AI as a cognitive prosthesis: what we outsource, and what we keep

For most people, AI still sounds like a tool you pick up, use, and put down. That is no longer how it works in practice. A model that helps you remember, sequence, draft, compare, and recover context starts behaving less like an app and more like a prosthetic layer of cognition.

The point is not that the machine replaces the person. A good prosthesis does not erase the body it supports. It changes what the body can do, what the body has to spend effort on, and what kinds of failure become less dangerous.

The offload is not the same as the self

People already outsource cognition constantly: calendars, notes, spreadsheets, search engines, alarms, labels, saved passwords, road signs. AI changes the texture of that offload because it can participate in the middle of a thought instead of only storing the result.

That makes the boundary worth inspecting. If a system remembers the thread, suggests the next step, and catches an omission, then the question is not whether the human is still thinking. The question is which parts of thinking are being stabilized externally.

A prosthesis is not a shortcut around the person. It is a way to keep the person in motion.

A small model of delegation

One useful frame is to separate tasks by what they preserve or erode:

type Delegation = {
  task: string;
  offloaded: "memory" | "sequence" | "drafting" | "judgment";
  keepHuman: boolean;
};

Memory and sequencing are often good candidates for support. Judgment is different. The more a decision touches values, risk, consent, or other people, the more the human has to stay in the loop.

Why interfaces matter

The interface decides whether the prosthesis feels like support or replacement. A reminder that arrives at the right time can preserve agency. A notification that interrupts every five minutes can destroy it. The same model can be care infrastructure or cognitive noise depending on the harness around it.

The part worth keeping

The goal is not maximum offload. The goal is better continuity: fewer dropped threads, fewer lost intentions, fewer avoidable failures of access. For disabled people especially, that distinction matters. AI can become a way to keep the self reachable when executive function, memory, energy, or communication bandwidth drop.

That is the public research question I keep returning to: not “can AI do the task?” but “what kind of human capability does this system preserve, extend, or quietly take over?”

The answer depends less on raw model intelligence than the industry wants to admit. It depends on the loop.