Annual Review of Criminology

Artificial Intelligence,
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Predictive Policing, and

Risk Assessment for Law
Enforcement
Richard A. Berk
Departments of Statistics and Criminology, University of Pennsylvania, Philadelphia,
Pennsylvania 19104, USA; email: berkr@[Link]

Annu. Rev. Criminol. 2021. 4:209–37 Keywords

First published as a Review in Advance on
artificial intelligence, algorithms, machine learning, predictive policing,
November 13, 2020
risk assessment, criminal justice, fairness, transparency
The Annual Review of Criminology is online at
[Link] Abstract
[Link]
There are widespread concerns about the use of artificial intelligence in
012342
law enforcement. Predictive policing and risk assessment are salient exam-
Copyright © 2021 by Annual Reviews.
ples. Worries include the accuracy of forecasts that guide both activities, the
All rights reserved
prospect of bias, and an apparent lack of operational transparency. Nearly
breathless media coverage of artificial intelligence helps shape the narrative.
In this review, we address these issues by first unpacking depictions of ar-
tificial intelligence. Its use in predictive policing to forecast crimes in time
and space is largely an exercise in spatial statistics that in principle can make
policing more effective and more surgical. Its use in criminal justice risk as-
sessment to forecast who will commit crimes is largely an exercise in adap-
tive, nonparametric regression. It can in principle allow law enforcement
agencies to better provide for public safety with the least restrictive means
necessary, which can mean far less use of incarceration. None of this is mys-
terious. Nevertheless, concerns about accuracy, fairness, and transparency
are real, and there are tradeoffs between them for which there can be no
technical fix. You can’t have it all. Solutions will be found through political
and legislative processes achieving an acceptable balance between competing
priorities.

209
 INTRODUCTION
The use of artificial intelligence in the public and private sectors has of late generated consid-
erable controversy. There can be concerns about technical matters (Domingos 2015), jurispru-
dence (Coglianese & Lehr 2017, Kroll et al. 2017), ethics (Berk et al. 2018, Dwork & Roth 2014),
and even survival of the human species (Cellen-Jones 2014, Scharre 2018). Some treatments are
cross-cutting integrations (Kearns & Roth 2019, Mitchell 2019a). Although the issues are very real
and often judiciously addressed, many of the debates can be compromised by misunderstandings.
Key concepts too often are undefined or defined differently, assertions commonly are made that
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

are baseless or misrepresentations, and focus can be lost to a variety of other agendas that may
be political, economic, or narrowly self-serving. Mass-media treatments that otherwise are very
thoughtful (Mullainathan 2019, Pande 2018, Smith 2019) can at times get carried away (Chinoy
2019, Starr 2014).
Many of the same issues arise in criminal justice settings where the discussions can be no
less contentious or poorly informed. Perhaps the best example is the controversy surrounding
the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) risk in-
strument protocol. COMPAS is a widely used risk assessment tool deployed to inform criminal
justice decisions such as sentencing. Angwin and her colleagues published an analysis claiming
that the instrument was racially biased (Angwin et al. 2016). Their assertions were widely dis-
seminated and, in many circles, accepted. However, the critique was far too narrowly formu-
lated and poorly informed about the nature of racial bias and necessary tradeoffs between dif-
ferent kinds of unfairness (Berk et al. 2018, Spielkamp 2017). Furthermore, COMPAS became
the poster child for all statistical forms of risk assessment. At the very least, it was guilt by associa-
tion. A second equally controversial example is software used for facial recognition. The Guardian
found in an internet search more than 130 articles on facial recognition over the past several
years ([Link] Among the many issues
raised were privacy, politically motivated surveillance, and differential accuracy depending on race
and gender.
Within this larger context, the pages below address the use of data by law enforcement to
make predictions of various kinds. In some forms, the enterprise is called predictive policing. In
other forms, it is called risk assessment. In either case, the predictions are often said to be the
product of artificial intelligence, and controversy necessarily follows. There are, again, concerns
about technical matters, jurisprudence, and ethics, with a special emphasis on criminal justice
fairness. There will be no resolution to these issues in this review because the main disagreements
are largely unresolved. Rather, the goal is to clarify the discussion and convey empirical findings
when they can be found.

