Rethinking AI for Good Governance

Helen Margetts

This essay examines what AI can do for government, specifically through three gener-
ic tools at the heart of governance: detection, prediction, and data-driven decision-
making. Public sector functions, such as resource allocation and the protection of

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

rights, are more normatively loaded than those of firms, and AI poses greater ethical
challenges than earlier generations of digital technology, threatening transparency,
fairness, and accountability. The essay discusses how AI might be developed specifi-
cally for government, with a public digital ethos to protect these values. Three moves
that could maximize the transformative possibilities for a distinctively public sector
AI are the development of government capacity to foster innovation through AI; the
building of integrated and generalized models for policy-making; and the detection
and tackling of structural inequalities. Combined, these developments could offer
a model of data-intensive government that is more efficient, ethical, fair, prescient,
and resilient than ever before in administrative history.

F
rom the 2010s onward, data-fueled growth in the development of artificial
intelligence has made tremendous leaps forward in scientific advancements,
medical research, and economic innovation. AI research and development
is generally carried out by or geared toward the private sector, rather than gov-
ernment innovation, public service delivery, or policy-making. However, govern-
ments across the world have demonstrated strong interest in the potential of AI,
a welcome development after their disinterested approach to earlier digital sys-
tems.1 Security, intelligence, and defense agencies tend to be the most advanced,
but AI is starting to be used across civilian policy sectors, at all levels of govern-
ment, to tackle public good issues.2
What would a public sector AI look like? What might it offer to government in
terms of improving the delivery of public goods and the design of policy interven-
tions, or in tackling challenges that are specific to the public sector? Using a broad
definition of AI that includes machine learning (ML) and agent computing, this
essay considers the governmental tasks for which AI has already proved helpful:
detection, prediction, and simulation. The use of AI for these generic governmen-
tal tasks has both revealed and reinforced some key ethical requirements of fair-
ness, transparency, and accountability that a public sector AI would need to meet
with new frameworks for responsible innovation. The essay goes on to discuss

© 2022 by Helen Margetts

360 Published under a Creative Commons Attribution-
NonCommercial 4.0 International (CC BY-NC 4.0) license
https://doi.org/10.1162/DAED_a_01922
 Helen Margetts

where the development of a distinctively public AI might allow a more transfor-

mative model for government: specifically, developing internal capacity and ex-
pertise, building generalized models for policy-making, and, finally, going beyond
the development of ethical frameworks and guidance to tackle long-standing
inequalities and make government more ethical and responsive than it has ever
been before.
Computers were first adopted by the largest departments of the largest gov-
ernments in the 1950s.3 In the very early days, government was an innovator and

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

leader in digital technologies: the UK Post Office produced the world’s first digital
programmable computer in 1943, later used for code-breaking at Bletchley Park.4
But since then, in many or even most countries, governments’ digital systems
were progressively outsourced, often in very large contracts that stripped digital
expertise from the government. Partly for that reason, governments were slow
to adopt Internet-based services or communicate with citizens online; in gener-
al (there are exceptions), they have lagged behind the private sector in adopting
the latest generation of data-intensive technologies.5 However, there has recently
been much greater interest in the possibilities of data science and AI for govern-
ment. The number of UK government announcements that mentioned data sci-
ence and artificial intelligence rose from fifteen in 2015 to 272 in 2018. In the Unit-
ed States, a comprehensive study of the use of AI in the federal government found
that nearly half of federal agencies studied (45 percent) had experimented with AI
and related machine learning tools by 2020.6 AI has helped governments perform
three key tasks: detection, prediction, and simulation, all of which can improve
policy-making and service delivery.7 In a perhaps unanticipated way, AI also forc-
es governments to think about ethical issues and the ethos of the government’s
digital estate, often in ways that have not been explicitly discussed before.

