What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models' capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and government grants. Despite this, the notion of an ML model possessing a capability has not been interrogated: what are we saying when we say that a model is able to do something? (...) And what sorts of evidence bear upon this question? -/- In this paper, we aim to answer these questions, using the capabilities of large language models (LLMs) as a running example. Drawing on the large philosophical literature on abilities, we develop an account of ML models' capabilities which can be usefully applied to the nascent science of model evaluation. Our core proposal is a conditional analysis of model abilities (CAMA): crudely, a machine learning model has a capability to X just when it would reliably succeed at doing X if it 'tried'. The main contribution of the paper is making this proposal precise in the context of ML, resulting in an operationalisation of CAMA applicable to LLMs. We then put CAMA to work, showing that it can help make sense of various features of ML model evaluation practice, as well as suggest procedures for performing fair inter-model comparisons. (shrink)

In this paper, I examine whether agents who commit testimonial injustice are morally responsible for their wrongdoing, given that they are ignorant of their wrongdoing. Fricker (2007) argues that agents whose social setting lacks the concepts or reasons necessary for them to correct for testimonial injustice are excused. I argue that agents whose social settings have these concepts or reasons available are also typically excused, because they lack the capacity to recognise those concepts or reasons. Attempts to trace this lack (...) of capacity back to an earlier culpable wrongdoing will often fail, due to there being no point at which these perpetrators knowingly chose to develop their prejudices. Attempts to ground culpability under Talbert’s (2008) Attributionist account of moral responsibility will also fail. This is because perpetrators’ lack of awareness of what they are doing makes it the case that they are not expressing objectionable evaluative judgments in the way required for blameworthiness. Finally, I argue that our temptation to blame agents who commit testimonial injustice is not completely unfounded. Appealing to Watson’s (1996) attributability/accountability distinction allows us to make sense of how some responses to the jurors are appropriate, despite their being excused. (shrink)

In the book, I provide an account of what it is for an agent to have an ability. According to the Success View, abilities are all about success across possible situations. In developing and applying the view, the book elucidates the relation between abilities on the one hand and possibility, counterfactuals, and dispositions on the other; it sheds light on the distinction between general and specific abilities; it offers an understanding of degrees of abilities; it explains which role intentions and (...) performances play for the having of an ability, how abilities relate to success and failure, and gives an outlook at how the view might be applied to the free will debate and time travel. (shrink)

A proper understanding of agentive modals promises to clarify issues to do with free will, know how, and other philosophically interesting topics. In this paper I identify one constraint on, and one structural feature of, trying-based versions of the conditional analysis of the agentive modals. I suggest that the constraint and structural feature together provide a novel account of why the famous Lehrer-Chisholm objection to conditional analyses of ability modals is so powerful. I argue that Mandelkern et al.’s ‘Agentive Modals’ (...) (Philosophical Review,126/3, 301–343, 2017) conditional analysis of the agentive modals fails to avoid this problem. I also identify two further problems for their account. I close by summarising a number of criteria which any successful semantic analysis of the agentive modals should satisfy. (shrink)