THE MUDDLE OF ARTIFICIAL INTELLIGENCE

Before examining law enforcement applications, confusion about the nature of artificial intelli-
gence must be addressed. In part because there seems to be nothing more human than thinking,
virtually any collection of computer code becomes fair game if it produces results like those that
could arise from human mental processes. In an early formulation, Alan Turing (1950) argued that
if a machine could exhibit behavior that was indistinguishable from that of a human, the machine
would be executing intelligent behavior. A recent application of this perspective had a computer
make a telephone reservation at a restaurant through such a human-like conversation that the
person taking the reservation did not know he was speaking to a computer. “It had mastered the
cadence of a human conversation, replete with uh’s and hhhm’s that we sprinkle into conversations
while we’re thinking” (Krohn 2019, pp. 35–36). However, the Turing Test is as controversial as it

210 Berk
 is fascinating and has been met with a number of incisive critiques (Grahm & Dowe 2019). Among
the simplest is to question whether imitation by itself is enough. Should one not care about how
the imitation is assembled?
A rather different definition of artificial intelligence boils down to any device that is able
to sense its environment and then take actions increasing the likelihood that its objectives are
achieved (Poole et al. 1998). However, this formulation represents a very low bar because even
the common household thermostat qualifies. An even lower bar only requires that an intelligent
device compete favorably with humans on the same cognitive tasks. For example, if my goal is to
travel by car to a supermarket to which I have never been, the GPS-based application Waze qual-
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

ifies if its route and mine get me to my destination in effectively the same elapsed time. It would
not matter if we took the same path or different paths. A very recent and spectacular illustration is
the victory of Google’s AlphaGoZero over the best Go player in the world (Krohn 2019). In both
these examples, there is no requirement that the machine processes and the human processes be
the same. All that matters is comparable output.
A related conception works backward from particular tasks whose execution exemplifies artifi-
cial intelligence (Goodfellow et al. 2016, p. 1).

In the early days of artificial intelligence, the field rapidly tackled and solved problems that were intel-
lectually difficult for human beings but relatively straightforward for computers—problems that can be
described by a list of formal, mathematical rules. The true challenge to artificial intelligence proved to
be solving tasks that are easy for people to perform but hard for people to describe formally—problems
that we solve intuitively, that feel automatic, like recognizing spoken words or faces in images.

Artificial intelligence becomes nothing more than a means to tackle certain kinds of challenging
empirical exercises that are easy for humans and represent activities that seem to be distinctly
human such as speech.
In short, there is no commonly accepted definition of artificial intelligence and, so far at least, all
the candidates have substantial flaws. Part of the confusion is that the activities to which artificial
intelligence is applied are very heterogeneous and have become more diverse over time. In this
context, there is merit in Jon Krohn’s three kinds of artificial intelligence (Krohn 2019), which
distinguish between real-world applications and applications that currently are highly speculative.

 Artificial narrow intelligence performs specific tasks such as translating languages, recog-
nizing faces, providing medical diagnoses, guiding self-driving cars, and playing games such
as chess or Go. This form of artificial intelligence currently is available.
 Artificial general intelligence implements a single set of code able to perform effectively
across a range of tasks formerly undertaken separately under artificial narrow intelligence.
There is at present no such code, and experts predict that it is several decades away, if it can
be achieved at all. For example, the same code that translated spoken words into written
text would also run a self-driving car, search video for specified images (e.g., CCTV video
of a man in a blue sweater and tan sneakers who witnessed an industrial accident), and win
at games of Settlers of Catan ([Link]
 Artificial superintelligence, which is little more than a placeholder for machine intelligence,
is, at best, 50 years in the future and would surpass human intelligence. Current thinking
has centered on the many obstacles that may be insurmountable, an important one be-
ing the enormous computing capability required that even quantum computing may not
be able to provide (Arute et al. 2019, Metz 2019). Equally challenging are inherent con-
nections between different kinds of thinking, feeling, and reasoning that all contribute to
human intelligence (Mitchell 2019a). “Human intelligence is a strongly integrated system,

[Link] • Criminal Justice and AI 211

 one whose many attributes—including emotions, desires, and a strong sense of selfhood and
autonomy—can’t easily be separated” (Mitchell 2019b, p. A27).

There is more to intelligence than information processing. Furthermore, as machines acquire

increasing cognitive skill and neuroscience continues to make progress, conceptions of intelligence
will evolve.
To summarize, the artificial intelligence of today is computer code written to accomplish a
specific empirical task. Restated, it is just a computational procedure. In law enforcement settings,
these procedures are by and large enhancements of activities that humans are already performing.
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Arguably, computers can do them faster and more accurately, but the internal operations are far
removed from the way humans actually think. It is not even clear that the term intelligence applies.

