A.I. MPACT ASSESSMENTS

This report collates resources and guidelines on AI impact assessments to support practitioners and policymakers seeking to design or implement safeguard against possible harms posed by certain AI technologies.

A.I. MPACT

An overview of resources on AI impact assessments

A.I. Impact gives an overview of resources and frameworks for AI impact assessments to guide policymakers and practitioners seeking to protect human rights in the deployment of artificial intelligence.

Download

Mapping the state of AI policy and regulation in Africa

The growing appreciation of the risks posed by unchecked uses of AI has fuelled interest among policymakers and practitioners in the use of AI impact assessments as a safeguard against potential harms.

Impact assessments evaluate any potential or actual harm caused by something, such as a project or plan, and aim to correct or mitigate the harm.

While critics have argued that impact assessments are a limited tool that do not adequately protect against human rights abuses in the domain of AI, the complexity of AI regulation suggests that effective oversight is only possible through multi-layered approaches that use a range of tools and mechanisms to offset each one’s limitations.

There is a growing body of research on how to effectively develop and implement an impact assessment in the context of AI. We collate a selection of these resources, sourced from global experts on the human rights implications of AI, that provide guidance and insight to practitioners looking to design and implement impact assessments of their own.

Explore the resources below, or access the full report here.

Guiding practitioners seeking to protect human rights

KEY FINDINGS

UN Guiding Principles on Business and Human Rights (2011)

Guiding Principles on Business and Human Rights

United Nations (2011)

Part II of the guiding principles provides for the corporate responsibility to respect human rights. It recommends that businesses conduct human rights due diligence processes to identify and mitigate the adverse human rights impacts of their activities. It provides that these processes should involve meaningful consultation with potentially affected groups. Assessments should be taken at regular intervals.
UN Special Rapporteur report on AI and human rights (2018)

Report on AI and the rights to freedom of opinion and expression, privacy and non-discrimination (Report A/73/348)

UN Special Rapporteur on the promotion and protection of the right to freedom of opinion and expression (2018)

This report addresses AI technologies and their implications for human rights, and highlights that HRIAs can serve as a tool to enable transparency throughout the lifecycle of AI. It recommends that impact assessments should be performed before procurement, development, or use and involve both self-assessment and external review. It also states that public sector procurement of AI technologies from private vendors should be accompanied by a public consultation on the design and implementation of the AI system. It recommends that States should pursue regulatory or co-regulatory schemes designed to require businesses to undertake impact assessments and audits of AI technologies.
AI Now framework on Algorithmic Impact Assessments (2018)

Algorithmic Impact Assessments: A Practical Framework for Public Agency Accountability

AI Now (2018)

This report provides a practical framework to assess automated decision systems used by public agencies and to assess the impacts of these systems. The AIA framework proposed by the authors is designed to support affected communities and stakeholders to assess the claims made about these systems and to determine where – or if – their use is acceptable.
Canadian Directive on Automated Decision-Making (2019)

Canadian Directive on Automated Decision-Making

Government of Canada, 2019

The directive applies to federal government bodies that use or intend to use AI-driven decision systems. The directive requires the Assistant Deputy Minister responsible for the program deploying the system to complete an AIA prior to implementation. These requirements differ depending on the level of risk presented by the system and include peer review, public notice of the use of the system, that humans must be in the loop for certain decisions, and, in some circumstances, that approval is required for the system to operate. It also requires that the final results of the impact assessment be made public.
Algorithmic Impact Assessments and Accountability (2021)

Algorithmic Impact Assessments and Accountability

Data & Society, 2021

This paper analyses the history and nature of impact assessments from other domains, the emerging and existing proposals for AIAs, and findings from past AIAs. Its recommendations include that designers and operators of algorithmic systems should be obliged to explain and justify the potential harms of algorithmic systems; and that the communities affected by the algorithmic systems are the ‘foremost experts’ and therefore should be included in assessment design and implementation.

