This piece is part of OGP’s “Open Algorithms Blog Series.” Read other posts in the series here.

Over the last few years, the way in which governments function has undergone a significant transformation, with more services being digitized for efficient and effective delivery. Government agencies are increasingly using sophisticated algorithms and data for public decision-making on matters related to health, educationAccountability within the public education system is key to improving outcomes and attainment, and accountability is nearly impossible without transparent policies and opportunities for participation ..., migration, and other key policy areas. Making the public sector use of algorithms more transparent and accountable is crucial to ensure that digital technologies are not misused and that the rights of users and citizens, including the use of personal data in an ethical way, are protected. 

Heightened public interest about the use of algorithms for public decision-making, which can emerge in the wake of scandals such as the SyRI algorithms in the Netherlands and exam grading algorithms in the United Kingdom, provide cause for concern but also are a valuable opportunity to raise awareness within and outside the government about the importance of increasing transparencyAccording to OGP’s Articles of Governance, transparency occurs when “government-held information (including on activities and decisions) is open, comprehensive, timely, freely available to the pub... More and public accountabilityAccording to OGP’s Articles of Governance, public accountability occurs when ”rules, regulations, and mechanisms in place call upon government actors to justify their actions, act upon criticisms ... More of algorithms. 

Several OGP members including France, Netherland and New Zealand are now using their OGP action plans to advance transparency, participation and accountability of government algorithms. In November of 2020, OGP hosted an online exchange with government officials from implementing agencies from Canada, France, New Zealand, and the United Kingdom. This network met for the first time in May 2020 and will continue to gather on a quarterly basis, tackling a variety of policy questions related to algorithmic accountability. They will be joined by different civil society and expert partners, based on the topic of discussion.  

This most recent discussion focused on sharing country experiences on cross-governmental coordination and building capabilities in relevant government agencies to support the implementation of algorithmic accountability. The group was joined by Amba Kak, Director of Global Policy & Programs at New York University’s AI Now Institute, who shared examples of gaps and challenges that they had consulted with governments on. Here are some of the design and implementation questions highlighted in the discussion: 

1. Scope & definition of algorithms: Reaching a common understanding of the notion and scope of an algorithmic policy is among the greatest policy design challenges. There is not a “one-size-fits-all” policy. The content and scope depend on what service the government is delivering to the public and what problems are being tackled — making it hard to take a broader approach and to create comprehensive guidance for relevant government agencies. Some of the main questions that need to be tackled when designing algorithmic accountability are:

• Technology vs impact: Should the focus be on the technology used or the impact (or rather the decision)? The group noted the effectiveness of focusing on the impact to determine the rules, rather than the technology itself. In terms of what technologies should be involved in the scope, one of the key questions raised was the definitional challenges with artificial intelligence (AI). What has been observed is that several policy instruments use “automated decision-makingAs governments are increasingly turning to algorithms to support decision-making for public services, transparency and accountability of algorithm use is necessary to minimize their harm and increase ...” as the umbrella term rather than AI. This shifts focus to the function, use and impact of these systems as a whole rather than narrower definitions. Other key questions include:  What should and should not be considered exemptions? How can principles of data ethics be taken into consideration?
• Sectoral scope: Should we take a broad or narrow scope? Which sectors should be covered? Practitioners have identified both a sectoral and regulatory overlap in the implementation of design principles. 
• Design and implementation: Is the risk-based approach a good fit to guide agencies on what to focus on? How can we bring algorithmic accountability to the design of algorithms, and not only in the implementation? Most officials in the group find it equally important to bring these principles into both stages across government departments. One of the key challenges identified is determining how to harmonize new reforms with existing legal frameworks. Some countries find it useful to link reforms to their pre-existing data ethics and data protection frameworks when they have been co-created with civil society and experts. 

2. Mandated by law, policy or a voluntary approach: Enforcement of algorithmic accountability in the different government agencies that use algorithms in their work is also a key aspect to consider. Different models are in place:

• Mandated by law: France has a legal framework mandating open algorithms, so the policy scope is set in the law. Administrations should publish online a list of algorithms when they are used to make decisions that impact citizens’ life. 
• Mandated by policy: Canada has a Directive issued by the financial and operations oversight committee of the federal government, which sets requirements for how algorithms can be used to support service delivery to citizens. Compliance with this directive is required for departments in the Canadian federal government.
• Voluntary: In New Zealand, the approach to algorithms is based on a public commitment from government agencies, not a legal standard. The country created an algorithms charter (including as part of an OGP commitment) that was subject to public consultation and is currently working on its implementation. As in any voluntary approach, one of the challenges is that not all agencies have committed to implement the charter. It is important to consider that many of these changes are happening before legal frameworks are established. This is a constraint but also an opportunity that has allowed for quicker and more nimble movement. Testing approaches for algorithmic accountability within the public sector is seen to be possible.

3. Responsibility of dedicated agencies

One of the challenges the group highlighted was implementing accountability of algorithmic decision-making in the contexts where the delivery of services happened through a chain of different departments, each with different mandates and responsible for different elements across the delivery chain. In such cases it is possible for different agencies to be responsible for collecting data and others for using the data. While some use the data for internal government processes, others are service-delivery oriented. Therefore, one of the key policy questions that arises is at what stages accountability tools and measures should be put in place.

Culture change between government departments is another challenge that was identified: Initially it was thought that implementation of digital technologies needed IT departments alone. Now we have learnt that there is a need to take a multidisciplinary approach that includes legal, privacy, impact, user-centric design and delivery aspects.

Other challenges are that these agencies have competing policy priorities and deal with limited resources, which increases the need to find a balance between what is useful and what can be humanly demanded from them. There is also a perspective that dedicating resources to transparency and accountability may come at the cost of innovation.

4. Channels of engagement

Some of the participating governmentsOGP participating countries and locals – often referred to as OGP members – are committed to working with civil society to strengthen their adherence to open government values. To join OGP as a pa... have been proactively reaching out to relevant agencies or setting up working groups to coordinate the intra-governmental work. For example:

• Etalab in France has established a working group with relevant government agencies, including from local governments. 
• In New Zealand, Stats NZ and the Government Chief Data Steward connected with relevant agencies to promote the application of the Algorithm Charter and are working with the largest agencies in the country to coordinate its implementation. A total of 26 government organisations have signed up to the Algorithm Charter. 
• When drafting its Directive on Automated Decision-Making, Canada held consultations in different cities around the country with government representatives, academics, legal experts and civil society. 

The relevance of communicating the importance of algorithms and their use to both the public and fellow government officials is also essential. The government of France, for example, created a short video to explain an algorithm.    

Next steps

This group will continue to gather in 2021 and engage a broader group of governments, civil society and international partners from other regions to drive this discussion forward. We’ll also explore questions around transparency and accountability of government procurement of algorithms and AI systems, algorithmic bias, systems of accountability, international standards and concrete use cases.

Featured Image Credit: heylagostechie via Unsplash