G
overnments need detectors: instruments for taking in information. De-
tection is one of the “essential capabilities that any system of control
must possess at the point where it comes into contact with the world out-
side,”8 and governments are no exception. They need to understand societal and
economic behavior, trends, and patterns and calibrate public policy accordingly.
To do this, governments need to detect (and then minimize) unwanted behavior
by firms or individual citizens. For example, regulators need to be able to detect
harmful behavior in digital environments, where the machine learning capabili-
ties of large firms challenge traditional regulatory strategies and where the coun-
tering of online harms requires constant innovation.
Machine learning’s core competency in classification and clustering offers
government new capability in the detection and measurement of unwanted ac-
tivity in large data sets. For example, machine learning is valuable in the detection
of online harms such as hate speech, financial scams, problem gambling, bully-

151 (2) Spring 2022 361

Rethinking AI for Good Governance

ing, misleading advertising, or extreme threats and cyberattacks. Many agencies

or regulators either need to detect these harms, or to oversee firms in so doing,
requiring the building of machine learning “classifiers” trained on data generat-
ed by social media or other digital platforms. Growth of what is broadly called
“counter-adversarial technology” to counter online threats to state or society is
a particularly important task for “public” AI research and development, requir-
ing constant innovation, as offenders continually game platforms to evade detec-
tion.9 These techniques are of increasing importance to security and intelligence

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

agencies, going beyond the creation of dedicated red teams for adversarial test-
ing10 to the creation of generative adversarial networks (GANs), in which neural
networks are designed in tandem: one designed to be a generative network (the
forger) and the other a discriminative network (the forgery detector). Each net-
work can “train and better itself off the other, reducing the need for big labelled
training data.”11
Civilian agencies across sectors also benefit from enhanced detection capabil-
ities. For example, the U.S. Securities and Exchange Commission uses a historical
data set of past issuer filings and machine learning with a random forest model to
identify which filers might be engaged in suspect earnings management, relying
on indicators such as earnings restatements and past enforcement actions.12 De-
tection is enhanced by AI-powered developments in robotics, computer vision,
and spatial computing. Health research agencies have been particularly advanced
in the use of computer vision and machine learning models trained to detect early
signs of, for example, cancer. Law enforcement agencies have been early adopt-
ers of AI for detection, combining these tools with robotic devices and AI-related
technologies such as computer vision. The U.S Department of Homeland Secu-
rity’s Customs and Border Protection (CBP) agency has a long running program
of using facial recognition technology, growing out of the agency’s emphasis on
counterterrorism post 9/11, developed by a range of private vendors using deep
learning within their proprietary technologies.13

T
he predictive capacity of machine learning has much to offer regulatory agen-
cies and governments broadly, which are not known for their strength in
foresight or forecasting. Governments can use machine learning tools to
spot trends and relationships that might be of concern or identify failing institu-
tions or administrative units. For example, in 2020, the U.S. Food and Drug Admin-
istration used machine learning techniques to model relationships between drugs
and hepatic liver failure, with decision trees and simple neural networks used to
predict serious drug-related adverse outcomes. They utilized regularized regres-
sion models, random forests, and support vector techniques to construct a rank-
ordering of reports based on their probability of containing policy-relevant infor-
mation about safety concerns, allowing the agency to prioritize those most likely

362 Dædalus, the Journal of the American Academy of Arts & Sciences
 Helen Margetts

to reveal problems.14 More generally, the use of predictive risk-based models can
greatly enhance the prioritization of sites for inspection or monitoring, from water
pipes, factories, and restaurants to schools and hospitals, where early signs of fail-
ing organizations or worrying social trends may be picked up in transactional data.
Government agencies can use AI tools to predict aggregate demand, for exam-
ple, in schools, prisons, or children’s care facilities. Understanding future needs
is valuable for resource planning and optimization, allowing government agen-
cies to direct human attention or manpower where it is most required. Machine

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

learning models of COVID-19 spread during 2020–2021 might have been used to
direct resources such as ventilators, nurses, and drug treatments toward those ar-
eas likely to be most affected, and even to target vaccination programs. An inves-
tigation of data science in UK local government suggested that even in 2018, 15 per-
cent of local authorities in the United Kingdom were using data science to build
some kind of predictive capability, such as to target safety prevention measures at
the streets placing most demand on emergency services.15 Unsupervised learning
models are also utilized to categorize criminal activities from free-text data gener-
ated by complaints, of potential use across the UK criminal justice system.16
The use of prediction to deliver individual (as opposed to aggregate) risk scores
is much more controversial. For local authorities that have used predictive tech-
niques to identify the number of children that are likely to be at risk of abuse or
neglect, the next step from forecasting (say) demand for childcare places is likely
to be “which children?” Such a question would come naturally to social services
departments terrified of being held responsible for the next ghastly case of abuse
to hit the headlines, the next “Baby P.” But should a technique that is essentially
inductive be used in this way? A risk of 95 percent of being a victim of an abusive
incident means that there is still a chance that the event will not happen, and if
the figure is 65 percent, the meaning of the individual number is highly ambig-
uous. Social policy experts who advocate this kind of machine learning for deci-
sion support have built models to support childcare workers’ decision-making in
New Zealand, the United States, and Australia.17 But other studies have counseled
a more cautious and thoughtful approach, and noted the importance of the data
environment.18 The most feted version, in Pittsburgh, was built from a data-rich
environment providing a 360-degree view of all children’s and their families’ in-
teractions with state agencies throughout their lives, an environment that rarely
exists in local authorities. And such systems are extremely vulnerable to bias, es-
pecially where data are derived from the criminal justice system.
As with detection, the earliest examples of the use of machine learning for risk
prediction came from law enforcement agencies. In the United States, a promi-
nent example was the Correctional Offender Management Profiling for Alterna-
tive Sanctions (COMPAS) system, a decision support system for judges that as-
sesses the risk of an individual prisoner being likely to reoffend, and therefore in-

