Contemporary debates about AI failures often oscillate between attempts to repair algorithmic systems and calls to refuse or abandon them. Yet these diagnostic responses presuppose particular understandings of what it means for a technology to fail. This paper offers a philosophical and historical analysis of the concepts through which technological fallibility is understood. Drawing on Yuk Hui’s conceptual history of mechanism and organism and historical scholarship on cybernetics and machine learning, the paper traces the emergence of algorithmic error as a historically specific concept of technological fallibility. It argues that, unlike the (post-)phenomenological and mechanistic conceptions of breakdown, the concept of algorithmic error is organized around the recursive incorporation of contingency rather than its manifestation as rupture. This historical transformation does not turn the (post-)phenomenological or mechanistic understandings of breakdown obsolete. Still, it limits their ability to conceptually exhaust the forms of fallibility characteristic of contemporary algorithmic systems. The paper then examines the implications of this shift for concepts of repair and refusal, as they figure in the debates on AI failures. It argues that repair becomes less a matter of restoring a previous state than of intervening in the conditions under which recursive learning occurs, while refusal shifts from contesting erroneous outputs toward questioning the assumptions embedded in the formulation of computational problems.
AI alignment often involves the training of AI systems to better accord with the preferences of human annotators. Pattison argues that the comparative nature of such alignment methods can be both understood and justified through the lens of non-ideal theory. We contend instead that preference-based alignment cannot be characterized as fully non-ideal, in part because of the consequentialist commitments it carries, and in part because of the role that normative principles play in how humans express preferences over AI behavior. Reflecting on the ideal way to apply and adjudicate these principles, we argue, is useful and sometimes crucial for making these judgments, especially in hard cases with conflicting principles or harms. Furthermore, when determining the principles that AI systems should be aligned to, it pays to reflect upon the ideal roles or functions that AI should serve in a just society. While this may fall short of a fully specified ideal theory, we argue that this principle-based approach to AI alignment ultimately provides richer normative resources than approaches that eschew ideal theorizing or rely solely on consequentialist or harm-based justifications.
Large language models are increasingly taken to be capable of semantic practices. Debate about what to make of this has largely been conducted as a dispute about understanding. I argue that these standard framings, whichever side one takes, leave out an essential component of discursive reasoning: responsibility. To assert something is to make oneself answerable for what is claimed, to stand ready to defend it, and to give up a previously-asserted claim that cannot be defended. What settles whether a system is a participant in discursive practice is therefore not merely whether it understands, nor merely how well it performs, but whether it can be held to what it has claimed. Drawing on Brandom’s inferentialism, I develop five criteria for responsible reasoning, each of which admits of degrees. What they track is not independence from a community, since no speaker’s discursive capacities are independent of the community that institutes the norms she is answerable to, but the degree to which an agent does its own deontic scorekeeping. The central thesis I defend is accordingly agent-neutral: an agent whose performances fail these criteria is not, to that extent, reasoning responsibly.
Stone’s generative commons require independently observable mediation. This commentary distinguishes auditing exclusion rules from measuring informational visibility and argues that European disclosure and data-access obligations only partially support these assessments.
Behavioral data machine learning (ML) systems make up a common class of ML models designed to capture user behavioral data from retail websites, social media platforms, and other digital interfaces and to translate this data into behavioral predictions about and interventions on users. A common objection to such systems is that they undermine users’ autonomy by confining them to a feedback loop informed by desires they may not endorse from a reflective perspective. I complicate this picture in two ways. First, because behavioral data ML systems act as a kind of mirror of an agent’s desires, they can uncover desires of hers that were hitherto obscured from her own awareness. In so doing, they can foster a kind of self-knowledge. Second, although the agent might reflectively endorse or reject such revealed desires, she might instead maintain ambivalence towards them. I explore the implications of this observation. Though often regarded as irrational, ambivalence can be a warranted attitude to maintain towards inconsistent desires. Behavioral data ML systems, however, are likely to appeal to one rather than the other of two inconsistent desires towards which the user is ambivalent. One effect may be that an ambivalent user’s desires, or her higher-order attitudes toward them, subtly shift, leaving her more strongly inclined towards one inconsistent desire than the other. In the case of warranted ambivalence, this effect is troubling not because it compromises an agent’s autonomy but rather because it compromises an agent’s ambivalence itself, resolving it, but not aptly, not reasonably.