MODELS, ALGORITHMS, AND MACHINE LEARNING

An essential feature of artificial narrow intelligence is that it depends on algorithms, not models.
Algorithms are methods to compute things. When you balance your checkbook, you are using an
algorithm. When you compute a regression coefficient or undertake a singular value decomposi-
tion, you are using an algorithm. In contrast, a model is an explicit theory about how the world
works. Force = Mass × Acceleration is a model. The goal is explanation, perhaps involving cause
and effect. Algorithms should not be confused with models because a range of errors can easily
follow (Breiman 2001b). For example, model misspecification is a pervasive problem in criminol-
ogy and the social sciences more generally (Freedman 2009), but model misspecification is not
even defined for algorithms (Berk 2020b); it makes no sense to say an algorithm is misspecified.
Consider the pair of equations for the ubiquitous linear model with a numeric response
variable.

yi = β0 + β1 x1i + β2 x2i + · · · + β p x pi + εi , 1.

where

εi ∼ NIID(0, σ 2 ). 2.

Equations 1 and 2 are a theory of how each case i came to be. Each response value is a linear
combination of regressor variables weighted by their regression coefficients, a constant serving as
the y-intercept, and a perturbation drawn independently and at random from some probability
distribution, sometimes assumed to be normal.
Equations 1 and 2 are meant to capture exactly how the response values are generated; nature
literally adopts the conventional linear model (Berk et al. 2014). For Equation 1, this means that
the first-order conditions must be met; the mean function is correct. It then follows that one can
obtain unbiased estimates of coefficients in Equation 1. To obtain the unbiased standard errors
required for valid statistical tests and confidence intervals, Equation 2 must be correct as well.
These are the second-order conditions. All uncertainty is produced precisely as Equation 2 states.
Furthermore, if the first-order conditions are not met, the second-order conditions, as a formal
matter, cannot be met. Even when the most apparent inferential consequences of a misspecified
mean function are asymptotically addressed with robust standard errors such as the sandwich,
violations of the first-order conditions invalidate all statistical tests and confidence. Bias in the
regression coefficients systematically makes the coefficient estimates too large or too small, which
offsets the test statistics and systematically makes the p-values too large or too small.
Although in practice, Equations 1 and 2 can be used for forecasting, interest typically centers on
estimates of the regression coefficients. The primary goal is explanation, i.e., why things work the

212 Berk
 way they do. These issues are revisited below when predictive policing and criminal justice risk
assessment are discussed. For more background, didactic treatments are readily available (Berk
et al. 2019, Freedman 2009).
In conventional matrix notation, a common mathematical expression for the calculations by
which coefficient estimates from Equations 1 and 2 are obtained is

β = (X  X )−1 X Y . 3.

Equation 3 is a computational procedure that can serve as an algorithm. It makes absolutely no

Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

claims about how the world works and has no aspirations to explain anything. Because there are no
first- and second-order conditions to meet, concepts like model misspecification are inapplicable.
Its sole purpose is to provide values for the regression coefficients and from them to derive fitted
values. As addressed shortly, unbiased estimates are no longer the Holy Grail. They ideally are
balanced against the estimation variance through the bias–variance tradeoff (Berk 2020b).
Artificial narrow intelligence is consistent with this perspective, although the algorithms can be
far less transparent. The usual aim is to produce fitted values that can be used to inform decisions.
In facial recognition, for instance, a properly trained algorithm can be tasked with identifying par-
ticular individuals from CCTV images. Explanation is irrelevant, and no claims are made beyond
whether a match is found.
Because artificial narrow intelligence does not depend on models, a wide variety of algorithms
can be used. These in general fit the data better than models (Berk & Bleich 2013). Many are
classified as machine learning (ML). Here too, there are no commonly accepted definitions, but
discussions usually center on data-fitting procedures that are adaptive (Berk 2020b). For example,
the size of the window when a moving average is computed is determined on the fly by some mea-
sure of fit (e.g., mean squared error). Window size is not determined before the data are analyzed.
The lasso (Tibshirani 1996), when used to determine the included predictors in a regression anal-
ysis, is in the same spirit because the model specification is not known before the analysis begins.
Adaptive fitting causes fundamental problems for statistical inference, although in some situations,
good solutions exist (Berk 2020b).
Immediately below is an illustration of an algorithmic procedure that has many similarities
to an ML approach called boosting (Berk 2020b). It emphasizes the adaptive nature of the fitting
process. These are the kinds of tools that are central when artificial narrow intelligence is deployed
for criminal justice applications. Rarely is there anything remotely like a model.

