Skip to main content

AI for smarter legislation

Artificial intelligence can help improve some of the core functions of our democracies. Here are two ways the technology can help improve both the accuracy of new policies and the quality of debate about those policies in legislatures.

This article was derived from written testimony that Joe Mariani of Deloitte’s Center for Government Insights delivered to the United States House of Representatives Select Committee on the Modernisation of Congress on June 28, 2022.

“Government of the people, by the people, for the people.1”—plus an AI bot or two.

Legislation is an inherently human endeavour. But just as organisations across industries are unlocking new capabilities and efficiencies through artificial intelligence (AI), governments also can aid their legislative processes through the application of AI.

For the past five years, we’ve studied the potential impact of AI on government. We’ve looked at everything from how much time AI could save workers in each US federal agency to the rate of AI adoption in US federal, state and local governments.2 While AI can help many different areas of the legislative process—from AI assistants answering members’ questions about legislation to natural language processing analysing the US Code for contradictions—two key applications stand out.

AI as a microscope: Assess the impact of existing legislation

The same broad scope and volume of data that make assessing legislation a difficult problem for humans make it an ideal challenge for AI.

Machine learning (ML) models can find patterns in inputs and outputs without having to specify ahead of time how those inputs and outputs are likely to be linked. Just as a microscope can examine a leaf to find structures and patterns invisible to the human eye, ML models can find patterns in the outcomes of programmes that may be invisible to humans.

There are already examples of ML models examining public policy in exactly this way. For several years, researchers have been using ML to understand the risk factors for infant mortality in childbirth. With the data available in electronic health records, many of these models can predict the likelihood of complications with 95% or greater accuracy.3 Researchers from RAND then took those models to the next step. They used ML on data from Alleghany County, Pennsylvania, to evaluate which interventions had the biggest impact on reducing infant mortality.4

The strength of ML models is that they’re able to say not only what outcomes each intervention is likely to produce, but also among which groups those outcomes are likely to occur. These findings can then guide policy recommendations. For example, researchers found that mothers who used social services were less likely to use prenatal medical services and vice versa, pointing to a need for policies aimed at building awareness of other services aimed at improving child health.

While anything that improves the lives of infants is clearly a good policy outcome, other issues could be less clear-cut. ML models can uncover hidden outcomes of policies or programmes, but only humans can decide if those outcomes would qualify as successes or failures. In the spirit of human-machine teaming, once the ML model has uncovered the hidden outcomes of a programme or piece of legislation, members or staff can then look at those outcomes and determine: 1) if they are positive or negative and, 2) if the overall benefits are worth the cost and effort.

AI as a simulator: Test the potential impacts of future legislation

The ability of ML models to predict outcomes of policies begs the questions: What if we did something differently? How would things change? In essence, can AI be used as a “simulator” for problems? Think of Apollo 13. After an explosion, the crew had to figure out new interventions, new ways of doing things. They used the ground-based simulator to try procedure after procedure until they found one that worked. Imagine having an AI simulation run through hundreds of thousands of possible interventions in minutes, instead of locking astronaut Ken Mattingly in a dark box for days.

In place of ML models trained on historical data and projecting trends into the future, these simulations are designed to capture the dynamics of complex systems like the economy or the health care system. Simulations are based on models for how a portion of a complex system operates. For example, one form of simulation well-suited to legislative tasks are agent-based models that replicate how individual actors would respond to and interact in different situations. These models are good at capturing the “emergent properties” of complex systems where individual decisions add up in unusual ways. Think about flocking birds: Each bird just tries to stay close to the bird next to it, but together, they make intricate patterns in the sky as they avoid obstacles and predators. The big human systems that legislatures are often interested in like health care, the economy, or national defence exhibit similar traits.5

Researchers in Europe have created an agent-based model designed to help policymakers understand the likely impacts of different interventions on the Irish economy. The Innovation Policy Simulation for the Smart Economy uses data from patents, knowledge flows and other economic data to model how individual companies and investors are likely to react to different policies.6 For example, researchers can examine if different funding methods or tax incentives would help support the creation of new small businesses in a specific city or high-tech industry. Such models could be of great benefit as a government examines which policies could help spur domestic semiconductor manufacturing or other advanced technologies.

In 1964, economist George Stigler said, “We do not know the relationship between the public policies we adopt and the effects these policies were designed to achieve.”7 ML models can help uncover just this. But that relationship is only helpful in making future policies if we assume the future is like the past. So when we find ourselves in an era we know is different from the past (during a global pandemic, for example) or when we see a model based on historical assumptions drifting away from current data (such as with the widespread adoption of new technology, for example), then simulations become critical tools for understanding the likely outcomes of new public policies.

