon and proposes a conceptual framework that distinguishes between five types of anthropomorphism: established, behavioural, cognitive, emotional and social. Furthermore, two computational methods are compared to provide a transparent and scalable methodological approach for identifying and classifying anthropomorphism in large corpora of news articles. This research found that the two methods each identified different aspects of anthropomorphic language. The Dependency Parsing method (N = 8.062 articles, 50.900 sentences) identified active, agentic forms of anthropomorphism, while the AnthroScore method (N = 919 articles, 1.000 sentences) was able to classify more subtle and metaphoric cases. Although more research needs to be done define the phenomenon and refine the methods presented in this research, it can be concluded that different types of anthropomorphism are present, and that their use varies across news outlets and AI systems.