1. Specify a mean function as in Equation 1 and apply least squares as usual.

2. Compute the fitted values.
3. Compute the residuals.
4. Compute a measure of fit.
5. Apply Equation 1 again, but weight the data so that observations with larger absolute values
of the residuals are given more weight.
6. Update the fitted values obtained from the immediately preceding regression with the new
fitted values weighted by their measure of fit.
7. Repeat steps 2–6 until the quality of the fit does not improve (e.g., 1,000 times).
8. Output the fitted values.

The output of interest is the linear combination of weighted fitted values computed over many
passes through the data. The regression coefficients have no interpretative value because there
can be thousands of them calculated from differently weighted data sets. Equation 1 is not being
used as a model but, via Equation 3, is a component of an algorithm.

[Link] • Criminal Justice and AI 213

 Data
collection

Data Forecasting
management information

Trai
ning ANI data
and
test analysis
ing
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Decisions

Actions
Use

Figure 1
Artificial narrow intelligence (ANI) data analysis embedded in other activities and used to inform several criminal justice functions: data
collection, data management, decisions, and actions.

ARTIFICIAL NARROW INTELLIGENCE IN THE FIELD

Another important source of confusion is a failure to separate artificial narrow intelligence from
the other activities that surround it. Figure 1 underscores that there are essential steps before
the data analysis begins and after the data analysis ends. An ML algorithm is often blamed for
problems introduced by the data and data management or for decisions and actions taken after
the data analysis ends.
Typically, algorithms used in criminal justice settings must be trained. This requires training
data that include predictor variables, which are essentially regressors in the linear regression set-
ting, and one or more response variables to be fitted. Usually, there is a single response variable
that may be numeric or categorical. From the training, the algorithm learns how the predictor
variables are related to the response variable(s). Once these associations are established, they can
be used to construct imputed or forecasted values of the response variable(s) when those val-
ues are unknown. There should also be test data, generated in the same manner as the training
data but not used in the training. Because of the adaptive nature of learning algorithms, there
are complications for statistical inference that can sometimes be properly addressed with test
data. The test data are uncontaminated by adaptive procedures or “data snooping” in the training
process.
From Figure 1, an initial step is collecting and organizing relevant training and test data, and at
this point, a variety of difficulties may be introduced. For example, a prior record for misdemeanors
can contain arrests motivated by racial animus. Perhaps more fundamentally, the arresting charge
may not comport well with the actual behavior triggering the arrest. Regardless of an offender’s
race, police might in some cases overcharge to give prosecutors more leverage for plea bargaining.
There can also be recording or data processing errors so that, for instance, some birthdates are
incorrect. This can happen if, rather than obtaining the date of birth from a driver’s license, the
offender is asked for his or her date of birth. All errors, whatever their cause, are carried forward
in the training data and can affect results from the algorithm. There is growing evidence that this
is the major source of inaccuracies and unfairness in algorithms used by criminal justice decision-
makers (Berk 2020a). The algorithm is faultless and, indeed, there is ongoing work to design
algorithms that reduce the impact of problematic data (Berk et al. 2018).

214 Berk
 Once the trained algorithm reaches satisfactory performance in test data, it is ready for use.
Information used for imputation or prediction is presented to the trained algorithm. The output
is intended to inform criminal justice decisions that, in turn, affect actions by criminal justice
officials. Just as with the data, a variety of problems can be introduced. The algorithmic results for
particular cases might be inaccurate. Inappropriate actions can follow. Sometimes the algorithmic
results are ignored or misused. Again, inappropriate actions can follow. And the decisions and
actions may be inappropriate for reasons unrelated to an algorithm’s output. But, until there are
concrete actions, there can be no real harm and no real benefits. There is, therefore, an opportunity
to right obvious wrongs before actions are taken.
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

The key point is this: The algorithmic tools are embedded in several concomitant activities that
are typically far more consequential for the actions taken than the algorithms themselves. Reme-
dies for problematic actions should be introduced where they can be most effective. Currently at
least, this is unlikely to be for the algorithm.