As the Apollo 13 simulator helped astronauts, AI simulations can help policymakers to:

  • Uncover the drivers of a particular problem, whether that is the amperage limitations on the lunar module in Apollo 13 or the causes of regulatory noncompliance.8
  • Understand which interventions could be effective, whether sequencing systems start up for Apollo 13 or organising national airspace to allow for more on-time flights.9
  • Understand the trade space of a given issue. Of all the effective interventions, how much is required, at what cost and to achieve what outcome?10

Yet even these complex AI simulations can’t make value judgements. They can’t determine the best option. They can only assess the optimal choice for the given values and assumptions that humans specified at the start. However, by forcing the human side of the human-machine team to be specific about those values and assumptions, AI simulations may hold the potential to transform legislative processes.

For example, modelling the impact of different policies on economic development can help validate the assumptions that undergird our positions. We may find that we assume in the model that government research and development spending will crowd out private research & development in that industry. But this assumption can be tested, providing ground for more constructive debates. Similarly, simulations can help uncover human values that may not be well-articulated in a policy debate. For example, running a simulation to optimise economic growth may yield undesirable consequences, leading members to realise that, while economic growth is a goal, it may only be desirable when it improves living conditions for the public in a particular area. Testing assumptions and uncovering hidden values can help provide a firmer foundation for data-driven policy debates.

In fact, there is evidence that experimenting with models in itself may help drive consensus.11 As members examine the values, potential interventions and trade space of a topic, they’re likely to see more of the factors that they agree on, rather than the few on which they don’t. This certainly won’t bridge all ideological divides, but it can offer fertile ground for productive debate on evidence-based policies.

Potential challenges

While human-machine teaming has the potential to bring transformational benefits to the legislative process, it’s not without risk—with concerns over the quality of data, security and handling of the human component standing out. However, by focussing on the tasks we want those teams to perform, we can carefully control for risks while still realising the transformational benefits.

Data and model governance

AI’s outputs are only as good as the reliability of the model and the accuracy of the data. If the data isn’t accurate and fit for purpose or the model isn’t robust and explainable, it can create significant issues for the privacy, security, or fairness of an AI tool. There are already several frameworks for understanding and managing the risks of using AI in government. The National Institute of Standards and Technology and US Government Accountability Office have issued several important reports on the topic and organisations like the Department of Defense are already operationalising much of that guidance.12

These guidelines are by no means one-size-fits-all. The unique tasks that any given model performs can lead to different challenges that require different controls. For example, the central role played by historical data in the “microscope-like” use of ML to assess existing programmes means that those ML models need clean, accurate data that’s matched to their task. Open public data can help ensure the availability of good data.13 Similarly, tagging data sources with the use cases for which they are suitable can help avoid instances where data gathered in one context is used in another. For example, data that is representative for race and gender may not be representative for income level, so it shouldn’t be used in models where that’s an important parameter.14

While ML models are based on historical data, AI simulations are primarily based on assumptions of how factors relate to each other—whether that’s how individuals will react to a given choice or how smoking rates vary with rates of physical activity, for instance.15 These assumptions can and should be based on real data, but they’re still assumptions and there’s never a guarantee that the future will look like the past. Therefore, when we see models based on historical assumptions drifting away from expectations, it can be a sign that those assumptions need to be adjusted to better match a changing world. For example, many mass transit models assumed only a few major modes of transit such as car, bus, train, or bike. However, the sudden, massive growth of e-scooter ridership in 2018 and 2019 would have altered these models, forcing them to reevaluate assumptions about how people get around.16

As a result, in the case of AI simulations, special attention needs to be paid to the transparency of assumptions. Model governance procedures can help ensure the transparency of assumptions so that human team members can understand the context around how the simulation reached its conclusions.


Whether used to assess or shape legislation, AI tools need protection beyond typical cybersecurity considerations. The potential for adversaries to manipulate the outcomes of these AI models to tip policymaking to their advantage calls for careful safeguards.17

The centrality of data in “microscope-like” ML models means that they can be particularly vulnerable to the poisoning of training data—that is, tampering with data used to train ML models with the aim of influencing the results.18 Therefore, it’s critical to have controls on the access to and quality of the data. On the other hand, AI “simulator-like” models need safeguards placed on the variables, assumptions and even outputs of the models to avoid manipulation.

New processes, new skills, new training

The AI portion of the human-machine team isn’t the only aspect of the partnership that needs attention. Introducing new tools to the legislative process will require human team members to learn new skills, adapt to new processes and work together in new ways.

