Rachel Katharine Sterken

The development of sophisticated artificial intelligence (AI) conversational agents based on large language models raises important questions about the relationship between human norms, values, and practices and AI design and performance. This article explores what it means for AI agents to be conversationally aligned to human communicative norms and practices for handling context and common ground and proposes a new framework for evaluating developers’ design choices. We begin by drawing on the philosophical and linguistic literature on conversational pragmatics to motivate a set of desiderata, which we call the CONTEXT-ALIGN framework, for conversational alignment with human communicative practices. We then suggest that current large language model (LLM) architectures, constraints, and affordances may impose fundamental limitations on achieving full conversational alignment.

A difficulty in combating fake news is that it is not entirely clear what fake news is. This chapter provides an overview of recent definitions, questioning the essential role they assign to the beliefs of fake news consumers. We offer an alternative definition, which raises questions about how fake news is to be distinguished from fictional news (e.g., news satire or parody). We propose that the categories of fake news and fictional news in fact overlap, although works of fictional news are members of artistic genres while works that are merely fake news are not.

Online speech is structured rather differently than offline speech. One important aspect of this, we argue, is that online speech environments are amplificatory. That is, these speech environments are designed to make the speech act of amplification easy, make amplification of others’ speech a predictable side-effect of one’s own, or both. In this essay, we first clarify what the speech act of amplification amounts to. Then we investigate the design choices of our present online speech environments which serve to promote amplification, and how these lay what we call the ‘amplification trap’ when it comes to responding to bad speech. Finally, we consider some ways that online environments could be re-designed in order to mitigate the amplification trap.

The development of sophisticated artificial intelligence (AI) conversational agents based on large language models raises important questions about the relationship between human norms, values, and practices and AI design and performance. This article explores what it means for AI agents to be conversationally aligned to human communicative norms and practices for handling context and common ground and proposes a new framework for evaluating developers’ design choices. We begin by drawing on the philosophical and linguistic literature on conversational pragmatics to motivate a set of desiderata, which we call the CONTEXT‐ALIGN framework, for conversational alignment with human communicative practices. We then suggest that current large language model (LLM) architectures, constraints, and affordances may impose fundamental limitations on achieving full conversational alignment.

According to an influential research program in cognitive science, philosophy, and linguistics, there is a deep, special connection between generics and pernicious aspects of social cognition such as stereotyping. Specifically, generics are thought to exacerbate our propensity to essentialize, lead us to overgeneralize based on scarce evidence, and lead to other epistemically dubious patterns of inference. Recently, however, several studies have put empirical and theoretical pressure on some of the main tenets of this research program. The goal of this paper is to bring these results together in a comprehensive narrative and systematically evaluate their impact on the hypothesis that generics have a uniquely problematic effect on our social and cognitive capacities.

This Handbook brings together philosophical work on how language shapes, and is shaped by, social and political factors. Its 24 chapters were written exclusively for this volume by an international team of leading researchers, and together they provide a broad expert introduction to the major issues currently under discussion in this area. The volume is divided into four parts: Part I: Methodological and Foundational Issues Part II: Non-ideal Semantics and Pragmatics Part III: Linguistic Harms Part IV: Applications The parts, and chapters in each part, are introduced in the volume's General Introduction. A list of Works Cited concludes each chapter, pointing readers to further areas of study. The Handbook is the first major, multi-authored reference work in this growing area and essential reading for anyone interested in the nature of language and its relationship to social and political reality.

In addition to ordinary conversations among relatively small numbers of individuals, human societies have public conversations. These are diffuse, ongoing discussions about various topics, which are largely sustained by journalistic activities. They are conversations about news – what is happening now – that members of various groups (such as the residents of a certain country, a certain town, or practitioners of a certain profession) need to know about in their capacity as members of those groups, and about how to react to the news. Our topic in this chapter is a type of resistance to evidence that can arise at the level of these public conversations, rather than at the level of individual agents. We call it “relevance-based resistance to evidence”. A public conversation exhibits this kind of evidence resistance when it becomes overly focused on topics that members of the group that the conversation concerns do not in fact need to know about qua members of the group – topics which are, in a word, irrelevant. We argue that the risks of such relevance-based knowledge resistance are significantly amplified by certain structural features of online discourse.

