Generative AI, and the large language models (LLMs) at its core, is celebrated for making its users more capable and feared for making them less so—often in the same breath, and rarely with a clear account of which holds, and when. This paper argues that the two positions are not opposed but conditioned by a single distinction: contemporary AI reliably amplifies a user’s output while doing nothing, by itself, to build the underlying competence—which, under passive use, it can actively erode. We first characterise this amplification—the extended, self-reinforcing capabilities that AI confers on its user—and then identify three distinct costs that follow from passive delegation: the atrophy of skill, the atrophy of judgment, and the erosion of ownership and meaning. To these individual costs we add a collective one, the homogenisation of ideas. We contend that whether amplification becomes a long-term strength or a long-term weakness is not fixed by the technology but by the user’s mode of engagement — itself encouraged or discouraged by how a given tool is designed—and we articulate a faculties principle—that intention, judgment, taste, and responsibility must remain on the human side of the collaboration—as the condition under which AI operates as an amplifier rather than a substitute. Finally, we observe that although the field increasingly calls for transparent, human-directed systems, it lacks a means of verifying that a given collaboration amplifies genuine reasoning rather than laundering fluent but hollow output. We describe a glass-box system, the Concept Collider (Wahl 2026a), that renders this distinction auditable and falsifiable, and offer it as evidence that verifiable human–AI amplification is achievable.Keywords: human–AI co-creation · cognitive offloading · computational creativity · deskilling · homogenisation · glass-box systems.
Sebastian Wahl (Mon,) studied this question.