Implicit Fuzzy Specifications, Inferior to Explicit Balancing.

Abstract

Lukas J. Meier et al. offer the promise of a pathway for resolving clinical bioethical problems using an artificial intelligence (AI) interface (Meier et al. 2022). The ultimate goal, we assume, is to make sounder decisions that would reduce the need for the expertise currently required. The authors emphasize, such AI is not currently ready for prime time and that “...machine intelligence will inevitably be inferior to human ethicists... .” Although this proof of concept paper makes clear the plethora of choices to be made, resulting machine learning may be highly sensitive to such choices including the establishments of initial weights. Our aim is not to evaluate the various choices the authors make including selecting principlism as developed by Beauchamp and Childress (B & C) (Beauchamp 2013). Rather, we argue that the proof of concept using fuzzy specification does not accomplish the intended task and that the end goal of such technologies should be based on balancing that would provide tools to augment and error check human ethicists. In evaluating whether the AI program, METHAD, accomplishes the intended initial goal, four issues arise: (1) B & C do not advocate balancing, but instead rely on specification of their principles, as we more fully explain in (DeMarco and Ford 2006). We argue that specification is inferior to balancing. (2) The application of METHAD is an implicit use of specifications rather than B & C’s explicit specification. (3) METHAD is opaque and fuzzy in a way that fails Meier et al.’s own transparency requirement. (4) Finally, balancing is superior to the implicit specification of METHAD since it facilitates the development and teaching of sound clinical ethics decision making as tool for augmentation and error checking. In their endorsement of specification B & C describe its function: “Specification is a process of reducing the indeterminateness of abstract norms and providing them with action-guiding content” (Beauchamp and Childress 2001, 16). They endorse Henry S. Richardson argument that specification is superior to balancing. In his view, balancing is too intuitive or subjective (Richardson 1990, 298). B & C offer various examples of specification: For example: “‘Follow a patient’s advance directive whenever it is clear and relevant’” (Beauchamp and Childress 2001, 39). And: “It is morally prohibited to disrespect a parental refusal of treatment, unless the refusal constitutes child abuse, child neglect, or violates a right of the child” (270) [footnote, please see the balancing article for a more detail examination of Richardson’s position and B & C’s specification.] Notice that by definition, specification does not balance principles; instead it reduces their indeterminateness. Once reduced, the principles do no additional work. As Richardson points out, specifications are based on sustained argumentation, which may properly rely on balancing. (Richardson, 1990, 305–308) However, once the moral universe is populated by a plethora of specifications, the complex deliberative process is neither needed nor transparent. Specification’s inferiority to explicit balancing is based on its dependency on implicit balancing and the lack of transparency. Directly applying the balancing of basic values is, by contrast, transparent and so facilitates debate and teaching ethical decision making. The claim that fuzzy cognitive maps (FCMs) are a form of balancing seems to be based on the use of the weights between identified nodes, for example, age and autonomy (Meier et al. 2022). For METHAD training involves the assignment of weights to produce outcomes that match expert opinion. Thus, the assigned weights do not indicate the strength of a particular node in terms of producing the end result; rather each weight functions in a holistic system to produce the final decision (Felix et al. 2019, 1714) This is not balancing. If the weight of the relationship