Skip to main content

At the Boundary of Law and Software

Toward Regulatory Design with Agent-Based Modeling

Sebastian Benthall1, Michael Carl Tschantz, Erez Hatna, Joshua M. Epstein, and Katherine J. Strandburg

The use of agent-based models (ABMs) can improve software accountability by representing populations of actors (e.g., house buyers, job seekers, college and insurance applicants) affected by software. Unaccountable software can embed and exacerbate societal biases or have other pernicious effects. ABMs can be used to explore the societal effects of software systems to aid the design and enforcement of regulations. Currently, software accountability measures are largely defined very narrowly, as mechanical compliance with regulations, written without a systematic way of predicting societal impact. ABMs are a way to model the societal impact of software and to correct or design regulations accordingly. We argue, in essence, that agent-based models (ABMs) of the interactions between software systems, those systems’ social environments, and applicable regulations can help to improve software accountability. We focus on software systems that employ personal data, have significant impacts on individuals, and are subject to regulation or used within regulatory agencies.

For example, anti-discrimination regulations in the housing context focus on societal goals such as fairness in housing or the reduction of residential segregation. Some of these regulations are applicable to the software systems that target online housing ads. These systems, typically, are written not by regulators but by private sector software designers. The resulting deployed advertising systems may produce effects that are inconsistent with – even subverting – the regulatory intent. One possible reason for such an outcome is that the social environment in which the software is deployed is absent from the analysis of software impact. We want to include it and make it accessible to the regulators. If the social environment were represented, auditors could better anticipate whether the software is likely have the intended social effect. We propose to use ABMs to fill this gap, making software more accountable in this sense.

This sort of transformation has occurred in other fields. Infectious disease modeling is one. Faced with a novel pathogen, like Swine Flu or SARS-CoV-2, it is crucial to have some way to estimate how fast it may spread, how much vaccine to produce, whether to ban international travel or close schools and workplaces. Before the advent of infectious disease transmission models, doctors and public health official were operating in the dark. Now, we have disease simulation models at scales from the local to planetary for forecasting and mitigation. They are part of the fabric of public health decision making, and are used to inform policy at the CDC, NIH, WHO, and many national governments. They are not crystal balls and do not always agree, as in weather forecasting. But collectively, they can bound our uncertainties, estimate sensitivities, explore tradeoffs, and offer headlights in uncertain settings. Agent-Based Models specifically, which can include social networks and cognitive factors, are increasingly used. ABMs have the enormous advantage that they are also visual and rule-based (not equation-based) allowing non-technical audiences and domain experts to understand their results and, indeed, to participate in their construction. At the same time, they can be calibrated to epidemic data. The net effect of all this is that ABMs can have higher impact than complex differential equation models. We are proposing to do this for software accountability.

We see potential in this method because ABMs are legible to software engineers, social scientists, and regulators. We have also identified key challenges to this approach, which are the potential politicization of model choice, the selection of appropriate robustness metrics, and the design of the interface between the ABM and audited software.