Anthropomorphism
2 writings
Familiar Words, Unfamiliar Minds: Conventionalization without Calibration in the Vocabulary of Mind
Under review
Argues that the extension of the vocabulary of mind to language models pulls apart two processes that usually coincide: conventionalization, whereby a stretched word ceases to be marked as a stretch, and calibration, whereby the inferences it invites are fitted to what its new referents bear out. Much as a gene variant can go to fixation without being an adaptation, a word can become conventional without being calibrated. A corpus study of “hallucinate” in 8,568 abstracts and 31,866 forum comments shows conventionalization running most of its course within three years. Calibration need not keep pace: it presupposes a referent that is accessible, stable, and independent of the discourse about it, whereas language models are opaque, fast-changing, and trained on that very discourse. Where the invisible hand does not suffice, calibration has to be organized through a reciprocal division of linguistic labour, guided by a norm of targeted qualification: flag the specific inferences that an attribution invites but the evidence does not support.
Download PDF
Disciplined Anthropomorphism: How to Extend Cognitive Language to Artificial Intelligence
Under review
Argues that those who describe AI in cognitive terms such as “understanding” and “believing” and the critics who denounce this as wishful anthropomorphism share a subsumptive approach: both ask whether AI systems fall under concepts minted for human beings, and report as discoveries about those systems verdicts that are largely determined by conceptual frames. Yet the problem AI poses is precisely how to frame these systems conceptually, and the rise of anthropomorphism is best seen as a response to that problem which needs disciplining rather than discouragement. A disciplined anthropomorphism disentangles the inferential connections bundled up in our cognitive concepts and asks which to preserve, sever, or forge for AI. The discipline comes from the human conceptual needs that give these connections their point, and from empirical and especially mechanistic evidence that supports and constrains them. The resulting conceptions of understanding and belief are not just thinned-out versions of existing ones, but potentially richer and better fitted to the peculiarities of AI.
Download PDF