The European Union Artificial Intelligence Act (“AI Act”)

European Union (2021)

The AI Act is a proposed European law on AI. It uses a risk-based approach in which AI is classified according to its relative threat to fundamental rights: low/minimal risk, high risk, and unacceptable risk. AI systems considered as unacceptable risks are prohibited and include systems that have “significant potential to manipulate persons through subliminal techniques” or to “exploit vulnerabilities of specific vulnerable groups” as well as social scoring systems and real-time biometric identification systems.
Human Rights Impact Assessment and AI (2022)

Human Rights Impact Assessment and AI

Alessandro Mantelero (2022)

This chapter proposes a model for HRIAs as part of the broader Human Rights, Ethical and Social Impact Assessment model. The authors argue that the model was developed in line with the existing practices in HRIAs, but in a way that better responds to the specific nature of AI applications in terms of scale, impacted rights and freedoms, prior assessment of production design, and assessment of risk levels, as required by several proposals on AI regulation.
Mandating Human Rights Impact Assessments in the AI Act (2021)

Mandating Human Rights Impact Assessments in the AI Act

Data & Society and the European Centre for Not-for-Profit Law (2021)

In a commentary on the European Union’s efforts to develop a legally binding framework on AI based on the EU’s standards on fundamental rights, the authors provide a set of recommendations for establishing an HRIA process for AI.

European Commission Ethics Guidelines for Trustworthy AI (2020)

Ethics Guidelines for Trustworthy Artificial Intelligence

European Commission 2020

A High-level Expert Group of the European Commission published the Ethics Guidelines in 2019 to enable organisations and companies to conduct self-assessments on AI systems being developed, deployed, procured, or used. Following that, in 2020, the group developed the Assessment list for Trustworthy AI which translates the ethics guidelines into an accessible and dynamic checklist that guides developers and deployers of AI in implementing such principles in practice.
Recommendations for Assessing AI Impacts to Human Rights, Democracy, and the Rule of Law

Recommendations for Assessing AI Impacts to Human Rights, Democracy, and the Rule of Law

Data & Society and the European Centre for Not-for-Profit Law (2021)

The authors assess the limitations of HRIAs for algorithmic accountability and advance a framework for evaluating impact assessment processes that will help to create more accountable and rights-respecting AI systems. The recommendations include that impact assessments should be legitimised either through legislation or within officially recognised sets of norms.
Guidance for the Development of AI Risk and Impact Assessments (2021)

Guidance for the Development of AI Risk and Impact Assessments

Louis Au Yeung, Center for Long-Term Cybersecurity, UC Berkeley (2021)

This paper provides a comparative analysis of AI risk and impact assessments (“RIAs”) from five regions around the world: Canada, New Zealand, Germany, the European Union, and San Francisco, California.
The European Union Artificial Intelligence Act (2021)

The European Union Artificial Intelligence Act (“AI Act”)

European Union (2021)

The AI Act is a proposed European law on AI. It uses a risk-based approach in which AI is classified according to its relative threat to fundamental rights: low/minimal risk, high risk, and unacceptable risk. AI systems considered as unacceptable risks are prohibited and include systems that have “significant potential to manipulate persons through subliminal techniques” or to “exploit vulnerabilities of specific vulnerable groups” as well as social scoring systems and real-time biometric identification systems.
Microsoft Responsible AI Standard, v2: General Requirements (2021)

Microsoft Responsible AI Standard, v2: General Requirements

Microsoft (2022)

The Standard is intended to serve as concrete and actionable guidance to product development teams on how to develop responsible AI in the absence of effective laws and norms. Goal A1 of the Standard is that all Microsoft AI systems undergo an impact assessment, which is subject to a review before development starts. The Impact Assessment must be updated and reviewed at least annually when new intended users are added, and before advancing to a new release stage. The Standard provides specific requirements for providing information about the capabilities and limitations of AI systems to stakeholders and notifying users that they are interacting with an AI system, and provides procedures to minimise the risk of bias or discrimination and steps to be taken to prepare for failures and remediations.