Inga hears there is an exhibition at MoMA, recalls that the museum is on 53rd Street, and goes. Otto has Alzheimer's and carries a notebook everywhere, in which he writes down what he learns; he consults it, reads "53rd Street," and goes. Andy Clark and David Chalmers (1998) argued that both believed where the museum was before recalling or consulting: Otto's notebook plays the role of Inga's memory. Hence their parity principle: if something outside the head does a job that, done inside, we would unhesitatingly call cognitive, it is part of the cognitive process. The mind does not stop at the skull.
Not every tool qualifies. In Otto's case, the notebook is a constant in his life, its information is available without difficulty, he automatically endorses what he reads, and he wrote it down himself in the first place. The authors already warned that the Internet would probably fail several of these conditions, except for someone who relied on computers heavily, used them fluently and trusted them. Adams & Aizawa (2001) replied with the coupling-constitution fallacy: that a process is causally coupled to something external does not make that thing part of cognition.
The thesis became an everyday matter with AI. On the human side, a conversational assistant could be Otto's notebook; Clark (2025) has extended the idea to generative AI. But the condition of automatically endorsing what the resource says is precisely the one that hallucinations make dangerous. On the machine side, a RAG system separates parametric memory (the weights) from non-parametric memory (a document index) (P. Lewis et al., 2020, sec. 1), and an agent with tools, memory files and web search spreads its computation across network and environment.
Clark and Chalmers's question becomes a technical one: where does an artificial agent end, in its weights, its context window, its vector database, or the Internet? And where do we end when we think with one?