How to Tell When Your Numbers Are Lying to You

A short, practical guide to Goodhart’s Law and using AI to think past your stats and trackers.

TL;DR
Goodhart’s Law is a simple mental model that explains why your metrics start to drift away from reality the moment you treat them as the goal. When a measure becomes a target, people and systems learn to game it, and the number stops telling you the truth. In real life, that looks like hitting subscriber targets, revenue goals, or step counts while the underlying reality stays the same or quietly gets worse.

This guide turns Goodhart’s Law into something you can actually use, including AI prompts that help you stress test your metrics, see how they might be misleading you, and design small bundles of numbers that stay closer to what you really care about.

Use this guide to stop treating a clean number as the full story and to use AI as a second mind that keeps your measurements honest.

1. The Clean Number Problem

There is a special kind of relief that comes from a clean, impressive number.

A big green “+23%” on a dashboard.
A round subscriber milestone.
A benchmark score at the top of a chart.

You know it is just one slice of reality, but in the moment it feels like the whole truth.

Maybe you grow your newsletter list from 500 to 5,000 people. The graph looks great and the milestone feels real. But when you look closer, opens and replies have barely moved. You have more subscribers, but not many more readers.

Or a new AI model jumps to the top of a leaderboard. Screenshots of the score fly around. Then you try it on your own messy tasks and it feels only slightly better, if at all. The model has learned to perform on the test, not in the conditions you actually care about.

The same pattern shows up with step goals, grades, revenue targets, follower counts. You pick a number as a shortcut for something richer, then slowly start treating the shortcut as the goal. The map becomes more important than the territory.

You can hit the target and miss the point.

Goodhart’s Law is the name for this whole pattern. It explains why a metric that started as a useful shortcut can end up pulling you away from the thing you originally wanted.

AI lives right next to this problem. You can use it to polish the numbers, chase the benchmark, and make the graph look better. Or you can use it as a second mind that asks what the metric really means and how someone could hit it without changing the underlying reality. Used that way, AI helps you notice when a metric starts to lie and pulls your attention back to the real goal.

First we will make Goodhart’s Law simple. Then we will turn it into a few AI prompts you can use in the moments that matter.

2. Goodhart’s Law in plain language

The classic line is:

When a measure becomes a target, it stops being a good measure.

In plain language:

If you start chasing the score instead of the thing the score was meant to represent, the score stops telling you the truth.

Most of the time you pick a metric as a shortcut.

You track steps to be healthier, then end up pacing around your living room at night to hit a number while your sleep and stress stay the same. The step count moves. The underlying reality does not.

Goodhart’s Law is a reminder to keep asking:

“What did I hope this number would stand in for, and is it still doing that job?”

When it helps

Use Goodhart’s Law when:

• You are choosing what to track or what to tie rewards and praise to

• A metric looks good but your gut tells you something is off

• You notice your behavior changing just to make a graph look better

In those cases, Goodhart’s Law is a small brake pedal. It nudges you to check whether the metric is still connected to the thing you actually care about.

When it should not be used

Be careful with this lens when:

• You are tempted to say “all metrics are fake” and ignore them completely

• The main issue is obvious noise, a bad goal, or politics and incentives, not the metric becoming a target

• You are in a hard constraint situation where a single blunt number really does decide survival for now, for example cash in the bank or runway

Goodhart’s Law is a first lens, not a reason to throw away numbers or become cynical about all data. It simply reminds you that any metric is a map, not the territory.

Next, we bring this into practice with three prompts you can use with any AI assistant.

3. Prompts to use Goodhart’s Law with AI

The simplest way to use this mental model with AI is to literally say:

Apply Goodhart’s Law and help me check how this metric could be lying to me.

You describe the situation, add that one line, and let the assistant push back on your first interpretation of the number.

Pro tip: if your AI assistant can read images, you do not have to type everything out. Take a screenshot of your dashboard, paste it in, and say:

Look at this dashboard, apply Goodhart’s Law, and help me see what these numbers might be missing or hiding.

The AI can then reason about the whole picture instead of just one isolated metric.

These are copy paste prompts. Save or screenshot them so you can reach for them the next time a graph looks good, feels wrong, or starts to control your behaviour more than you like.

