Cognitive Bias
Systematic patterns of deviation from norm or rationality in judgment, which can influence AI decision-making if present in training data.
Definition
Human biases (anchoring, confirmation, availability) can taint data labeling, feature selection, and objective setting. Recognizing cognitive biases requires structured data-governance reviews, blind-labeling protocols, and diverse labeling teams. Organizations must audit for bias sources not only in data distributions but also in human-in-the-loop processes.
Real-World Example
A survey-response classifier mislabels neutral sentiment as negative because labelers, primed by recent news events, over-interpret neutrality as pessimism (availability bias). The team implements blind labeling and rotates labelers to mitigate this cognitive bias in future annotations.
.svg)
