Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

AI Impact Assessment

nounAI Ethics, Safety and Governancealso algorithmic impact assessment, AIA, fundamental rights impact assessment

In one line

An AI impact assessment is a structured review carried out before and during the use of an artificial intelligence system to identify its likely effects on people, rights, safety and society, and to decide what safeguards are needed.

An AI impact assessment is a structured review of how an artificial intelligence system is likely to affect people, their rights, safety and society, carried out before launch and repeated as the system changes. Related names include algorithmic impact assessment and fundamental rights impact assessment. The practice borrows from environmental impact assessment and data protection impact assessment and is a key tool of responsible AI, AI risk management and AI governance.

Purpose

An assessment forces a team to ask hard questions early: Is AI needed here? Who could be helped or harmed? What happens if the system is wrong? What safeguards, alternatives and remedies are available? It produces a record that decision makers, auditors, regulators and the public can examine, which supports AI accountability and AI transparency.

Where assessments are required or encouraged

  • European Union. Under Article 27 of the AI Act, certain deployers of high-risk systems, including public bodies and some private providers of public services and banks and insurers using specific systems, must carry out a fundamental rights impact assessment before first use [1]. Data protection law separately requires data protection impact assessments for high-risk processing; see AI privacy.
  • Canada. The Directive on Automated Decision-Making has required federal institutions to complete an online Algorithmic Impact Assessment, which scores risk and sets requirements that scale with the score [2].
  • United States. Federal guidance and several state and city laws call for assessments or bias audits for certain automated decision tools, and the NIST framework encourages mapping impacts.
  • Voluntary. Many companies, standards bodies and public agencies publish their own templates.

Rules vary and change, so readers should consult current texts and legal advice.

Typical steps

### 1. Describe the system and purpose

State what the system does, who uses it, the data it relies on, the decision it supports and the legal basis. Name the owner.

### 2. Identify stakeholders

List everyone affected, including people who do not use the system but are subject to its outputs, and vulnerable groups. Plan how to hear from them.

### 3. Assess impacts and risks

Examine effects on rights and interests such as equality, privacy, health, education, work, access to services, freedom of expression and the environment. Consider both likelihood and severity, and look at direct and indirect effects; see algorithmic bias and data bias.

### 4. Evaluate necessity and alternatives

Ask whether the goal could be met by a simpler or less intrusive method, and whether the benefits justify the risks.

### 5. Plan mitigations

Select controls such as better data, testing, human review, user notices, appeals, limits on use or a decision not to deploy; see human oversight.

### 6. Decide and document

Record who approved the system, conditions attached and residual risks accepted.

### 7. Monitor and revisit

Set indicators, schedule reviews and update the assessment when the model, data, population or purpose changes; see AI observability.

Contents of a good report

A strong assessment is specific, evidence-based and honest about uncertainty. It usually includes the system description, a stakeholder map, a list of identified impacts with ratings, test results by group, the mitigation plan, the monitoring plan, the contact route for complaints and the approval record. Some jurisdictions or organizations publish summaries so the public can read them.

A worked example

A city plans to use a model that ranks households for a home energy subsidy. The team describes the system, which uses past bills and building data, and lists affected groups: renters, older adults, people without internet access and households in older buildings. Testing shows the model ranks renters lower because landlords control some data. The assessment finds that this could widen existing inequality, so the team changes the data sources, adds a paper application route and requires a caseworker to review every denial. Residents are told that automated tools are used and how to appeal. A review after six months compares approvals by neighborhood and housing type, and the report is posted online.

Strengths and limits

  • Strengths. Early discovery of problems, shared understanding, a documented rationale and a basis for audit.
  • Checklist risk. Forms can become box-ticking exercises with no effect on decisions.
  • Who participates. Without real input from affected communities, important harms are missed.
  • Independence. Self-assessment may undervalue risks, so independent review adds credibility.
  • Measurement. Social harms are hard to quantify.
  • Timing. An assessment done after development is complete rarely changes design.

Good practice

Start during design, involve diverse experts and affected people, tie the assessment to decision gates so that a poor result can stop or change the project, share results where possible and keep the document alive after launch. Link it with data protection, security and equality reviews to avoid duplication.

Impact assessments for generative AI

Generative systems complicate assessment because their uses are open ended and may change after release. A provider can assess general risks, such as fabricated content, privacy leakage, harmful instructions and bias in outputs, while each organization that deploys the model assesses its own use case, for example a customer chatbot, a tool that drafts legal letters or a school writing assistant. Assessments of these systems often include scenario testing, review of the provider’s documentation, rules about which data may be entered and plans for handling user reports. Because behavior can shift when a supplier updates a model, teams should repeat key tests after each major change; see AI evaluation and red teaming.

Public participation and publication

Many scholars argue that impact assessments are most meaningful when affected people take part in them and when results are shared. Participation can take the form of community meetings, advisory panels, surveys, interviews with workers and advocacy groups, and open comment periods for public-sector systems. Publishing a summary, along with what changed in response to feedback, lets others check whether concerns were addressed. Where full publication is not possible because of security or trade secrets, regulators or independent auditors can be given access under confidentiality, so that the assessment remains a source of real accountability and not only an internal paper.

FAQs

What is an AI impact assessment? It is a structured review of how an AI system may affect people, rights and society, done before use and repeated over time.

Is an AI impact assessment required by law? In some cases, such as certain high-risk uses under the EU AI Act and federal automated decisions in Canada.

How is it different from a risk assessment? A risk assessment can focus on organizational risk. An impact assessment centers on effects on people and rights and on stakeholder input.

Who should carry it out? The accountable owner with legal, technical, domain and affected-community input, ideally with independent review.

When should it be updated? Whenever the system, data, purpose or context changes significantly, and on a regular schedule.

More in AI Ethics, Safety and Governance

Key facts

  • Article 27 of the European Union's AI Act requires certain deployers of high-risk AI systems to carry out a fundamental rights impact assessment before first use1.
  • Canada's Directive on Automated Decision-Making, in force since 2019, requires federal institutions to complete an Algorithmic Impact Assessment questionnaire2.
  • The 2021 report Algorithmic Impact Assessments by Moss and colleagues at Data & Society examined how such assessments can create accountability3.

Related concepts

Quick checkWhat is an AI impact assessment?Show answer

An AI impact assessment is a structured review carried out before and during the use of an artificial intelligence system to identify its likely effects on people, rights, safety and society, and to decide what safeguards are needed.

Sources

  1. European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 27, https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  2. Government of Canada, Algorithmic Impact Assessment tool, https://www.canada.ca/en/government/system/digital-government/digital-government-innovations/responsible-use-ai/algorithmic-impact-assessment.html
  3. Emanuel Moss et al., Assembling Accountability: Algorithmic Impact Assessment for the Public Interest, Data & Society, https://datasociety.net/library/assembling-accountability-algorithmic-impact-assessment-for-the-public-interest/

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue