Transparency as design publicity: explaining and justifying inscrutable algorithms.

Area of Science:

• Computer Science
• Philosophy of Technology
• Artificial Intelligence Ethics

Background:

• Current methods for explaining machine learning algorithms, particularly 'black box' models, often rely on post-hoc techniques like counterfactual explanations.
• These methods oversimplify complex algorithms and can mislead the public about a model's true nature and normative implications.

Purpose of the Study:

• To propose a novel framework for algorithmic transparency called 'design publicity.'
• To differentiate design publicity from existing post-hoc explanation methods.
• To establish how design publicity can provide a more robust justification for algorithmic design and decision-making.

Main Methods:

• Critique of current post-hoc explanation techniques for machine learning.
• Proposal and definition of 'design publicity' based on Daniel Dennett's design stance.
• Application of John Rawls's concept of procedural justice to algorithmic decision-making.

Main Results:

• Design publicity defines transparency by explaining algorithms as intentional products with specific goals and measurable outcomes.
• This approach offers a more direct link to the justification of an algorithm's design and individual decisions.
• Design publicity addresses the need for impersonal justification, distinct from explaining individual decisions.

Conclusions:

• Design publicity provides a superior form of transparency compared to post-hoc explanations.
• Justifiable algorithmic goals, including fairness and bias avoidance, when pursued consistently, can justify all resulting decisions, even those with incorrect predictions.
• Algorithmic transparency through design publicity aligns with principles of procedural justice.