151 (2) Spring 2022 363

Rethinking AI for Good Governance

forming sentencing decisions. The judges receive risk scores in low, medium, and
high risk buckets, and feed this evidence into the decision-making process. A 2016
study by ProPublica showed that COMPAS exhibited racial bias, a claim that has
generated much discussion over this use of machine learning in legal judgments.19
The system also demonstrates some of the subtle but deep shifts in perceptions
within the policy-making system that occur when machine learning technologies
are introduced, bringing with them notions of statistical prediction to a “situa-
tion which was dominated by fundamental uncertainty about the outcome be-

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

fore,” according to one thoughtful case study on the implementation of COMPAS,
showing that practitioners within the system valued what they perceived as the
“research-based” nature of COMPAS results, which they felt reduced uncertainty
in the system.20

T
he third area in which AI-related technologies can help policy-makers in
the design of policy interventions and evidence-driven, data-intensive
decision-making is simulation. Governments need ways of testing out inter-
ventions before they are implemented to understand their likely effects, especially
those of costly new initiatives, major shifts in resource allocation, or cost-cutting
regimes aimed at saving public resources. In the past, the only option for trying
out initiatives was by running field experiments: randomized trials in which the
intervention is applied to a “treatment group” and the results are compared with
a “control group.” But such trials are expensive and take a long time, challenge
notions of public equity, and sometimes are just not possible due to attrition or
ethical constraints.21 In contrast, the availability of large-scale transactional data,
and innovative combinations of agent computing and machine learning, allow the
simulation of interventions so unintended consequences can be explored without
causing harm.
Like AI itself, agent computing is a form of modeling that has been in existence
for a long time but has been revolutionized by large quantities of data. The agent-
based method was developed within economics in the 1960s and 1970s for the
purposes of simulation, but these were “toy models”: formal models with hardly
any data, and when tested on data generated by real-life situations, they tended to
perform very badly indeed. In contrast, the kind of agent computing models used
now are based on large-scale data, which can replicate whole economies, with 120
million firms and workers.22 A modern agent-based model like this consists of
individual software agents, with states and rules of behavior, and large corpuses
of data pertaining to the agents’ behavior and relationships. Some computer sci-
entists have called for such models to be developed ex ante–“agent-based mod-
eling as a service”–so that in an emergency, it could be rapidly employed to feed
in key variables and model possible policy interventions. Mainstream economics
has been resistant to such innovations, and political systems have inbuilt tenden-

364 Dædalus, the Journal of the American Academy of Arts & Sciences
 Helen Margetts

cies to try out hurried policy decisions, such as not having enough police, or doc-
tors or nurses, and learning the hard way. But the disadvantages of this on-the-
hoof policy-­making were illustrated during the first stage of the COVID-19 crisis
in 2020, when in many countries, policies regarding masks, social distancing, and
lockdown measures were made in an ad hoc and politically motivated fashion.
Agent computing has gradually gained popularity as a standard tool for trans-
port planning, or to provide insight for decision-makers in disaster scenarios such
as a nuclear attack or pandemic.23 Researchers working with police forces are tri-

