Michael Zahorec

This dissertation stakes out a position against prevailing views on natural and scientific kinds by responding to three questions: (1) Are kinds real? (2) What are kinds like? (3) Which kinds are there? Before responding to these, I establish the significance of philosophy of kinds. Chapter 1 shows how philosophy about kinds, traditionally taken up under the heading ‘natural kinds’ but now often ‘scientific kinds’, naturally arises in core theoretical philosophical inquiries (e.g., in ontology and in dealing with the question of realism and the nature of concepts) and, conversely, how philosophical positions on kinds are simultaneously positions on other foundational philosophical issues. Chapter 2 defends a non-metaphysical realism, positioned between more popular alternatives. Metaphysical kind realism goes farther than we need to go to secure belief in our successful practices and their objects, while anti-realist theories like domain relativism, kind eliminativism, and merely epistemic accounts misconstrue and undermine our epistemic practices. My position preserves the domain independence and thus the objectivity of kinds, while rejecting that they are mind independent. Chapter 3 argues that no general theory of kinds will succeed. I show that two enduringly popular theories – essentialism and property cluster theory – fail, and argue that kinds are too diverse to admit of a general, informative treatment. Chapter 4 defends the reality of many kinds typically taken as apparent or merely conventional. Taking examples from biology, computer science, psychology, and everyday life, I show how standard approaches – what I call “restrictive” approaches – exclude indispensable kinds which obviously exist. I also show how popular permissive accounts fail: by still trying to offer a principle or account that informatively says what makes kinds real. I suggest this pursuit is destined to fail, and advocate for a novel permissive account which abandons the search for a general formula. Permissiveness fits naturally with my rejections of metaphysical realism and unified theories of kinds, which are the main motivations for restrictiveness.

Nearly everything about understanding AI is contested, including what understanding AI is and why it matters. With two epistemic distinctions—interpretable AI versus explanations of AI and internal-process explanations versus black-box (behavioral) explanations—I clarify this terrain. Drawing an analogy to motivating reasons for human action, I show that internal AI processes are normatively significant. Accordingly, black-box explanations face in-principle limitations, rendering them insufficient for responsible AI use and deployment. I show that black-box explanation, including LLM self-explanation, is especially limited for generative AI. Efforts to achieve the knowledge needed to overcome these limitations, such as mechanistic interpretability, are promising but currently face their own challenges. Recent research warrants cautious optimism about overcoming these challenges but, for now, certain forms of trust in and reliance on complex AI systems are ruled out.

As generative artificial intelligence (AI) systems are embedded in research and education, as well as in our professional, creative, and everyday lives, we need frameworks for responsible use and deployment. This essay argues we should treat generative AI systems as idea generators. The term "idea" suggests modesty, as when we say something is "just an idea". Two independent arguments support this practical conclusion. First, we face severe epistemic limitations: interpretability remains infeasible, black-box explanations (including self-explanations) are liable to misrepresent the nature of AI processes generating outputs, and internal-process explanation strategies face serious obstacles. We do not understand generative AI systems well enough to treat their outputs as we treat those from human experts, reliable information sources, or reliable information processors. The idea-generator paradigm responds appropriately. As with inspiration from a poem or landscape, receiving an idea from AI warrants neither belief nor endorsement. Alternative paradigms, such as AI-as-expert or AI-as-neutral-information-processor, allow and often encourage unwarranted trust. Second, even if these epistemic issues were resolved, excessive delegation to AI remains concerning. Delegating tasks to AI undermines authorship and can erode valuable abilities. Treating AI as an idea generator – rather than as, say, an expert, surrogate decision-maker, or co-author – preserves our ownership over our intellectual, creative, and other actions and projects, and preserves relevant valuable abilities. By preserving both epistemic responsibility and authorship, the idea-generator paradigm allows us to benefit from AI while maintaining individual and collective claims to responsibility, competence, and expertise.

The fact that Gilbert Ryle and J.L. Austin seem to disagree about the ordinary use of words such as ‘voluntary’, ‘involuntary’, ‘voluntarily’, and ‘involuntarily’ has been taken to cast doubt on the methods of ordinary language philosophy. As Benson Mates puts the worry, ‘if agreement about usage cannot be reached within so restricted a sample as the class of Oxford Professors of Philosophy, what are the prospects when the sample is enlarged?’ (Mates 1958, p. 165). In this chapter, we evaluate Mates’s criticism alongside Ryle’s and Austin’s specific claims about the ordinary use of these words, assessing these claims against actual examples of ordinary use drawn from the British National Corpus (BNC). Our evaluation consists in applying a combination of methods: first aggregating judgments about a large set of samples drawn from the corpus, and then using a clustering algorithm to uncover connections between different types of use. In applying these methods, we show where and to what extent Ryle’s and Austin’s accounts of the use of the target terms are accurate as well as where they miss important aspects of ordinary use, and we demonstrate the usefulness of this new combination of methods. At the heart of our approach is a commitment to the idea that systematically looking at actual uses of expressions is an essential component of any approach to ordinary language philosophy.

The fact that Gilbert Ryle and J.L. Austin seem to disagree about the ordinary use of words such as ‘voluntary’, ‘involuntary’, ‘voluntarily’, and ‘involuntarily’ has been taken to cast doubt on the methods of ordinary language philosophy. As Benson Mates puts the worry, ‘if agreement about usage cannot be reached within so restricted a sample as the class of Oxford Professors of Philosophy, what are the prospects when the sample is enlarged?’ (Mates, Inquiry 1:161–171, 1958, p. 165). In this chapter, we evaluate Mates’s criticism alongside Ryle’s and Austin’s specific claims about the ordinary use of these words, assessing these claims against actual examples of ordinary use drawn from the British National Corpus (BNC). Our evaluation consists in applying a combination of methods: first aggregating judgments about a large set of samples drawn from the corpus, and then using a clustering algorithm to uncover connections between different types of use. In applying these methods, we show where and to what extent Ryle’s and Austin’s accounts of the use of the target terms are accurate as well as where they miss important aspects of ordinary use, and we demonstrate the usefulness of this new combination of methods. At the heart of our approach is a commitment to the idea that systematically looking at actual uses of expressions is an essential component of any approach to ordinary language philosophy.