We’ve all seen some version of the Dunning–Kruger graph; that “Mount Stupid” meme where confidence shoots up early, crashes, and then slowly rises again as people become experts.

The funny part is:

• That exact red curve was never in the original paper.
• It was drawn later, without direct supporting data.

But the shape still feels very familiar.

The question that started this project was:

If the shape feels right, could it actually be mathematically real viewed from a different angle?

This is the “fake” but famous version of the Dunning–Kruger curve:

• Confidence starts near zero,
• Peaks early when knowledge is still low,
• Drops into a valley,
• Then rises again toward expert level.

On the next slide I recreate that shape in R so we have a clear reference.

In the actual Dunning–Kruger study, the main figure looked more like this:

• One line for Perceived Ability (what people think their percentile is).
• One line for Actual Test Score (their true percentile).
• The x-axis is split into bottom, 2nd, 3rd, and top quartile groups.

We don’t need the raw dataset to work with this. We can read approximate values off the graph and reconstruct those quartile points in R.

Here I approximate the quartile values from the original-style figure:

• Bottom quartile: perceived ≈ 60, actual ≈ 12
• 2nd quartile: perceived ≈ 62, actual ≈ 35
• 3rd quartile: perceived ≈ 70, actual ≈ 60
• Top quartile: perceived ≈ 75, actual ≈ 90

Now comes the key step: how to define confidence.

Instead of taking self-reported confidence directly, I model perceived confidence as the distance from truth:

Let:

• ( T ) = true value (actual test score)
• ( E ) = estimate (perceived ability)

Then define:

[ C = k , |E - T|]

where (k > 0) is just a scaling constant.

• People can overestimate (too high) or underestimate (too low).
• In both cases, they can still feel very sure.
• The magnitude of (|E - T|) captures how misaligned belief is with reality.

Using the reconstructed quartile values:

• Bottom quartile: (|60 - 12|)
• 2nd quartile: (|62 - 35|)
• 3rd quartile: (|70 - 60|)
• Top quartile: (|75 - 90|)

I also add a conceptual point at true ignorance:

• Knowledge = 0, Confidence = 0

This point doesn’t come from the study; it comes from the idea that someone who truly knows nothing and admits it has zero perceived confidence.

conf_quartile <- abs(perceived - actual)

# skill positions mapped to [0,1]
skill_q <- c(0.25, 0.50, 0.75, 1.00)

# anchor points including true ignorance
key_skill <- c(0, skill_q)
key_conf  <- c(0, conf_quartile)

data.frame(skill = key_skill, confidence = key_conf)

##   skill confidence
## 1  0.00          0
## 2  0.25         48
## 3  0.50         27
## 4  0.75         10
## 5  1.00         15

Now I interpolate a smooth curve through those anchor points:

• (0.00, 0) – true ignorance
• (0.25, |60−12|) – bottom quartile
• (0.50, |62−35|) – 2nd quartile
• (0.75, |70−60|) – 3rd quartile
• (1.00, |75−90|) – top quartile

This gives an expected confidence at any knowledge level between 0 and 1.

Here is an interactive version of the same curve so you can explore points and regions more closely.

On average, the derived confidence curve can be thought of as a simple function that is high when misalignment is high, and lower when belief is closer to truth.

In words:

\[ C(x) \text{ is large when } |E - T| \text{ is large,} \] where people are confident but still far from reality, and

\[ C(x) \text{ is small when } |E - T| \text{ is small.} \]

The exact formula is less important than the idea:

• Confidence is driven by distance from reality,
  
• Not by how much we actually know.

This doesn’t just apply to psychology. Any system that learns, human or AI, uses some notion of confidence to decide:

• What to trust,
• What to ignore,
• And what to update.

If confidence tends to increase the farther we drift from truth:

• Errors become harder to correct.
• Misinformation is more likely to “stick.”
• The system (or person) becomes harder to train over time.

Healthy learning needs confidence that tracks truth, not just the strength of our beliefs.

Even accurate data can mislead when the frame around it is incomplete.

The Dunning–Kruger story isn’t just about confidence, it’s about how interpretation shapes truth.

We often think we’re looking at evidence, when we’re really looking at it through ourselves.

This study also highlights something deeper:

Ignorance isn’t about how much you know, it’s about how true what you know actually is.

A person can be fully confident in something that’s objectively false, and that confidence still feels like knowledge.

It’s this very misalignment between belief and truth that reveals how to properly gauge confidence, not by how much one knows, but by how closely that knowledge aligns with reality.

That misalignment is both the source of overconfidence and the proof of this model’s validity.

What began as an investigation into a fake graph ended up revealing a real truth hidden in plain sight. By redefining confidence as distance from reality, the same “Mount Stupid” curve emerged naturally from the original Dunning–Kruger framework itself. And with it came a deeper lesson:

• Even accurate data can lead to false conclusions
• A study about misjudging knowledge was misjudged itself.
• A graph condemned may be right from another angle.
• The illusion of confidence became its own proof.

In the end, confidence didn’t just explain the Dunning–Kruger effect, it embodied it.

Confidence, it seems, loves irony.

What I love most about this study is how it proved its own point.

It began as a correction to a misunderstanding, and ended up revealing that even a misunderstanding can hold truth when seen from the right angle.

It reminded me that confidence isn’t the opposite of ignorance; it’s often what bridges the two.

Because sometimes the only thing separating truth from error is how willing we are to question what feels right.