Ethics and Information Technology最新文献

This paper argues that one problem that besets black-box AI is that it lacks algorithmic justifiability. We argue that the norm of shared decision making in medical care presupposes that treatment decisions ought to be justifiable to the patient. Medical decisions are justifiable to the patient only if they are compatible with the patient's values and preferences and the patient is able to see that this is so. Patient-directed justifiability is threatened by black-box AIs because the lack of rationale provided for the decision makes it difficult for patients to ascertain whether there is adequate fit between the decision and the patient's values. This paper argues that achieving algorithmic transparency does not help patients bridge the gap between their medical decisions and values. We introduce a hypothetical model we call Justifiable AI to illustrate this argument. Justifiable AI aims at modelling normative and evaluative considerations in an explicit way so as to provide a stepping stone for patient and physician to jointly decide on a course of treatment. If our argument succeeds, we should prefer these justifiable models over alternatives if the former are available and aim to develop said models if not.

Voting Advice Applications (VAAs) are interactive tools used to assist in one's choice of a party or candidate to vote for in an upcoming election. They have the potential to increase citizens' trust and participation in democratic structures. However, there is no established ground truth for one's electoral choice, and VAA recommendations depend strongly on architectural and design choices. We assessed several representative European VAAs according to the Ethics Guidelines for Trustworthy AI provided by the European Commission using publicly available information. We found scores to be comparable across VAAs and low in most requirements, with differences reflecting the kind of developing institution. Across VAAs, we identify the need for improvement in (i) transparency regarding the subjectivity of recommendations, (ii) diversity of stakeholder participation, (iii) user-centric documentation of algorithm, and (iv) disclosure of the underlying values and assumptions.

Supplementary information: The online version contains supplementary material available at 10.1007/s10676-024-09790-6.

Newly powerful large language models have burst onto the scene, with applications across a wide range of functions. We can now expect to encounter their outputs at rapidly increasing volumes and frequencies. Some commentators claim that large language models are bullshitting, generating convincing output without regard for the truth. If correct, that would make large language models distinctively dangerous discourse participants. Bullshitters not only undermine the norm of truthfulness (by saying false things) but the normative status of truth itself (by treating it as entirely irrelevant). So, do large language models really bullshit? I argue that they can, in the sense of issuing propositional content in response to fact-seeking prompts, without having first assessed that content for truth or falsity. However, I further argue that they need not bullshit, given appropriate guardrails. So, just as with human speakers, the propensity for a large language model to bullshit depends on its own particular make-up.