Thinking About Thinking With Machines That Think
Abstract
Intellectual work increasingly occurs in settings where large language models (LLMs) can produce competent text across a wide range of tasks. We argue that the central design problem is not capability but the preservation of plurality: sustaining individual perspective, judgment, and authorship in human–AI collaboration. Building on the view of LLMs as simulators (janus, 2022; Shanahan et al., 2023), we treat model outputs as samples from a latent distribution over possible authorial personae that is further shaped, in deployment, by post-training into a dominant assistant mode. We characterize process flattening as a failure of the developing scientific persona: the scholar’s constructed capacity to differentiate their own epistemic commitments from the statistical consensus of the training distribution. In the absence of directed steering, output regresses toward the distribution’s central tendency, thereby bypassing the developmental friction through which an independent scholarly identity is ordinarily forged. We analyze these dynamics in a prototype doctoral intensive in which neuroscience PhD students (N=7) wrote essays on thinking with AI at every stage. Using a blind-prompt protocol that records intentions before process begins and compares what the process produced against the model’s best output from the same frozen seed, we document differences in how participants narrowed the output space and committed to positions in ways attributable to individual judgment. We argue that effective collaboration is reciprocal: humans supply selection and commitment, while AI extends search and synthesis; decomposition is therefore required not to protect human relevance but because the combined system depends on it.