From understanding to justifying: Computational reliabilism for AI-based forensic evidence evaluation.

Area of Science:

• Philosophy of Science
• Forensic Science
• Artificial Intelligence

Background:

• AI algorithms are increasingly used in forensic biometrics but often function as 'black boxes'.
• The inclusion of these AI systems in criminal law raises significant ethical, legal, and philosophical questions.
• A multidisciplinary debate is needed to address the suitability of AI in forensic evidence evaluation.

Purpose of the Study:

• To present a philosophy of science perspective on the use of AI in forensic evidence evaluation.
• To introduce computational reliabilism (CR) as a framework for assessing AI reliability in forensics.
• To argue for the justification of AI output in voice and facial comparisons within a legal context.

Main Methods:

• Applying the framework of computational reliabilism (CR) to AI in forensic science.
• Classifying grounds for believing AI reliability into three types of indicators: technical, scientific, and societal.
• Analyzing the suitability of AI for forensic voice and facial comparison based on these indicators.

Main Results:

• AI outputs in forensic voice and facial comparison can be considered reliable under CR.
• Technical indicators involve algorithm validation, forensic application validation, and case-based validation.
• Scientific indicators include inherent identifying information in faces/voices and established forensic metrics.
• Societal indicators encompass scientific consensus and practitioner expertise.

Conclusions:

• Computational reliabilism provides a robust framework for evaluating AI in forensic evidence beyond mere explainability.
• Expert witnesses may prioritize technical indicators, while triers-of-fact may rely on societal indicators for AI-supported evidence.
• The study supports the justified belief in AI-driven forensic comparisons of voices and faces.