A Model Is a Useful Lie

By the end of this lesson, you will be able to...

• Explain why a model is a purposeful simplification rather than a copy of reality.

• Identify the purpose, assumptions, output, and blind spots of a small model.

• Build a model card that another person can inspect and challenge.

Idea in one sentence: A useful model leaves most of reality out so that one question becomes easier to answer, but we must keep the missing parts visible.

Core Insight

Imagine Mara has a job interview at 09:00. The office is across town. She opens a transit app, sees a journey time of 30 minutes, and decides to leave home at 08:30.

The number looks precise. It is also incomplete.

The app has not placed Mara, the bus, every traffic light, the weather, and every other traveler inside a second city made of data. It has selected a few features: the route, the timetable, recent traffic, and expected walking time. It has ignored many others.

At 08:40, rain slows traffic. The bus arrives eight minutes late. Mara reaches the office after 09:00.

Was the app useless? Not necessarily. Was its 30-minute estimate reality? Also no.

The estimate was a model: a simplified representation built for a purpose. It compressed a messy journey into an answer to one question:

Roughly how long should this trip take under expected conditions?

The model gave Mara leverage. She could plan without inspecting every road herself. The same simplification created a blind spot: unusual delay mattered to her decision, but the estimate did not represent it well.

That combination is the central idea of this track:

A model is useful because it is smaller than reality, and dangerous when we forget that it is smaller.

The World Is Too Detailed to Carry Whole

Think about what a complete description of Mara's journey would require.

It might include the position of every bus, the timing of every signal, road works, rain intensity, driver behavior, passenger boarding time, lift failures, and Mara's walking speed. Even that list would be incomplete.

Mara does not need all of it. She needs a decision: when to leave.

A model performs compression. It keeps details that seem relevant to a question and discards the rest. This is not careless by default. Without compression, reasoning would never finish.

Plain meaning:

A model is a smaller thing that stands in for part of the world.

In this scenario:

The transit estimate stands in for the future journey.

Technical name:

This is a representation. It represents selected relationships in a target situation. It is not the target itself.

The target may be physical, such as a journey or a bridge. It may also be social, such as a team, a market, or a policy. A diagram, equation, story, dashboard, simulation, or verbal explanation can all function as models when they help us describe, explain, predict, or decide.

Why Call It a Useful Lie?

The title is deliberately sharp, but it needs care.

A model is not a lie in the ordinary sense of deliberate deception. Its individual statements may be true. Calling it a “useful lie” points to a different fact: every model acts as if some missing details do not matter for the current purpose.

A printed city map may show a road as a clean line. The real road has width, cracks, slopes, traffic, and changing conditions. The line is not a failed road. It is a useful simplification for navigation.

The phrase becomes harmful if it suggests that evidence does not matter. A model cannot excuse invented data, hidden manipulation, or claims known to be false. Simplification and deception are different.

Use this safer test:

What has this model simplified, and is that simplification acceptable for this purpose?

The answer depends on purpose. A route map may be useful for choosing streets and dangerous for checking wheelchair access. The same representation can be adequate for one question and inadequate for another.

The Smallest Inspectable Model

Before trusting a model, make its structure visible. A compact model card can do this.

This card is not a complete theory of model building. Later lessons will examine boundaries, variables, proxies, feedback, uncertainty, and competing models in more depth.

For now, notice the basic discipline: the output is not presented alone. It remains attached to a purpose and a set of conditions.

Worked Path: From Situation to Decision

Let us build Mara's first model step by step.

Starting point

Mara must arrive by 09:00. She wants a simple departure time, not a complete traffic simulation.

Step 1: Select the parts

She keeps three expected durations:

• walk from home to the stop: 5 minutes;
• wait and ride on the bus: 18 minutes;
• walk from the final stop to the office: 7 minutes.

Everything else is temporarily outside the calculation.

Step 2: Define the relationship

The naive model adds the durations:

5 minutes + 18 minutes + 7 minutes = 30 minutes

Step 3: Produce an intermediate result

Expected journey time: 30 minutes.

If Mara subtracts 30 minutes from 09:00, the model recommends leaving at 08:30.

Step 4: Expose the assumption

The calculation assumes expected conditions are good enough for this decision. That assumption was present before we wrote it down. Writing it down makes it inspectable.

Step 5: Compare with observation

Heavy rain adds eight minutes. The observed journey takes 38 minutes.

The difference does not tell us that addition is broken. It tells us that the selected inputs did not represent an important condition.

Step 6: Decide what changes

Mara has several choices:

• keep the simple model for ordinary trips;
• add a buffer when arrival time is important;
• use a weather or live-delay signal;
• choose a different route;
• leave much earlier when the cost of lateness is high.

The model has moved through a full path:

question
-> selected details
-> relationship
-> estimated journey
-> decision
-> comparison with reality
-> revision choice

The naive failure would be to treat 30 minutes as a fact about the future. The stronger approach treats it as an output produced under assumptions.

So far, we have seen that a model is not only an answer. It is a small reasoning machine that turns selected information into an output for a purpose. Its assumptions connect that machine to the world.

Purpose Decides What “Good” Means

People often ask whether a model is accurate. Accuracy matters, but the question is incomplete.

We should also ask: accurate enough for what?

Consider three purposes for the same journey:

1. Planning an ordinary commute. A typical-time estimate may be enough.
2. Arriving at a critical interview. The cost of being late makes tail delays and buffers important.
3. Evaluating accessibility. Average travel time says little about stairs, lift reliability, or walking distance.

One model does not automatically serve all three purposes. More detail can improve one use while making another use slower or harder to understand.

