Mindful Explanations: Prevalence and Impact of Mind Attribution in XAI Research

Mind-attributing language in AI explanations is widespread: The study analyzed 3,533 XAI research articles and found three dominant types of mind-attributing language—metaphorical terms like "learn," awareness terms like "consider," and agency terms like "make decisions"—suggesting this framing is common throughout the field.

How AI is described affects who gets blamed: When explanations used mind-attributing language, participants were significantly more likely to hold the AI system responsible for harm, even when they were reminded that human experts created and designed it. This represents a "responsibility-concealing effect" that shifts accountability away from developers.

Non-technical language choices have real consequences: The study's experiment with 199 participants demonstrated that simply changing how explanations are worded—using phrases that suggest the AI "thinks" versus more neutral descriptions—measurably changed how people perceived the AI's awareness and who should be held accountable for its actions.

Current XAI practices may undermine accountability: By frequently using mind-attributing language, XAI research may inadvertently encourage the public to view AI systems as independent agents rather than tools created by humans, potentially obscuring the responsibility of AI developers and organizations for system failures or harms.

Careful language in explanations is essential for responsible AI: The research highlights that responsible AI communication requires deliberate attention to how systems are described, as imprecise language can fundamentally alter public understanding of accountability and agency.