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

aling the use of large-scale, real-time transactional data from daily activities of in-
dividual police in an agent-based model that would allow police managers to try
out different levels of police resourcing and measure the potential effects on deliv-
ery of criminal justice.24 If viable, such models could have potential for other ar-
eas of the public sector, where large quantities of trained professionals are needed,
such as in education or health care. In this way, agent computing can be another
good way of optimizing resources, by testing out the impact of different levels of
manpower without experiencing unintended consequences. Similarly, the United
Nations Development Programme is using an agent computing model to help de-
veloping countries work out which policies–such as health, education, transpor-
tation, and so on–should be prioritized in order to meet their sustainable develop-
ment goals.25 Researchers have started to explore the possibilities of “societal dig-
ital twins”: a combination of spatial computing, agent-based models, and “digital
twins,” or virtual data-driven replicas of real-world systems. These have become
popular for physical systems in engineering or infrastructure planning, although
proponents warn that the complexity of social systems renders the social equiva-
lent of digital twins “a long way from being able to simulate real human systems.”26

G
overnments of the progressive era of public administration from the late
nineteenth and early twentieth centuries stressed the need for a “pub-
lic service ethos” to limit corruption, waste, and incompetence. Such an
ethos prioritized values of honesty and fairness in an attempt to distinguish pub-
lic officials from the “inherently venal” nature of politicians and an increasingly
corrupt private sector.27 But as state operations became increasingly automated,
and personnel were replaced with digital systems, which were then outsourced to
computer services providers, there was a diminishing sense in which this ethos
could be said to apply to government’s digital estate.28 The advent of AI, howev-
er, has forced a rethink about the need to address issues of fairness, accountabili-
ty, and transparency in the way that government uses technology, given that they
pose greater challenges to these values than earlier generations of technology used
by government.
It is around ethical questions such as fairness that the distinctiveness of the
public sector becomes stark. If (say) Amazon uses sophisticated AI algorithms to

151 (2) Spring 2022 365

Rethinking AI for Good Governance

target customers in a biased way, it can cause offense, but it is not on the same
scale as a biased decision over someone’s prison sentence or benefits application.
Users of digital platforms know very little about the operation of search or news-
feed algorithms, yet will rightly have quite different expectations of their right to
understand how decisions on their benefit entitlement or health care coverage
have been made. The opaqueness of AI technology is accepted in the private sec-
tor, but it challenges government transparency.
From the late 2010s onward, there has been a burgeoning array of papers, re-

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

views, and frameworks aimed at tackling these issues for the use of AI in the pub-
lic sector. The most comprehensive and widely used across the UK government is
based on the principles of fairness, accountability, trustworthiness, and transpar-
ency, and a related framework was applied to the use of AI in the COVID-19 crisis.29
Policy-makers are starting to coalesce around frameworks like these, and ethics
researchers are starting to build the kinds of tools that can make them usable and
bring them directly into practice. It might be argued that progress is greater here
than it has been in the private sector. There is more willingness to contemplate
using less innovative–or differently innovative–models in order, for example, to
make AI more transparent and explainable in the process of high-stakes decisions
or heavily regulated sectors.30
The development of such frameworks could lead to a kind of public ethos for
AI, to embed values in the technological systems that have replaced so much of
government administration. Such an ethos would not just apply to AI, but to the
legacy systems and other technologies that first started to enter government in
the 1950s, and could be highly beneficial to the public acceptance of AI.31 There
is a tendency to believe that the technological tide will wash over us, fueled by
media and business school hype over “superintelligent” robots and literary and
cinematic tropes of robots indistinguishable from humans, powered by general
AI. If we do not design appropriate accountability frameworks, then politicians
and policy-makers will take advantage of this blame-shifting possibility. This will
range from cases like the UK prime minister blaming poor statistical processes to
calculate public examination results after school closures in the 2020 pandemic
prevented exams from taking place as a “mutant algorithm,” to the more nuanced
and unconscious shifting of responsibility to statistical processes involved in judi-
cial decision-making with AI observed above. A public sector AI in which fairness,
accountability, and transparency are prioritized would be viewed as more trust-
worthy, working against such perceptions.

