Model bias arises when a system's outputs are systematically skewed, typically because training data over-represents or under-represents certain patterns. It can appear as disparate treatment of demographic groups or as uneven performance across document types, formats, and terminology.
In legal applications, the practically significant form is often performance variation rather than demographic bias, such as weaker extraction from handwritten records or unfamiliar provider formats.
Alternative Names:
AI Bias, Algorithmic Bias
Why it Matters?
Bias in legal AI usually surfaces as uneven reliability rather than obvious unfairness, which makes it easy to miss. A tool that performs well on typed hospital records and poorly on handwritten nursing notes creates silent gaps in a chronology. Evaluating performance across the actual mix of documents a practice handles, not on vendor benchmarks, is what surfaces the problem.
Frequently Confused with
Related terms
Frequently asked questions
How does bias show up in legal AI tools?
How is it detected?
