SECAI Core · phase 6 of 15 · noun

Model bias

An error that can be introduced when approximating a real-world problem with a simplified model or a dataset that does not include all relevant factors to make accurate predictions.

The Explain card

Plain English
Model bias is systematic error introduced when a real-world problem is approximated with a simplified model, or trained on a dataset that misses factors needed for accurate predictions.
Example
A fraud detection model is trained mostly on transactions from one region. It learns that unfamiliar geographies look suspicious and starts flagging legitimate customers abroad while missing fraud patterns it never saw.
Why it matters
Biased models fail in predictable directions, which attackers can learn and exploit and which can harm specific groups of users. Defenders examine training data coverage and test performance across segments, not just overall accuracy.
Hook
Model bias is a blind spot baked in at training time; the model cannot see what it never learned.

Word knowledge

How the word is built, where it came from, and what it sits beside in memory.

In a sentence

The evaluation notebook on the malware pipeline reports model bias when Linux ELF samples miss more often than Windows PE files on the same holdout set.

Why these words
  • model French modèle, from Latin modulus, a small measure marks the slant as belonging to the mathematical stand-in rather than to a person
  • bias French biais, a slant or oblique line supplies the head noun, the slant the phrase pins on that stand-in
Where it came from
Origin
English statistical compound from French modèle (Latin modulus, a small measure) and French biais (a slant, an oblique line)
Entered the language
1950s
What changed
Model had meant a small measure, then a miniature to copy, then a mathematical stand-in. Bias had meant a diagonal cut, then the weight that makes a bowl curve, then a one-sided mind, then a systematic lean in an estimator. Mid-century statisticians joined the two words, and AI later used the pair for the lean a trained stand-in carries from its training set.
How it is spelled
Pattern
Latin modulus drops the u in English model
Pattern
-el is a two-letter ending from a Latin diminutive
Pattern
French biais keeps its final s in English bias
Breaks the pattern
the noun phrase stays two words; a hyphen appears only when the pair modifies another noun
Breaks the pattern
English model keeps one l, unlike Italian modello

Spelled like

  • module
  • modest
  • novel
Broken into chunks
  • mod
    • modest
    • modify
    • module
  • el
    • novel
    • channel
    • panel
  • bias
    • biased
    • unbiased
What it sits beside

Same subject

Same shape

Model compounds

Bias compounds

  • confirmation bias
  • selection bias
  • sampling bias

Where it sits in the deck

Phase 6: Model Quality, Ethics, and Responsible AI Principles

Before deploying a model, understand the quality and ethical dimensions — bias, fairness, explainability, and the human oversight needed to govern outputs responsibly.

Neighbours in this phase