S
o in what areas might government do more with AI? By 2021, government’s
use of AI was starting to speed up; the large-scale study of the use of AI by
the U.S. federal government concluded in 2020 that “though the sophistica-
tion of many of these tools lags behind the private sector, the pace of AI/ML de-

366 Dædalus, the Journal of the American Academy of Arts & Sciences
 Helen Margetts

velopment in government seems to be accelerating.”32 However, there are various

ways that AI could have a more transformative effect.
First, governments could prioritize the development of expertise and capacity
in AI to foster innovation and overcome some of the recurring challenges. As not-
ed above, the history of government computing has been characterized by large-
scale contracting to global computer services providers, but AI does not lend itself
to this kind of outsourcing, whereby governments lose control of key features. For
example, the U.S. CBP was criticized in 2020 for being unable to explain failure

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

rates of biometric scanning technology “due to the proprietary technology being
used.”33 Similar issues have dogged the adoption of facial recognition technolo-
gies by police agencies, with moratoria announced in several cities. There is ev-
idence that government agencies realize the importance of developing capacity:
the same U.S. study also found that “over half of applications were built in-house,
suggesting there is substantial creative appetite within agencies.”
An area with great scope is the use of data-intensive technologies to develop
new generalized models of policy-making. Governments have little tradition of
using transactional data to inform decision-making. In the classic Weberian mod-
el of bureaucracy, data are compressed within files, available for checking indi-
vidual pieces of information, but generating no usable data for analytics.34 This
characteristic of governments’ information architecture persisted into the era of
computerization, with a lack of usable data remaining a feature of the “legacy sys-
tems” of many governments. This point was well illustrated during the first wave
of the COVID-19 pandemic, when many countries discovered that they lacked the
kinds of data and modeling that could help design interventions. Key data flows
did not exist in real time; in the United Kingdom, for example, it turned out that
data for deaths were available only several weeks after the death had occurred.
Data were not fine-grained enough; the design of a stimulus package requires
sectoral-level data in order to target resources to those firms most in need. Mod-
eling took place in silos such as public health, health care, education, or the econ-
omy, meaning that interventions were targeted only at (say) economic recovery
or the health crisis, rather than an integrated approach taking account of the fact
that the domains were intertwined. Resilient policy-making would involve build-
ing such data flows and using agent computing, machine learning, and other AI
methodologies to create integrative models to both recover from the current crisis
and face future shocks.35
Finally, perhaps the most ambitious use of AI would be to tackle issues of
equality and fairness in governmental systems in a profound and transforma-
tive way, identifying and reforming long-standing biases in resource allocation,
decision-making, the administering of justice, and the delivery of services. Many
of the causes of bias and unfairness in machine learning, for example, come from
training data generated by the existing system. The COVID-19 pandemic revealed

151 (2) Spring 2022 367

Rethinking AI for Good Governance

many structural inequalities in how citizens are treated – for example, in the de-
livery of health care to people from different ethnic groups – just as the mobiliza-
tion around race has revealed systemic racism in police practice. Data and model-
ing have made these biases and inequalities explicit, sometimes for the first time.
Some researchers have suggested that we might develop AI models that incorpo-
rate these different sources of data and combine insights from a range of models
(so-called ensemble learning) aimed at the needs of different societal groups.36
Such models might be used to produce unbiased resource allocation methods and

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

decision support systems for public professionals, helping to make government
better, in every sense of the word, than ever before.

A
rtificial intelligence can help with core tasks of government. These tech-
nologies can enable real-time, transactional data to enhance govern-
ment’s armory of detecting tools, to build predictive models to support
decision-making, and to use simulation to design policy interventions that avoid
unintended consequences. They face distinct ethical challenges when used for
these public sector tasks, requiring new frameworks for responsible innovation.
As policy-makers become more sophisticated in their use of AI, these technologies
might be developed to overcome fragilities exposed in the COVID-19 pandemic,
to create new, more resilient models of policy-making to face future shocks, and
to “build back better,” the catchphrase of many governments in the postpandem-
ic era. AI can reveal and perhaps mitigate some structural biases and might even
be used to tackle some profound inequalities in the distribution of resources and
the design and the delivery of public services such as education and health care.
This would require a specific branch of AI research and development, geared at
distinctively public sector tasks and needs. Such a remit would be no less complex
or challenging than for any other field of AI. Indeed, some deep learning experts
suggest that even where machine learning has had success, as in medical diagnosis
of X-ray images, models are still outperformed by human radiologists in clinical
settings.37 But the potential public good benefits are huge.

about the author

Helen Margetts OBE FBA is Professor of Society and the Internet at the Oxford In-
ternet Institute at the University of Oxford and Programme Director for Public Pol-
icy at The Alan Turing Institute, London. She is the author of, among other publi-
cations, Political Turbulence: How Social Media Shape Collective Action (with Peter John,
Scott Hale, and Taha Yasseri, 2016), Digital Era Governance: IT Corporations, the State,
and E-Government, 2nd ed. (with Patrick Dunleavy, Simon Bastow, and Jane Tinkler,

368 Dædalus, the Journal of the American Academy of Arts & Sciences
 Helen Margetts

2008), and The Tools of Government in the Digital Age (with Christopher Hood, 2007),
and editor of Paradoxes of Modernization: Unintended Consequences of Public Policy Reform
(with Perri 6 and Christopher Hood, 2010).

endnotes
1 Helen Margetts, Information Technology in Government: Britain and America (New York: Rout-
ledge, 1999).

