Glossary term
AI Impact Assessment
Subject: Technology, Cyber and Data, AI Governance and Risk
Also known as: Algorithmic impact assessment
An AI impact assessment is a structured review, done before and during deployment, of how an AI system could affect people, groups and society, and what steps will reduce the harm.
Explanation
The assessment looks outward at the people on the receiving end, not at the organisation's own balance sheet. The team describes the system's purpose and data, lists who is affected, and asks what could go wrong for them: unfair refusals, loss of privacy, wrong decisions that are hard to challenge, or effects on vulnerable groups. Each risk is rated and matched with a safeguard such as human review, narrower use or better testing, and the result is a written record with a named owner. Some laws require one for risky uses; the EU AI Act, for example, asks certain deployers for a fundamental rights impact assessment, and Canada has used algorithmic impact assessments for public-sector systems. Many firms borrow the method of a data protection impact assessment and extend it beyond privacy. Reviewers look for a written record showing that a harm was identified and then changed something in the design.
Example
A city council wants to use a scoring tool to prioritise housing inspections. Before buying it, staff complete an AI impact assessment. They find the training data comes mostly from older neighbourhoods and could send inspectors disproportionately to low-income tenants. The council asks the supplier to retrain the model, adds a human check on each flagged case, and plans a review after six months.
Not to be confused with
- Data Protection Impact Assessment (DPIA)
- A DPIA focuses on privacy risks from personal data use; an AI impact assessment also covers fairness, safety and wider social effects.