Clark2025ExtendingMinds
Andy Clark (2025) Extending Minds with Generative AI
Bibliographic info
Clark, A. (2025). Extending Minds with Generative AI. Nature Communications, 16, 4627. DOI: 10.1038/s41467-025-59906-9.
Commentary
The essay relies on a concept it does not fully explore. Clark aims at the self-image that connects you to your biological brain. He argues that generative AI extends thought instead of replacing it. However, his way to protect against real risks, such as poor outputs, rigid methods, and dependence, is metacognitive judgment. This is the ability to know what to trust and when. That skill needs a home; it cannot be handed off. If you assign the verification to another AI, you must develop a new skill to determine when to trust it. In dealing with danger, the extended mind reaches for a non-transferable, brain-based supervisor—the very simple self the essay intends to break down.
The idea of "borderline-you" complicates this further. A personal AI that shapes your decisions from a young age and is considered "implicitly trusted" is something you cannot verify. It becomes part of the self that is supposed to do the verifying. Clark also warns early about "algorithms that do not have our best interests at heart," but then builds his optimistic view on commercial products without revisiting that point.
Excerpts & Key Quotes
The Friendliest Possible Test Case
- Page 2:
"it seems as if the AI-moves helped human players see beyond centuries of received wisdom so as to begin to explore hitherto neglected (indeed, invisible) corners of Go playing space."
Comment:
Go seems to be the best possible example that he could use, and this is the problem. There is a clear binary outcome in a closed system with an immediate and objective response: you either win or lose. And this means that the game will confirm the predictions made by the AI without the need for judgment. Then Clark applies the idea of "extension but not replacement" to areas such as art, medicine, and writing a complaint letter. He makes his point about trust in one situation where trust does not matter and extends it to situations where it matters and is hard to judge.
A Safeguard That Only Moves the Problem
- Page 3:
"We must also learn to adjust our levels of trust according to our choice of portals and wrap-arounds. Some of these, like FunSearch, can already help mitigate at least some of the risks."
Comment:
Clark’s solution to “Can I trust the AI?” is to employ a second tool to vet the first one. But this does not solve the issue of trust; it merely shifts it: Now the issue is how you will know when to trust the wrapper. Layer on yet another wrapper, and the issue shifts again. Ultimately, the regression ends in an evaluation by a human that cannot be delegated, which is why his framework implicitly requires the evaluator who is not extended.
The Off-Switch Someone Else Holds
- Page 3:
"You survive their loss or deletion, but much as you would a minor stroke... only in the same attenuated sense as you 'use' your hippocampus or frontal lobes."
Comment:
Metaphor works silently here. An AI loss does not parallel any minor stroke since a stroke cannot be inflicted upon you or taken away at any time by anyone for gain. A cloud-based "borderline-you" can be turned off, re-priced, subpoe-naed, or re-trained by its owner. When comparing an outside service to something integral to your being, Clark equates them cognitively, making sure the only distinguishing factor between the two is lost – an ability to turn off a part of your brain.