AI bias
Systematic distortions in AI systems that can disadvantage particular cases or groups of people.
Where can one-sided results come from?
AI bias refers to systematic distortions in data, models or their use. These can lead to worse treatment of particular cases or groups of people. Causes include data selection and labelling, the learning objective or existing business practices. Such distortion does not require deliberate disadvantage. A good average score can hide frequent errors in part of the intended use.
A fictional model for routing service requests is trained mainly on messages from the central workshop. These use official product names; local abbreviations from other branches are barely represented. Requests from those branches could therefore be assigned incorrectly more often. To investigate, examples from the affected branches are labelled under the same business rule, and misassignments are examined separately alongside case counts. This helps assess whether data selection, labelling or model behaviour needs improvement. More training messages help only if they actually cover the relevant gap.
Sources
Related terms
- Machine learningModels learn patterns from data and apply them to new inputs.
- Supervised learningA learning method that uses inputs together with their corresponding target values or classes as training examples.
- Training dataThe examples from which a model learns relationships or parameters during training.
- Test data in machine learningData reserved for final assessment whose results were not used to select the model variant being assessed.
- Explainable AIApproaches that make an AI system’s results or operation understandable to the intended recipients.
- Confusion matrixA table comparing reference classes with model predictions and counting correct and incorrect assignments.
Your project with ASES
Describe your project to us. We will get back to you with an initial assessment.