What is Algorithmic Bias in AI Governance?
Algorithmic bias is systematic, unjustified difference in how an automated system treats people, often arising through facts that correlate with a protected characteristic, such as a postcode standing in for ethnicity.
Where it comes from
Historical decisions that were themselves biased, data that represents some groups poorly, and rules that use a proxy for a protected characteristic without anyone intending it.
What explicit rules make visible
When decisions are taken by sealed rules, every fact a rule reads is named in the artifact. A rule that reads a postcode, or a fact derived from one, can be seen and questioned before it is sealed, and its effect can be measured against outcomes.
What the EU AI Act asks
For high-risk systems, the Act sets requirements for the data used to train, validate and test them (Article 10), and deployers who control the input data must ensure it is relevant and sufficiently representative for the system's intended purpose (Article 26(4)).
Example
Approval rates differ sharply between postcodes.
The model's inputs are not named anywhere a reviewer can see, so the cause is guessed at.
The sealed rules show which facts are read; one rule reads a postcode-derived risk band. It is reviewed, and removed or justified, before the next version is sealed.
Frequently asked questions
Related concepts
See it in practice
Upload a policy, confirm its rules, seal them: any system can then check an action against them.