Automated Decision-Making Systems in Municipal Governments

Overview

This past year, I was part of the inaugural Techtopia cohort at the Berkman Klein Center (note: this program later merged with the “Assembly” program), organized by Jonathan Zittrain within the Ethics & Governance of AI Initiative. The program involved participating in a series of resesearch seminars alongside an interesting cohort of other students from across the university, as well as the completion of a group project in the spring semester.

For the project phase, I helped develop the “Technology Procurement Playbook Project” alongside a great team¹. Our team’s deliverable was a playbook intended for municipal governments to help guide the procurement and development of automated decision-making systems (i.e. algorithms & machine learning products). The aim was to provide broad guidance for local government nationwide by examining the principles and processes required to ensure that municipal technology empowers, rather than harms citizens, and respects various moral/ethical considerations like equity, fairness & accountability. We developed the playbook through conversations with stakeholders in Boston city governemnt as well as academic experts and independent research.

While the project so far has not made it out of the “draft” phase, I wanted to reflect a bit on what I learned from engaging with it². I will highlight two main buckets of learnings.

Tech Ethics in Municipal Government

First, I learned three main things about the context of deploying automated decision-making systems in local governments, and addressing tech ethics & governace questions in this setting:

1. There is a clear need for this type of technical (and tech ethics) guidance in local governments,
2. Academia & civil society can play a meaningful role in addressing this need, and
3. Procurement processes are a very reasonable site of intervention for technology ethics & governance questions in this context.

The learnings around (1) and (2) mainly arose out of our conversations with local government stakeholders, who are facing real questions / needs around deploying automated decision-making systems, and have both very limited resources (in terms of time, technical background etc.), and a strong interest in deploying these systems responsibly.

On point (3), procurement is a relevant point of intervention for two reasons. First, local governments are often working with third-party contractors to develop or procure these systems rather than building them in-house (due to technical limitations). Second, the procurement process puts government stakeholders face-to-face with contractors and is a high-leverage moment (before a purchase commitment is made) to ask questions or extract concenssions from contractors around moral / ethical / governance issues that are relatively more important to government stakeholders in the negotiation³.

Centering Algorithmic Accountability

Second, I learned that making general recommendations around the responsible deployment of automated decision-making systems – that will apply across various complex sociotechnical contexts – is hard. This is true for all of the same reasons that social science and policy are hard: social contexts are complex and particular, and the specific issues that might arise in deploying a new technical system–and the best ways to address them–are likewise varied and difficult to predict ex ante.

So what can we do? In my understanding of our approach in the Playbook, we address this issue first by outlining a broad range of potential algorithmic harms for stakeholders to consider up front, and second by detailing generalizable processes for creating algorithmic accountability – that is, processes for (i) detecting, (ii) creating responsibility for, and (iii) redressing/mitigating any harms that might arise in the deployment of automated decision-making sytems.

The basic idea in focusing on algorithmic accountability is that, while it might be difficult to predict up front everything might go wrong in a particular context (especially with limited technical resources), governments can at least take steps to ensure that there are robust systems for detecting and addressing these issues (possibly with help from academia & civil society) before they become bigger problems. Hence our recommendations around accountability included taking steps toward public disclosure of algorithmic processes⁴ & ensuring wherever possible that they are open and auditable by independent third parties⁵. Further, we emphasized the importance of iterative development in collaboration with affected stakeholders & subject-matter experts⁶ wherever possible. An iterative approach is helpful here (if possible) since it allows for unintended consequences to be surfaced (and addressed!) via feedback from affected stakeholders. In practice, this may mean moving away from rigidly-specified RFP (“request for proposal”) models of procurement towards more collaborative models of engagement between governments, developers & stakeholders.

A final piece of the puzzle that we discussed was the importance of creating a public “chain of responsibility” for algorithmic systems. This seemed important to highlight since it can be otherwise difficult to assign responsibility in complex socio-technical systems. When something goes wrong with a predictive model in the field, who is to blame? Is it the user of the model? The model’s original developer? The team that procured the model? The model itself? Without clarity here, it’s sure to be difficult to create accountability for issues. There is a risk of what a related Data & Society Report calls “moral crumple zones”, where “one entity is held legally liable for errors, even if they aren’t in full control of the system” (21).

Footnotes