Counterfactual Harm: A Counter-argument.

Area of Science:

• Artificial Intelligence Ethics
• Medical Decision Making
• Causal Inference

Background:

• AI systems increasingly guide decisions, necessitating adherence to ethical principles like non-maleficence.
• The counterfactual definition of harm, common in binary settings, is widely used.
• Ethical AI requires robust definitions of harm, especially in complex scenarios.

Purpose of the Study:

• To identify limitations of the counterfactual definition of harm in multi-option treatment settings.
• To propose an alternative definition of harm that ensures transitive rankings.
• To address challenges in justifying clinical decisions based on AI recommendations.

Main Methods:

• Analysis of a formal definition of harm based on counterfactual reasoning.
• Illustration using a hypothetical example with three tuberculosis treatment options (A, B, C).
• Comparison with an interventionist definition of harm utilizing expected utility.

Main Results:

• The counterfactual definition of harm can yield intransitive results (e.g., B < A, C < B, but C > A).
• This intransitivity complicates justification of AI-guided clinical decisions.
• The interventionist definition ensures transitive rankings, offering a more robust approach.

Conclusions:

• The counterfactual definition of harm is problematic for AI in multi-option medical decision-making.
• An interventionist definition based on expected utility provides transitive and justifiable treatment rankings.
• This research contributes to developing more reliable ethical AI for healthcare.