The aim of this chapter is to explore various ways of thinking about the concept of manipulation in order to capture both current and potentially future instances of machine manipulation, manipulation on the part of everything from the Facebook advertising algorithm to super-intelligent AGI. Three views are considered: a conservative one, which slightly tweaks extant influence-based theories of manipulation; a dismissive view according to which it doesn't matter much if machines are literally manipulative, provided we can classify them as so doing to make sense of our interactions with them; and an ameliorative analysis, according to which we should change our concept of manipulation better to make sense of machine manipulation. We tentatively favor the latter.

In this paper, we consider how the notion of metalinguistic negotiation interacts with various theories of generics. The notion of metalinguistic negotiation we discuss stems from previous work from two of us (Plunkett and Sundell). Metalinguistic negotiations are disputes in which speakers disagree about normative issues concerning language, such as issues about what a given word should mean in the relevant context, or which of a range of related concepts a word should express. In a metalinguistic negotiation, speakers argue about such issues implicitly, via competing “metalinguistic” usages of terms. Here, we argue that some disputes involving generics are best thought of as metalinguistic negotiations, and that these cases can be illuminating for our more general theorizing about generics. Specifically, we argue that the “contextualist” theory of generics that one of us (Sterken) has developed in other work is best equipped to account for these metalinguistic negotiations, relative to other leading views of generics. We thus argue for a “package deal” view of generics: a view that combines Plunkett and Sundell’s account of metalinguistic negotiation with Sterken’s contextualist view of generics.

In a recent paper (Haslanger 2016), Sally Haslanger argues for the importance of structural explanation. Roughly, a structural explana- tion of the behaviour of a given object appeals to features of the struc- tures—physical, social, or otherwise—the object is embedded in. It is opposed to individualistic explanations, where what is appealed to is just the object and its properties. For example, an individualistic explanation of why someone got the grade they did might appeal to features of the essay they wrote—its being well-written, answering the set question, etc. But if the class is graded on a curve, then a better explanation will appeal to features of the class—of the social structure in which the student is embedded. That she wrote a better paper than 90% of the class explains better than that she wrote a well-argued paper. In this paper, I get clear as to various candidate concepts of structure that we might appeal to in structural explanations, argue that Haslanger’s preferred account is lacking, and present an alterna- tive that is more conducive to social structural explanation.

What words we use, and what meanings they have, is important. We shouldn't use slurs; we should use 'rape' to include spousal rape (for centuries we didn’t); we should have a word which picks out the sexual harassment suffered by people in the workplace and elsewhere (for centuries we didn’t). Sometimes we need to change the word-meaning pairs in circulation, either by getting rid of the pair completely (slurs), changing the meaning (as we did with 'rape'), or adding brand new word-meaning pairs (as with 'sexual harassment'). A problem, though, is how to do this. One might worry that any attempt to change language in this way will lead to widespread miscommunication and confusion. I argue that this is indeed so, but that's a feature, not a bug of attempting to change word-meaning pairs. The miscommunications and confusion such changes cause can lead us, via a process I call transformative communicative disruption, to reflect on our language and its use, and this can be further, rather than hinder, our goal of improving language.

The term "fake news" ascended rapidly to prominence in 2016 and has become a fixture in academic and public discussions, as well as in political mud-slinging. In the flurry of discussion, the term has been applied so broadly as to threaten to render it meaningless. In an effort to rescue our ability to discuss—and combat—the underlying phenomenon that triggered the present use of the term, some philosophers have tried to characterize it more precisely. A common theme in this nascent philosophical discussion is that contemporary fake news is not a new kind of phenomenon, but just the latest iteration of a broader kind of phenomenon that has played out in different ways across the history of human information-dissemination technologies. While we agree with this, we argue that newer sorts of fake news reveal substantial flaws in earlier understandings of this notion. In particular, we argue that no deceptive intentions are necessary for fake news to arise; rather, fake news arises when stories which were not produced via standard journalistic practice are treated as though they had been. Importantly, this revisionary understanding of fake news allows us to accommodate and understand the way that fake news is plausibly generated and spread in a contemporary setting, as much by non-human actors as by ordinary human beings.

The standard view amongst philosophers of language and linguists is that the logical form of generics is quantificational and contains a covert, unpronounced quantifier expression Gen. Recently, some theorists have begun to question the standard view and rekindle the competing proposal, that generics are a species of kind-predication. These theorists offer some forceful objections to the standard view, and new strategies for dealing with the abundance of linguistic evidence in favour of the standard view. I respond to these objections and show that their strategies fail. I offer a novel argument in favour of the standard view that I call the binder argument. The upshot of this argument is that if one rejects the existence of Gen, then one is committed to rejecting the existence of covert structure in general.