For example, as “microscope-like” ML models uncover new outcomes of public policies, policymakers will quickly find themselves consuming new types of data beyond bar charts and budget trends. New forms of information such as geospatial data, statistical relationships and more are likely to become important for decision-making. To ensure that these new sources of information are easily consumable, legislators and their staff may need new data visualisation tools. Similarly, staff members will likely need more data science skills to analyse, create and present the visualisations.

The “simulator-like” AI models may bring even more radical changes. In place of an “analyse then present” form of giving information to decision-makers, these models can allow for real-time decision support where policymakers can sit with staff to adjust models and examine conclusions as new data comes in. This shift has already taken place in industries such as auto racing, where Formula One race teams adjust strategy models in real time based on thousands of data points collected as cars race around the track.19 The shift to this mode of decision support can bring significant changes to how staffers spend their time. When Deloitte applied this concept to prototypical analysts in the intelligence community, for example, our model suggested that analysts could spend up to 39% more time advising decision-makers with the adoption of AI at scale.20

The way forward

Implementing AI in the legislative process can seem like a seismic transformation, but the shift is possible with the right commitment and investment. The experiences of other industries and even other parts of government already using AI models highlight that while the change is eminently possible, it will take considerable leadership—not just to put the technologies in place, but also to incorporate the training, education and business practices needed to make them work.

Lessons learned from other industries can help policymakers get started on their AI journey:

Don’t try to model everything

The scale of the issues that legislatures tackle is often tremendous and trying to model every aspect of each issue is practically impossible. The Formula One example shows that even relatively simple models can quickly get out of hand: For a single race, there are more race outcomes possible than there are electrons in the universe.21 This is where the human part of the human-machine team can help. Rather than trying to model everything, using human value judgements prior to modelling can help identify the core aspects of the problem that need to be modelled. In short, it all starts with deciding what the problem is and understanding what’s important. Then the technology can get to work.

Make a platform, not a solution

As the controller of the nation’s finances, the government also has a financial duty to the public. How can legislatures get the most out of AI without having to build a new tool from the ground up for every new policy debate? The answer is to build an AI-enabled platform, rather than a single-point solution. This is the approach Singapore took with its Virtual Singapore 3D model of the city-state. Virtual Singapore not only models the 3D layout of the city, but also allows for hosting of all other manner of data sources such as census and geospatial data.22 That way, when a new problem emerges, developers can simply create a new app within Virtual Singapore to run simulations about the new issue. Such an approach would allow legislatures to tap into AI in a way that’s cost-effective, efficient and able to evolve as technology changes.

Invest in the human dimension

Finally, the human element of the human-machine team is critical to the long-term success of digital transformations. AI is a powerful tool. It can run thousands of extremely precise calculations on mountains of data but, importantly for the purposes of legislating, AI cannot make value judgements.23 AI can calculate the fastest, cheapest, or largest solution to a problem. But it cannot tell you if that solution is good or bad, right or wrong, desirable or undesirable. So, the human element will always be critical to legislative decision-making. As a result, leaders should pay attention to the new tasks that may take up more time, new skills that may require difficult retraining and even new career paths that may change employees’ life goals. Taking care of the people will help take care of the technology.

AI is a powerful tool for the assessments and simulations that legislatures around the world need in their legislative processes. Pairing AI with the right people and the right processes can help provide a common foundation for debate, encourage consensus and deliver meaningful results for the public.

About the Deloitte AI Institute

The Deloitte AI Institute helps organisations connect all the different dimensions of the robust, highly dynamic and rapidly evolving AI ecosystem. The AI Institute leads conversations on applied AI innovation across industries, using cutting-edge insights to promote human-machine collaboration in the Age of With™. The Deloitte AI Institute aims to promote dialogue about and development of artificial intelligence, stimulate innovation and examine challenges to AI implementation and ways to address them. The AI Institute collaborates with an ecosystem composed of academic research groups, start-ups, entrepreneurs, innovators, mature AI product leaders and AI visionaries to explore key areas of artificial intelligence including risks, policies, ethics, future of work and talent and applied AI use cases. Combined with Deloitte’s deep knowledge and experience in artificial intelligence applications, the institute helps make sense of this complex ecosystem and, as a result, delivers impactful perspectives to help organisations succeed by making informed AI decisions. No matter what stage of the AI journey you’re in—whether you’re a board member or C-suite leader driving strategy for your organisation, or a hands-on data scientist bringing an AI strategy to life—the Deloitte AI institute can help you learn more about how enterprises across the world are leveraging AI for a competitive advantage. Visit us at the Deloitte AI Institute for the full body of our work, subscribe to our podcasts and newsletter and join us at our meetups and live events. Let’s explore the future of AI together. 


