Is This Work, or Is This the Gym?

Bruce Schneier offers a useful way to decide when to use artificial intelligence: ask whether the task is work or the gym.

The analogy, which Schneier credits to AI researcher Daniel Miessler, is simple.

When the objective is moving something heavy across a warehouse, use the forklift. The result matters more than the physical effort required to produce it.

At the gym, moving the weight is not really the objective. Lifting it yourself is the point because the effort develops strength. Having a robot lift it may produce the same movement, but it eliminates the benefit.

The same distinction applies to AI.

If a task simply needs to be completed reliably, using AI may be sensible. Reformatting data, producing repetitive documentation, converting notes into a consistent structure or handling routine administrative work can be treated as work.

But some tasks are valuable because of what happens while we struggle through them. Learning a programming language, debugging an unfamiliar system, developing an argument, writing to discover what we believe or practicing an artistic skill are closer to the gym.

When AI performs those tasks for us, it may complete the immediate assignment while denying us the understanding we were supposed to gain.

Schneier applies this to student writing. A policy memo assigned in a class is not needed because the world lacks another policy memo. The process of researching, outlining, arguing, revising and finding the right words is intended to develop the student’s ability to think.

This is also a useful way to think about software development.

Sometimes I need code produced, tested and shipped. Other times I am experimenting because I want to understand WebGL, Swift, networking or how a particular system behaves. Those may look like the same activity from the outside, but they have entirely different goals.

Most real tasks are not purely work or purely gym. AI can remove repetitive effort while leaving the important thinking with the person. It can review an argument after it has been written, suggest tests after someone understands the implementation or explain an error without replacing the act of debugging it.

A useful variation of Schneier’s question might be:

Which part of this task do I want completed, and which part do I need to practice?

There is also a second test. Even when a task is work, the AI still has to be reliable enough for the consequences involved. Its output must be reviewable, its mistakes correctable and the information given to it appropriate to share. Schneier makes clear that the work-versus-gym distinction only applies after those trust and security questions have been addressed.

AI should help remove drudgery. It should not quietly remove every opportunity to think, learn and maintain the skills we still care about having.

Read “Should You Use AI for a Task? Here’s a Simple Way to Decide.”