What does the bias risk index mean?

It is a weighted 0-to-100 score where 0 means the bias controls are well in place and 100 means they are largely absent. Each question is weighted by materiality, so missing a high-impact control such as subgroup evaluation raises the score more than a minor gap.

Which regulations does this align to?

The questions map to EU AI Act Annex III high-risk requirements — data governance, human oversight, transparency, and post-market monitoring — and to UK Equality Act 2010 obligations around direct and indirect discrimination. References are shown on each item.

What is a proxy variable and why does it matter?

A proxy is a feature that correlates with a protected characteristic even though it is not that characteristic — for example postcode standing in for race, or shopping history for gender. Models can discriminate indirectly through proxies, so they must be tested and justified or removed.

Is this a formal conformity assessment?

No. It is an indicative self-assessment to help you find and prioritise gaps. EU AI Act high-risk systems require a formal conformity assessment and, in many cases, a Fundamental Rights Impact Assessment carried out before deployment.

What is disaggregated evaluation?

It means measuring model performance separately for each demographic subgroup — accuracy, false-positive and false-negative rates, and selection rate — rather than only overall. Aggregate metrics can look fine while a subgroup is badly served, which is how bias hides.

Algorithmic Bias Self-Audit Tool

This tool helps you self-audit an AI or automated decision-making system for bias risk. Answer a set of governance questions and it produces a weighted risk index plus a prioritised list of gaps, each tied to an EU AI Act or UK Equality Act reference.

How it works

Each question maps to a recognised obligation and carries a weight reflecting how material it is to bias. Answers are scored:

Yes (fully in place) = no risk contribution
Partial = half the question’s weight
No = full weight

The total is normalised against the maximum possible weight to give a 0-to-100 risk index:

risk index = (sum of scored weights ÷ total weight) × 100

A score of 60 or more is flagged as high bias risk, 30 to 59 as medium, and below 30 as low. Every control that is not fully in place is added to the gap list with its regulatory reference and a concrete remediation step.

What the questions cover

Theme	Reference
Representative training data	EU AI Act Art.10 data governance
Proxy correlation testing	Equality Act 2010 indirect discrimination
Subgroup (disaggregated) evaluation	EU AI Act Annex III accuracy by group
Human oversight & override	EU AI Act Art.14
Explainability & transparency	EU AI Act Art.13
Post-market bias monitoring	EU AI Act Art.17
Complaint and redress route	Equality Act 2010 / EU AI Act Art.86
Fundamental rights impact assessment	EU AI Act Art.27

Notes

This is an indicative self-assessment, not legal advice or a conformity assessment. If your system falls within EU AI Act Annex III — recruitment, credit scoring, access to essential services, and similar — it is classed as high-risk and needs a formal conformity assessment and, often, a Fundamental Rights Impact Assessment before it goes live. Use the gap list to prioritise remediation ahead of independent review. The audit runs entirely in your browser.