A cochlear implant is a thin electrode array that a surgeon threads into the spiral cavity of the inner ear. It’s delicate work, and applying the right amount of force is critical. Pushing too hard can damage the internal structures, which is part of why residual hearing loss is a real risk for this procedure.

So researchers need a way to test designs without using a person. Cadaver tissue is scarce, variable, and doesn’t always mimic real patients accurately. The next best option is to physically model it as closely as possible. Scan a real cochlea, 3D print the spiral, insert the electrode, measure the force.

Even though the geometry was right, the mechanics were wrong.

The researchers had to adjust the surface properties and add a lubricant before the model began behaving more like the anatomy it was supposed to represent.

I like this example because it shows why calling a model “realistic” does not tell us very much. A model can be highly accurate in one dimension while being misleading in another.

It’s always important to consider if a model or proxy is representative of reality in the ways you actually care about, rather than just the ones that are easy to measure.

Let’s consider a subway map as an example.

It distorts distance, direction, and geography. Two stations that appear close together may actually be far apart, and a curved route may be shown as a simple straight line.

None of that makes it a bad model.

The map preserves the information passengers usually need, like the order of the stops, the connections between lines, and the places where they can transfer. It is useful for navigating the subway, even though it would be a poor tool for estimating walking distance.

The same distinction applies to prototypes, simulations, forecasts, dashboards, and benchmarks.

A 3D-printed part might be good enough to confirm it will fit in its intended location, but that does not mean it can predict how the final part will flex, wear, transfer heat, or fail under repeated loading.

A financial forecast might help estimate whether a business can cover its expenses. It may be much less useful for predicting exactly how much revenue the business will produce in a particular month.

A model should be judged relative to the job it is being asked to do.

Before relying on a model, it helps to ask:

The fourth question can really help for making adjustments after initial testing.

This does not mean that every model should include every possible variable.

A perfect reproduction of reality would no longer be a useful simplification. It would also be expensive, difficult to understand, and nearly impossible to build.

The goal is to include enough detail to support the intended decision without adding complexity that does not improve the result.

That makes model fidelity something we have to define, rather than something a model simply has.

Whenever you use a score, forecast, prototype, simulation, or dashboard, it is worth asking what that model preserves and what it ignores.

The answer may matter more than how accurate the model appears.

The next wave of millionaires will be people who figured out how to make AI work for them.

The window to get ahead is still open. But not for long.

Here are 200+ proven ways to make money with AI in 2026.

Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.

Claim your free list

Got a second? Give some feedback on today’s article so we can keep making improvements to The Manifold.

PS — I love 3D printing as much as the next guy, but like anything, it has its limitations.

The Future Today: From the president of OpenAI to the CEO of Google, we talk to the actual people changing the world with AI. And then we put all of their insights, advice, and business ingenuity into our newsletter. Join 75K Readers →