  1. Abraham Lincoln, "Gettysburg address delivered at Gettysburg Pa. Nov. 19th, 1863," Library of Congress, November 19, 1863

    View in Article
  2. Peter Viechnicki and William D. Eggers, How much time and money can AI save government? Deloitte Insights, April 26, 2017; Edward Van Buren et al., Scaling AI in government—How to reach the heights of enterprisewide adoption of AI, Deloitte Insights, December 13, 2021.


    View in Article

    Muhammad Nazrul Islam et al., "Machine learning to predict pregnancy outcomes: a systematic review, synthesizing framework and future research agenda," BMC Pregnancy and Childbirth 22, no. 348 (2022).



    View in Article
  4. Evan D. Peet, Dana Schultz, and Susan L. Lovejoy, Using an innovative database and machine learning to predict and reduce infant mortality, RAND Corporation, 2021.

    View in Article
  5. Melissa Tracy, Magdalena Cerdá, and Katherine M. Keyes, "Agent-based modeling in public health: Current applications and future directions," Annual Review of Public Health 39, (2018), pp 77-94.


    View in Article
  6. Nesta, "Innovation policy simulation for the smart economy," 2022.

    View in Article
  7. George J. Stigler, "The danger of making policy based on assumption," Chicago Booth Review, June 11, 1964.

    View in Article

    Will Douglas Heaven, "An AI can simulate an economy millions of times to create fairer tax policy," Technology Review, May 5, 2020.



    View in Article
  9. Aaron Parrott, Brian Umbenhauer, and Lane Warshaw, Digital twins—Bridging the physical and digital, Deloitte Insights, January 15, 2020.

    View in Article

    Cerdá Magdalena, Tracy Melissa, and Katharine M. Keyes, “Reducing urban violence: A contrast of public health and criminal justice approaches,” Epidemiology 29, no.1 (January 2018): pp. 142–150.



    View in Article
  11. Juliette N. Rooney-Varga et al., Building consensus for ambitious climate action through the world climate simulation, Advancing Earth and Space Science, December 2021.

    View in Article
  12. NIST, AI risk management framework: Initial draft, March 17, 2022; U.S. Government Accountability Office, Artificial Intelligence: An accountability framework for federal agencies and other entities, June 30, 2021; Department of Defense, Responsible artificial intelligence strategy and implementation pathway, June 2020.

    View in Article
  13. Tasha Austin et al., Trustworthy open data for trustworthy AI, Deloitte Insights, December 10, 2021.

    View in Article
  14. Ibid.

    View in Article

    Benjamin Yarnoff et al., "Validation of the Prevention Impacts Simulation Model (PRISM)," Preventing Chronic Disease 18 (2022).



    View in Article

    Abubakr Ziedan et al., "Complement or compete? The effects of shared electric scooters on bus ridership," Transportation Research Part D: Transport and Environment 101, (2021): pp. 103098.



    View in Article
  17. The Director of National Intelligence has included “minimizing the potential for adversarial influence” as one of the key design principles for AI in the Intelligence Community. See: Office of the Director of National Intelligence, “Principles of artificial intelligence ethics for the Intelligence Community,” accessed September 7, 2022; for an introduction to adversarial AI, see: William Dixon, “What is adversarial artificial intelligence and why does it matter?” World Economic Forum, November 21, 2018.

    View in Article
  18. Jacob Steinhardt, Pang Wei Koh, and Percy Liang, Certified defenses for data poisoning attacks, accessed September 8, 2022.

    View in Article
  19. Joe Mariani, “Racing the future of production: A conversation with Simon Roberts, operations director of McLaren’s Formula one team,” Deloitte, January 22, 2018


    View in Article

    Kwasi Mitchell, Joe Mariani, Adam Routh, and Akash Keyal, "The future of intelligence analysis," Deloitte Insights, December 11, 2019. 



    View in Article
  21. Formula 1, “What does an F1 strategist do?.” YouTube video, 10: 29, posted January 9, 2020

    View in Article
  22. National Research Foundation, "Virtual Singapore," Singapore Government, accessed September 6, 2022

    View in Article

    Tamra Lysaght et al., "AI-assisted decision-making in healthcare," Asian Bioethics Review 11 (2019), pp. 299–314. 



    View in Article

I would like to thank Thirumalai Kannan D. and Pankaj Kamleshkumar for their exhaustive background research, as well as Davin O’ReganBrenna Sniderman, and Bill Eggers for their help in telling a compelling story. Finally, I would like to thank Peter BrownHannah Shea, and the whole team for making the testimony a realit

Did you find this useful?

Thanks for your feedback

If you would like to help improve further, please complete a 3-minute survey