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

2 Thomas M. Vogl, Cathrine Seidelin, Bharath Ganesh, and Jonathan Bright, “Smart Tech-
nology and the Emergence of Algorithmic Bureaucracy: Artificial Intelligence in UK
Local Authorities,” Public Administration Review 80 (6) (2020): 946–961, https://online
library.wiley.com/doi/10.1111/puar.13286.
3 Margetts, Information Technology in Government.
4 Mar Hicks, Programmed Inequality: How Britain Discarded Women Technologists and Lost Its Edge
in Computing (Cambridge, Mass.: MIT Press, 2017).
5 Margetts, Information Technology in Government; Patrick Dunleavy, Helen Margetts, Jane
Tinkler, and Simon Bastow, Digital Era Governance: IT Corporations, the State, and E-Gov-
ernment (Oxford: Oxford University Press, 2006); and Helen Margetts and Cosmina
Dorobantu, “Rethink Government with AI,” comment, Nature, April 9, 2019, 163–165.
6 David Freeman Engstrom, Daniel E. Ho, Catherine M. Sharkey, and Mariano-Florentino
Cuéllar, Government by Algorithm: Artificial Intelligence in Federal Administrative Agencies
(Washington, D.C.: Administrative Conference of the United States, 2020).
7 Margetts and Dorobantu, “Rethink Government with AI.”
8 Christopher Hood and Helen Margetts, The Tools of Government in the Digital Age (London:
Macmillan, 2007), 3.
9 Abhijnan Rej, “Artificial Intelligence for the Indo-Pacific: A Blueprint for 2030,” The
Diplomat, November 27, 2020; and Bertie Vidgen, Alex Harris, Dong Nguyen, et al.,
“Challenges and Frontiers in Abusive Content Detection,” Proceedings of the Third Work-
shop on Abusive Language Online, Florence, Italy, August 1, 2019.
10 National Security Commission on Artificial Intelligence, Final Report (Washington, D.C.:
National Security Commission on Artificial Intelligence, 2021), 383.
11 Ibid., 607; and Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, et al., “Generative
Adversarial Nets,” Neural Information Processing Systems 27 (2014), https://papers.nips.cc/
paper/5423-generative-adversarial-nets.pdf.
12 Engstrom et al., Government by Algorithm, 23.
13 Ibid., 32.
14 Ibid.
15 Jonathan Bright, Bharath Ganesh, Cathrine Seidelin, and Thomas M. Vogl, “Data Sci-
ence for Local Government” (Oxford: Oxford Internet Institute, University of Ox-
ford, 2019); and Vogl et al., “Smart Technology and the Emergence of Algorithmic
Bureaucracy.”

151 (2) Spring 2022 369

Rethinking AI for Good Governance

16 Daniel Birks, Alex Coleman, and David Jackson, “Unsupervised Identification of Crime
Problems from Police Free-Text Data,” Crime Science 9 (1) (2020): 1–19.
17 Rhema Vaithianathan, Emily Putnam-Hornstein, Nan Jiang, et al., Developing Predictive
Models to Support Child Maltreatment Hotline Screening Decisions: Allegheny County Methodology
and Implementation (New Zealand: Centre for Social Data Analytics, AUT University, 2017);
and Rhema Vaithianathan, “Five Lessons for Implementing Predictive Analytics in Child
Welfare,” The Chronicle of Social Change, August 29, 2017, https:// chronicleofsocialchange
.org/opinion/five-lessonsimplementing-predictive-analytics-child-welfare/27870.
18 David Leslie, Lisa Holmes, Christina Hitrova, and Ellie Ott, Ethics Review of Machine Learning

