Justifying Our Credences in the Trustworthiness of AI Systems: A Reliabilistic Approach

Area of Science:

• Philosophy of Artificial Intelligence
• Epistemology
• Human-Computer Interaction

Background:

• Justifying epistemic attitudes towards artificial intelligence (AI) systems is crucial for appropriate reliance in human-AI interactions.
• AI trustworthiness is a complex, time-dependent property with both actual and perceived dimensions.
• Existing frameworks lack a robust method for evaluating the justification of beliefs about AI trustworthiness.

Purpose of the Study:

• To propose a novel framework for justifying epistemic attitudes towards AI trustworthiness.
• To conceptualize AI trustworthiness as a time-relative, composite property.
• To establish a reliabilistic account for justified credences in AI trustworthiness.

Main Methods:

• Conceptualizing trustworthiness as a composite property with actual and perceived facets.
• Utilizing credences (beliefs held with confidence) as the appropriate epistemic attitude.
• Deriving a reliabilistic account from Tang's probabilistic theory of justified credence.
• Defining justified credence based on a calibrated assessment process.

Main Results:

• Credences are identified as the suitable attitude for representing AI trustworthiness over time.
• A reliabilistic account is introduced, linking justified credence to a calibrated assessment process.
• The proposed account calibrates actual and perceived AI trustworthiness.
• The framework provides a method for measuring the reliability of user perception of AI trustworthiness.

Conclusions:

• The developed framework offers a philosophical basis for understanding and justifying beliefs about AI trustworthiness.
• It provides actionable insights for enhancing human-AI interaction by ensuring appropriate reliance on AI systems.
• The research contributes to the ethics of AI by offering a method to assess the reliability of AI trustworthiness perception.