3.1 Quick check for a metric you rely on

Use this when there is a number you look at often and treat as a verdict. That can be steps, revenue, subscribers, hours worked, a score on a screen, or anything similar.

Apply Goodhart’s Law to this metric.

1. Restate the metric and what it is supposed to stand for.

2. List a few ways I could hit this number without actually improving the underlying reality.

3. Suggest one or two extra signals I could track next to it so I see more of the full picture.

Metric and context:
[describe the metric, how you track it, and what you think it means]

If you can see many ways to hit the number without doing the real work, you are probably asking too much of that single metric.

3.2 Noise or Goodhart for spikes and slumps

Use this when a number suddenly jumps or falls and you feel a strong urge to react right away. That could be a spike in followers, a drop in sales, a sudden change in open rates, or a sharp move on any dashboard.

Apply Goodhart’s Law to this spike or slump.

1. Summarize what changed and over what time period.

2. List a few normal explanations that do not mean anything deep yet.

3. List a few Goodhart-style explanations where the metric could move while the underlying reality does not improve.

4. Suggest one simple check or small experiment I can run before I change my plan.

Situation:
[describe the metric, the change, and why it bothers or excites you]

You are not asking AI to tell you if the spike or drop is good or bad. You are asking how many stories this one move can support and what tiny bit of extra evidence would make your next decision less reactive.

3.3 Triangulation: a small bundle instead of one target

Use this when you want to keep using metrics, but you no longer trust one number to carry the whole meaning. This works for health, money, learning, creative work, or teams.

Help me design a small bundle of metrics using Goodhart’s Law.

1. Restate what I actually care about in plain language.

2. Suggest 3–5 simple metrics that each capture a different side of that goal.

3. Mark which one should be the main signal and which ones should be guardrails that warn me if I am damaging something important.

4. Give one example of how things could look if I were accidentally Goodharting the main metric while the guardrails show the damage.

Goal and current metric:
[describe what you really care about and the main number you currently look at]

Think of this as a small triangulation bundle: a few metrics that together point at the underlying reality better than any single number.

You do not need an elaborate dashboard. A small bundle might mean something as simple as tracking both “hours worked” and “hours of deep work,” or both “new customers” and “customers who are still with me after six months.”

4. Principles and traps

A few simple rules keep this lens practical.

Principles

• Name the real goal next to the metric
  “This is a shortcut for: [health, trust, learning, profit, connection].”

• Prefer “active” over “total”
  Active readers over total subs. Returning customers over signups. Deep work over hours at desk.

• Pair numbers that check each other
  Growth with retention. Speed with satisfaction. Revenue with refunds.

• Notice when you are acting for the graph
  If you would not do it without the dashboard, pause.

• Let AI be the skeptic
  “Apply Goodhart’s Law and help me see how this metric could be misleading me.”

Traps

• Metric nihilism
  “All numbers are fake.” You still need signals, just not blind faith.

• Metric overload
  Twenty stats no one really watches. Better to have three you respect.

• AI as a spin machine
  Asking it to justify the number instead of question it.

• Blaming the metric for an incentive problem
  Sometimes everyone already knows the target is bad but the bonus depends on it.

• Using Goodhart to ignore clear, long term data
  A bad pattern for months is not “just metrics,” it is information.

Used well, Goodhart’s Law keeps your numbers in their proper place: helpful, visible, and never the whole truth.

5. From prompt to practice

At first you will remember Goodhart’s Law only in hindsight.
Then you will remember it when you are already staring at a graph.
Later you will notice the feeling, open your AI, and run one of these prompts before you decide what the number means.

Goodhart’s Law only really helps if you use it on a real number in your life.

Pick one metric you actually check:

• A health app number

• A work dashboard metric

• A follower or subscriber count

• A revenue or savings target

Then take it to your assistant, describe what you track and what you hope it means, and ask it to apply Goodhart’s Law and show you how the metric might be misleading you or getting gamed. If you have a dashboard, you can add a screenshot instead of describing everything.

Read the reply and pick one thing to change. That might be:

• A second metric you start watching

• A behavior you stop doing just for the graph

• A clearer sentence for what “success” really is here

You are not trying to fix all your metrics. You are just practicing one move:

Use AI to question the number before you let it define the story.

Most tools use AI to optimize your metrics. You are using it to keep them honest.