Operationalizing algorithmic fairness with decision support tools at the NYC Office of Technology & Innovation
Written by Isabel Corpus
The Challenge
In December 2025, the New York City Council passed a set of local laws intended to ensure the transparency, accountability, and fairness of algorithmic systems operated by New York City agencies. In particular, these laws, which came to be known as the GUARD Act, assign New York City’s central technology agency,the Office of Technology and Innovation (OTI), responsibility to provide oversight over public-impacting AI tools that are used by mayoral agencies. To enact meaningful oversight in a city with 50+ mayoral agencies and 86 algorithmic tools reported in 2025, there will need to be rules promulgated under New York City Charter to govern how agencies procure, use, and monitor AI.
While the language of the GUARD Act enshrines principles like fairness, the laws leave the specifics of operationalizing these principles to future policy. The next challenge is to define the processes required to satisfy them.
My Project
As a Rubinstein PiTech PhD Impact fellow with OTI, I conducted research about how principles of algorithmic fairness can be operationalized. In particular, I studied and synthesized AI policy that has been deployed by other jurisdictions (including cities, states, and countries) and academic literature that defines best practices for algorithmic transparency, accountability, and fairness.
There are many decision points along the road to operationalizing a concept like algorithmic fairness. Academic literatrue often glosses over these decision points, assuming that the parties conducting bias assessments have sufficient time, access, compute, and technical capacity to conduct a broad suite of statistical tests. Meanwhile, in existing public policy, bias testing often sits within the anti-discrimination requirements of strictly regulated domains like credit, housing, and employment, rather than in a single centralized agency that is responsible for tracking the development of algorithmic bias across sectors and use cases. Taken together, the limitations of existing academic research and the contextual differences between the GUARD Act and existing policies mean that operationalizing the Act will require novel work from both academic researchers and policymakers.
I translated these broader questions into a concrete operational problem: determining the levels of access required to conduct any fairness testing on an algorithmic system. This is important groundwork, because an oversight office might not directly own the algorithmic tools that it assesses. For example, according to Local Law 35 reporting, most of the algorithmic tools that are eligible for assessment are owned by government agencies outside of OTI.
There are two distinct access challenges. First, communication between government entities is neither automatic nor guaranteed. It may require mediation through intergovernmental affairs teams or City Hall. Second, access to the algorithmic systems and data needed for an assessment is not automatic. Facilitating sufficient access may require legal agreements, like a memorandum of understanding or data usage agreement. To determine the level of agreements that are necessary to facilitate bias testing and to define the technical methodologies that can be used to conduct those tests, we first need clarity about exactly what level of access is required.
To address this, I created a taxonomy defining the types of information and system access necessary for different levels of fairness testing, and mapped tiers of system access to the rigor of testing that would be feasible at each tier. In other words, more rigorous testing requires greater access. This taxonomy can then inform formal protocols for data and system access, thus helping an oversight office establish the conditions necessary to conduct fairness testing.
Isabel Corpus
Ph.D. Student, Information Science, Cornell University
Impact and Path Forward
Translating algorithmic fairness research into actionable, enforceable rules is a significant task.
I learned a lot about the limits of current academic research, which may not account for the realities of new oversight agencies and the legal processes required to enable even basic bias tests. My experience was uniquely valuable, as a rare hands-on experience translating law into tangible fairness assessment practices. I would strongly recommend the PiTech fellowship to future students to understand how their academic contributions are (or are not) grounded in the reality of civic processes, and thus better tailor their future work toward real-world impact.
My experience with OTI will also shape my future academic work, particularly by ensuring that recommendations for public policymakers are actionable. One future research direction could be to analyze public-facing policy documents to study how the language of legal requirements for fairness and bias testing align with real-world constraints on conducting such tests.