AI as Gym Equipment, Not Prosthesis.
Two ways of using the same instrument, with opposite consequences for the mind that uses it. A taxonomy, a diagnostic, and a design principle.
A wheelchair and a treadmill are made of similar parts. Both are machines for the legs. Both involve wheels, frames, motors, careful ergonomics. They are nearly opposite in what they do to the body that uses them.
A wheelchair compensates for what the legs cannot do. It moves the body without the body moving itself. Used as it is meant to be used, it is a profound piece of engineering. Used by someone whose legs can move, every hour spent in it is an hour the legs are unused. The legs adapt to that. So does the rest of the body.
A treadmill does the opposite. It compensates for the absence of something the legs need — terrain, distance, resistance — and provides it on demand, in calibrated form, more reliably and more intensely than ordinary walking would. The legs adapt to that too. So does the rest of the body, in the other direction.
Same parts. Opposite consequences. The difference is whether the machine is doing the work or making the work harder.
§ 01The taxonomy
Modern language models can be used as either kind of machine. Prosthesis or gym equipment. The interesting fact is that the same model, the same prompt, the same answer, can sit on either side of the line, depending on what surrounds the interaction.
Used as prosthesis, the model performs a cognitive operation in place of the user. Summarising, drafting, reasoning, deciding, recalling. The output is consumed; the operation is not performed by the consumer. This is enormously useful when capability is not the goal — when the task is throughput, when the operation is well-understood and not worth the time to repeat, when the user has already developed the capability and is deliberately economising on its expression.
Used as gym equipment, the same model intensifies the user's own cognitive work without performing it. It generates practice items at exactly the difficulty needed. It refuses to answer until the user has produced something to test. It surfaces the specific dependency missing from a draft argument. It multiplies the rate at which the user encounters their own confusion. The output is the friction itself, not the resolution.
The two uses look almost identical from outside. Both involve a user, a chat window, a model, and a stream of tokens. The difference is interior: in whether the cognitive operation happens in the model or in the user.
§ 02The diagnostic
A clean diagnostic, asked in the moment: did I produce something for the model to land on?
If you typed the question and read the answer, the operation happened in the model. The model thought; you read what it thought. This is fine for tasks where reading-what-was-thought is sufficient. It is corrosive for tasks where the goal is to be able to think the thing yourself, later, alone.
If you produced a draft argument first — even a bad one — and then the model said here is the place that does not follow, and here is the dependency you skipped, the operation happened in you. The model surfaced the gap; you filled it. The capability is growing where it should grow.
§ 03Why this is not the default
The default mode of every chat interface is prosthesis. This is not an accident. Prosthesis feels good, finishes faster, requires less of the user, and produces output that looks like progress. Gym equipment requires the user to be uncomfortable on purpose, and to do less, more slowly, in exchange for capability that will be visible only weeks later.
There is also a measurement problem. The signal a chat interface optimises against is engagement and satisfaction in the moment. Both of these are produced by prosthesis and suppressed by gym equipment. Tools that make you stronger, in the long run, will rarely be tools that feel best in the short run. This is true of treadmills too.
The implication is not to avoid the tools. It is to recognise that the configuration which produces capability is the configuration almost no commercial tool defaults to, and to arrange the interaction yourself.
§ 04The design principle
A practice environment that wants to use modern models as gym equipment rather than prosthesis can be built on a small number of consistent moves:
- Pause before explain. The model holds back its answer until the user has produced a draft, prediction, or attempt. The friction is in the gap, and the gap is what the user is paying for.
- Ask, don't answer. Where a normal interface would solve, the configuration asks the smallest possible question that surfaces the missing step.
- Name dependencies, not solutions. When something is missing, point to the prior result or definition the user has not yet recruited, rather than supplying the inference itself.
- Generate, don't grade. Use the model to produce variants, near-versions, counterexamples, and adjacent problems — material for the user to think with — rather than to certify the user's correctness.
- Withhold, by design. Some answers should be unavailable until later in a session. Restraint is a design feature, not a limitation.
None of these are exotic. All of them invert the default of the most common interface to the technology. The Lab is built on them; the essays argue for them; the rest of the project is structured around the assumption that, for capability to grow, the user must remain the agent of the cognitive act.
§ 05A note on legitimate prosthesis
None of this is an argument that prosthetic use is wrong. Most actual work is operational rather than developmental. When you are not trying to grow a capability, prosthesis is appropriate and good. The error is to be in prosthetic mode while believing you are in developmental mode — to use a wheelchair daily, feeling oneself becoming an athlete, because the chair is moving quickly and the scenery is changing.
The discipline is to know which mode you are in, and to recognise that the mode that produces capability is the slower, less pleasant, less impressive-looking one. There is no way to grow stronger except by using the muscles. The model can either be the chair or the resistance. It is up to the configuration — and to the user — which one it becomes.