Algorithmic Impact Assessment in Mitigating AI Harms

IBM’s 2022 Global AI Adoption Index Report notes a four-point increase (35%) in the adoption of AI and its impact on businesses and Society in comparison to the previous year(2021). The Global AI Adoption Index report commissioned by IBM gives an overview of AI deployment across businesses in the world. However, the report notes that the increase in adoption is not complemented by steps to ensure trustworthy and responsible AI adoption, particularly in reducing unintended bias, a significant consideration in mitigating such harms lies in the development and implementation of AI governance structures.¹

Various approaches have been instituted in governing AI to ensure trustworthy, reliable, and responsible AI. These include but are not limited to the development of AI ethical frameworks, laws governing the impact of AI and robotics, and technical approaches like Algorithmic Impact Assessments. (AIA).² A core pillar of developing and deploying AI accountability. Algorithms form the core functioning element in the development and functioning of AI, some of Its benefits include; personalizing experiences, enhancing productivity, and improving decision-making. These benefits may however have negative consequences particularly where systems using algorithms undermine fairness by reproducing societal biases.

Stakeholders in the AI ecosystem have proposed AIAs as a regulatory strategy for addressing algorithmic harms. It requires the assessment of socially harmful impacts prior to implementation and the creation of documentation for accountability and policy development.³ The potential for harmful effects increases with the continued automation of systems and the complexity of algorithms.⁴ One of the technical aspects of mitigating biases in algorithms is algorithmic accountability. Algorithmic accountability is a process of holding developers and entities responsible or accountable in cases where the algorithm they develop or operate makes decisions that result in unfair outcomes.⁵ Several algorithmic accountability mechanisms are being used in both the public and private sectors to hold people and institutions accountable to those affected by the negative impacts of AI.⁶ AIA is one of the emerging mechanisms being utilized as a means of building algorithmic accountability.⁷ The use of AIA as an accountability mechanism in gaining popularity and being adopted into governance structures by government, public bodies, and AI developers. The aim is to create accountability for potential benefits and mitigate harms from AI systems.⁸

“Algorithmic accountability concerns a networked account for a socio-technical algorithmic system, following the various stages of the system’s life cycle. In this accountability relationship, multiple actors (e.g. decision makers, developers, users) have the obligation to explain and justify their use, design, and/or decisions concerning the system and the subsequent effects of that conduct.”⁹ Contents of an AIA would include an assessment of,¹⁰

• • Data: Sensitivity, appropriateness, and timeliness of the data used in the AI system

  • Nature of the impact: this covers the impact of affected persons. The impact could be in terms of physical, mental, economic, and ecological harms as well as fundamental and legal rights.

  • The scale of impact: severity of the impact and the number of people affected.

  • The permanence of harmful impacts: Are the impacts long-lasting and what reversibility of the effects

  • Likelihood of the impact occurring

  • Role of the system in making decisions

  • Transparency of the system

This list is not exhaustive as AIAs may be structured with different components in the course of assessment to fit the needs of the industry or sector conducting the assessment. Governments actively involved in developing and implementing AIAs include the Canadian government, having established an online AIA tool, a 2019 framework proposed by Germany’s Data Ethics Commission, a legal framework proposed by the European Commission in April 2021, New Zealand with the “Algorithm charter for Aotearoa New Zealand” for use by government agencies, set up by the New Zealand government, the “Ethics and Algorithms Toolkit,” which developed by the City and County of San Francisco and its partner organizations, GovEx, Harvard DataSmart, and Data Community DC; additionally we the proposed US Accountability Algorithmic Accountability Act. ¹¹

As AIAs become more entrenched in regulation, its effectiveness will be characterized when big tech companies implement the practice of assessment and publish the assessments. This way, redress mechanisms would be easily established as there would already be in place a means of establishing accountability. AIAs, unlike most impact assessments, must be published publicly as they are designed to engage end users and providers on the areas of concern through various comments and reviews allowing individuals, communities, and policymakers to participate in establishing and upholding accountability.¹² Development, adoption, implementation, and use of AI continue to grow globally. In turn AI regulation is also on the rise owing to the numerous harms AI technology has already had socially and ecologically. The effectiveness of AI regulation relies on the anticipation of the ways in which it can be filtered through private and public sector environments and the government owing to the contributions both sectors make in the AI ecosystem; this also applies to AIA regulation. It will therefore be crucial that regulators establish a technical understanding of the industry, the organizational culture as well as emerging documentation standards to ensure that the law keeps up with ever-evolving technology.¹³

From an African perspective, AIAs are yet common within the continent, however, the continent is witnessing a rise in the number of AI national strategies with countries such as Mauritius and Egypt establishing extensive strategies. Regional bodies have also shown their commitment to establishing AI regulation in the continent through the AU -AI Continental Strategy. AIAs may not be far off reality and established regulatory practice as the continent has made advancements in its utilization of AI technologies, further, big tech companies are setting up regional technological hubs in the continent. While counties in the continent continue to develop their strategies, AIAs would be beneficial to consider and add as implementation procedures while drawing up how and in what ways AI can be leveraged in the different countries and in particular, establishing the impact of AI socially, economically, and culturally. With this accountability is established as a non-negotiable principle and standard in the development, adoption, implementation, and use of AI thus establishing responsible AI. Whereas accountability is not the only principle/standard in establishing responsible AI it anchors the principles of fairness, transparency, privacy, and bias elimination alongside the existing ethical AI principles.

Image Source: CAIDP Linkedin