What Is the Turkey Problem? Nassim Taleb's Definition

The Turkey Problem is an epistemological challenge: a turkey is fed by the butcher for 1,000 days. Each day confirms with increasing statistical confidence that the butcher loves turkeys. On day 1,001 — Thanksgiving — the turkey's model of the world is catastrophically wrong.

The turkey's mistake: mistaking absence of evidence for evidence of absence. Past calm was taken as proof of future calm.

The Epistemological Error

The turkey has data. 1,000 days of data. The data is consistent. The model fits perfectly.

But the data was generated by a process with a hidden structural break (day 1,001). The turkey's probability model is correct given the data, but the data is irrelevant to the actual distribution the turkey faces.

More data and higher confidence can lead to more catastrophic error if the data is unrepresentative.

Real-World Examples

Before 2008: Historical housing price data showed only rises. Models built on this data predicted continued rises. The data didn't capture the regime change.

Before the pandemic: Historical pandemic data showed large outbreaks every few decades. Models predicted this pattern would continue. They didn't capture how a novel pathogen could spread globally.

Employees before layoffs: 20 years of evidence that the job is secure. The data doesn't capture the acquisition or market shift that changes everything.

Non-Turkey Thinking

The antidote isn't better prediction. It's different question-asking.

Instead of: "What does the history say?" ask "What would invalidate my current model? What would a breakdown look like?"

Instead of: "How confident am I based on past data?" ask "How catastrophic would I be if I'm wrong?"

Turkey thinking says: long history of safety = safe.

Non-turkey thinking says: long history of safety until regime change = fragile.

Go deeper:

For the full breakdown of the Turkey Problem and non-turkey wisdom, read The Turkey Problem: When Past Data Fails.