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

in Children’s Social Care (London: What Works for Children’s Social Care, 2020), https://
www.turing.ac.uk/research/publications/ethics-machine-learning-childrens-social-care.
19 Alex Chohlas-Wood, “Understanding Risk Assessment Instruments in Criminal Justice,”
Brookings Institution’s series on AI and Bias, June 19, 2020, https://www.brookings
.edu/research/understanding-risk-assessment-instruments-in-criminal-justice/.
20 See Aleš Završnik, “Algorithmic Justice: Algorithms and Big Data in Criminal Justice
Settings,” European Journal of Criminology 18 (5) (2019).
21 Helen Margetts and Gerry Stoker, “The Experimental Method,” in Theory and Methods in
Political Science, ed. Vivien Lowndes, David Marsh, and Gerry Stoker (London: Mac-
millan International Higher Education, 2017); and Peter John, Field Experiments in Polit-
ical Science and Public Policy: Practical Lessons in Design and Delivery (New York: Routledge,
2019).
22 Robert Axtell, “Endogenous Firm Dynamics and Labor Flows via Heterogeneous Agents,”
in Handbook of Computational Economics, 4th ed., ed. Cars Hommes and Blake LeBaron
(Amsterdam: North-Holland, 2018), 157–213.
23 M. Mitchell Waldrop, “Free Agents,” Science 360 (6385) (2018): 144–147.
24 Julian Laufs, Kate Bowers, Daniel Birks, and Shane D. Johnson, “Understanding the Con-
cept of ‘Demand’ in Policing: A Scoping Review and Resulting Implications for De-
mand Management,” Policing and Society 31 (8) (2020): 1–24.
25 Omar A. Guerrero and Gonzalo Castañeda, “Policy Priority Inference: A Computational
Framework to Analyze the Allocation of Resources for the Sustainable Development
Goals,” Data & Policy 2 (2020).
26 Dan Birks, Alison Heppenstall, and Nick Malleson, “Towards the Development of Soci-
etal Twins,” Frontiers in Artificial Intelligence and Applications 325 (2020): 2883–2884.
27 Christopher Hood, “A Public Management for All Seasons?” Public Administration 69 (1)
(1991): 3–19; and Christopher Hood, Explaining Economic Policy Reversals (Buckingham:
Open University Press, 1994).
28 Margetts, Information Technology in Government; and Dunleavy et al., Digital Era Governance.
29 Engstrom et al., Government by Algorithm; and David Leslie, “Tackling COVID-19 through
Responsible AI Innovation: Five Steps in the Right Direction,” Harvard Data Science
Review (2020).
30 United Kingdom Information Commissioner’s Office and The Alan Turing Institute,
Explaining Decisions Made with AI (Wilmslow, United Kingdom: Information Commis-
sioner’s Office, 2020), https://ico.org.uk/for-organisations/guide-to-data-protection/
key-data-protection-themes/explaining-decisions-made-with-ai/.

370 Dædalus, the Journal of the American Academy of Arts & Sciences
 Helen Margetts

31 Helen Margetts, “Post Office Scandal Reveals a Hidden World of Outsourced IT the
Government Trusts but Does Not Understand,” The Conversation, April 29, 2021.
32 Engstrom et al., Government by Algorithm, 55; and Vogl et al., “Smart Technology and the
Emergence of Algorithmic Bureaucracy.”
33 Engstrom et al., Government by Algorithm, 33–34.
34 Patrick Dunleavy and Helen Margetts, Digital Era Governance and Bureaucratic Change
(Oxford: Oxford University Press, 2022).
35 Jessica Flack and Melanie Mitchell, “Uncertain Times,” Aeon, August 21, 2020, https://

Downloaded from http://direct.mit.edu/daed/article-pdf/151/2/360/2054395/daed_a_01922.pdf by guest on 24 November 2022

aeon.co/essays/complex-systems-science-allows-us-to-see-new-paths-forward; and B.
MacArthur, Cosmina Dorobantu, and Helen Margetts, “Resilient Policy-Making Re-
quires Data Science Reform,” Nature (under review).
36 MacArthur et al., “Resilient Policy-Making Requires Data Science Reform.”
37 Tekla S. Perry, “Andrew Ng X-Rays the AI Hype,” IEEE Spectrum, May 2, 2021, https://
spectrum.ieee.org/view-from-the-valley/artificial-intelligence/machine-learning/
andrew-ng-xrays-the-ai-hype.

151 (2) Spring 2022 371