A good model is therefore not the largest model. It is a model whose simplifications fit its question.

Check: A restaurant uses last month's average lunch demand to order food for tomorrow. Is the average a model, a fact about tomorrow, or both?

Think first, then reveal.

Answer: It is a model input based on past facts. Tomorrow's demand is not yet a fact. Treating the average as a certain forecast would hide assumptions about season, weather, events, and changing behavior.

Assumptions Are Load-Bearing

An assumption is a condition the model accepts without representing or testing fully inside the model.

Assumptions are not automatically mistakes. Every usable model needs them. The problem is an assumption that is hidden, weak, or no longer true.

In Mara's model, “the route operates normally” is load-bearing. If the route closes, the 30-minute output loses its connection to the journey. The arithmetic remains correct, but the model is no longer useful for the decision.

This gives us a practical distinction:

• A calculation error means the model processed its selected pieces incorrectly.
• A model error means the selected pieces or relationships did not fit the target well enough.
• A use error means someone applied the model to a purpose or condition it was not designed for.

The three failures need different repairs. Recalculating does not fix a missing weather effect. Adding weather data does not help if the real question is accessibility. Better use begins by diagnosing which connection failed.

Trade-offs and Limits

The central trade-off is simple: simplification buys speed, clarity, and the ability to compare possibilities, but it also creates omission risk.

Adding detail can reduce a known blind spot. It also costs time, data, maintenance, and attention. A detailed model may become harder to inspect. It may even create false confidence because complexity looks like realism.

This lesson does not provide a formula for the perfect amount of detail. There is no purpose-independent perfect model. Instead, look for signals that the current simplification has reached its boundary:

• predictions repeatedly miss in the same direction;
• the decision changes when an omitted factor appears;
• two people use the same output for different purposes;
• an assumption is no longer true;
• the model cannot explain an important observation;
• people defend the output without naming how it was produced.

The goal is not to remove every blind spot. The goal is to make important blind spots easier to notice and revise.

Common Confusions

Confusion: A model is just an opinion

Why it is tempting:

Both models and opinions can be expressed in words, and both can be wrong.

Better model:

A useful model exposes structure: purpose, selected evidence, relationships, assumptions, and outputs. An unsupported opinion may have none of these. A model is still contestable, but it gives us something specific to inspect.

Confusion: More detail always makes a model better

Why it is tempting:

Reality contains detail, so copying more detail feels like progress.

Better model:

Detail is valuable when it improves the model's purpose enough to justify its cost. Irrelevant detail can hide the important relationship and make revision harder.

Confusion: If a model fails once, it is useless

Why it is tempting:

A failed prediction feels like proof that the whole representation was wrong.

Better model:

Failure is evidence about a boundary, assumption, input, relationship, or use. Some failures call for repair. Others show that the model should be replaced. The diagnosis comes before the verdict.

Confusion: The model's output is the world

Why it is tempting:

Numbers, diagrams, and dashboards look authoritative.

Better model:

An output is what the model says after processing selected information. Reality can still contain unrepresented conditions.

Check Your Understanding

Check: A hiring model ranks candidates using years of experience. The calculation is correct. Later, the team discovers that years of experience poorly represents skill for the role. Is this mainly a calculation error, model error, or use error?

Think first, then reveal.

Answer: It is mainly a model error: the selected proxy does not represent the target well enough. It may also become a use error if the team keeps applying the ranking after learning about the mismatch.

Check: Two people use the same weather forecast. One chooses whether to carry an umbrella. The other decides whether to evacuate a coastal town. Should the same forecast quality be sufficient for both decisions?

Think first, then reveal.

Answer: No. The purposes and costs of failure differ. A model that is adequate for a low-cost umbrella decision may be inadequate for a high-stakes evacuation decision.

Practice: Build a Model Card

A public library wants to decide how many staff members should work on Saturday. It has visitor counts from the previous eight Saturdays and a list of planned events.

Build a small model card. Include:

1. the target situation;
2. the decision the model supports;
3. two or three inputs;
4. one relationship or rule;
5. the expected output;
6. at least two assumptions;
7. one blind spot;
8. one update trigger.

A good answer should make these points visible:

• The purpose is staffing, not explaining everything about library use.
• Past visitor counts are evidence, not guaranteed future demand.
• Planned events may change expected demand.
• Assumptions could include similar opening hours or no unusual local event.
• A blind spot could be severe weather, staff skill mix, or an unlisted school visit.
• An update trigger could be a new event booking or a repeated gap between predicted and observed visitors.

There is no single correct model card. The quality comes from making the simplification inspectable.

Connection to the Next Lesson

Our model card already contains a target and blind spots. But we have not yet asked a harder design question:

Where does the modeled situation begin and end?

That is the subject of the next lesson. A boundary decides which people, forces, times, and consequences the model can see.

Resources

• [BOOK] The Model Thinker — Focus: Why different models reveal different parts of the same situation.
• [BOOK] Thinking in Systems — Focus: How purpose, boundaries, and assumptions shape what a systems model can explain.
• [ARTICLE] Mental Models I Find Repeatedly Useful — Focus: Practical examples of models as reusable tools rather than universal truths.

Key Takeaways

• A model is a purposeful simplification, not a copy of reality.
• A model becomes inspectable when its purpose, inputs, relationships, assumptions, output, blind spots, and update triggers are visible.
• “Useful lie” means deliberate simplification, not deception or freedom from evidence.
• A model is good only relative to a question and the cost of being wrong.
• Failure can reveal a calculation error, a model error, or a use error; each needs a different repair.