PREDICTION
In this review, prediction and risk are the main focus. The other uses of algorithms need to
be briefly mentioned because some are inaccurately folded into predictive policing or risk as-
sessment. Algorithms are commonly used by law enforcement organizations for the following
tasks:
 administrative functions like payroll or equipment tracking;
 personnel matters such as performance evaluations of sworn officers;
 forensics such as crime lab case management, DNA matching, and metabolic profiling
(Dinis-Oliveira 2017);
 license plate recognition;
 organizing and analyzing evidence about specific crimes that have occurred;
 displaying crime statistics and time–space crime patterns on department websites;
 CCTV surveillance;
 organization, redaction, and retrieval of information, such as facial recognition, from CCTV
videos or body cameras;
 crime mapping and spatial analysis;
 the use of algorithms in smartphones and computer operating systems;
 scheduling and deployment of human and material law enforcement assets;
 voice to written report translators; and
 performance evaluations for geographical areas (e.g., precincts) as in COMPSTAT.
The algorithms employed do not have to be fancy or sophisticated. A simple two-way table
from an Excel spreadsheet qualifies. Also, there is no forecasting. The algorithms look back in
time to what has already happened or provide prescriptive guidance for future activities.
At least for administration and scheduling, there is similar software for prosecutors’ offices,
public defenders’ offices, administrative offices of the courts, and jail, prison, probation, and parole
administrations. Here too, none of the algorithms are engaged in forecasting, and the calculations
can be very simple.

Predictive Policing
Despite some wide-ranging expositions (Perry 2013), the foundations of predictive policing are
forecasts of when and where future crimes will occur. Based in part on the “criminology of place”

[Link] • Criminal Justice and AI 215

 (Weisburd et al. 2012), forecasts are made in time and space about crimes that will be committed.
The spatial units are locations that can be quite large, such as an entire precinct or census tract.
Locations also can be small, e.g., a particular street corner. Likewise, the temporal units can vary
dramatically in size: an administrative reporting period such as a month, a week, a day, or even a
particular time of day. One might forecast, for instance, a dramatic increase in assaults and rob-
beries on a particular block on a Saturday night just after the bars close. The spatial units also
can be types of locations, and the importance of the temporal units may be determined by social
circumstances. For example, one might forecast an increase in residential burglaries along major
bus routes on weekdays just after the school day ends. Or one might forecast an increase in jewelry
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

stores break-ins over long holiday weekends.

From such forecasts, police assets can be allocated to prevent crime or apprehend criminals.
An example of the former is patrolling at certain times and places to “show the flag.” An example
of the latter is staking out a particular jewelry store on Christmas Day. Sometimes both kinds of
tactics are implemented at once. For instance, parking garages at airports late at night may benefit
from a police presence to deter automobile thefts and/or catch would-be car thieves. In both cases,
policing can be more surgical.
There is nothing in such police practices that requires powerful data analyses. Indeed, police in
urban areas have been using their experience, tips from informants, complaints from the public,
police craft lore, and very simple mapping procedures to do predictive policing for well over a
century. Research on crime hot spots retrospectively helps justify this approach, although the
demonstrable reductions in crime are typically modest (Braga et al. 2014).
The advanced use of data and data analysis by police departments is a relatively recent phe-
nomenon. COMPSTAT procedures, employed primarily as an internal management tool to in-
crease police accountability, are a good, early benchmark (Weisburd et al. 2004). Implicit fore-
casts became common. For example, a precinct with a large number of reported crimes typically
was assumed to be a high-crime precinct in the immediate future. Performance was evaluated by
whether, after various reforms and interventions, crime dropped. It was then a short step to begin
formal forecasting coupled with efforts to improve the quality of local data.
As an empirical matter, virtually all modern predictive policing procedures begin with recorded
crimes in time and space. These are often displayed on maps, much in the spirit of pin maps used
by police for generations. For a given time period, the map shows where specific kinds of crimes
have occurred (e.g., armed robberies in the third precinct). If the crimes are spatially concentrated,
counts of the number of crimes are used. Such maps and the underlying data are available over
time. This is the raw material for the data analytics of predictive policing.
Commonly, there is temporal and spatial autocorrelation that can be exploited for prediction.
Crimes that are more proximate in space and/or time tend to be more similar. Crimes lagged in
time and offset in space can be used as predictors. That is, a crime count or crime event at location
i and time t becomes a function of lagged crime counts or crime events in time and offset crime
counts in space. The time units might be months, weeks, or even days. The spatial units might be
distance in miles or more simply, areas adjacent to location i. The metric of Euclidian distance has
much the same statistical properties as the metric of time and often can be analyzed in much the
same manner.
Expressed very generally,

Yit = f (Y(=i,=t ) , ε(it ) ), 4.

where ε is an IID (independent and identically distributed) perturbation that can also be lagged
in time and offset in space. Yit can represent crime in several different ways: as a count of the
number of crimes, the presence or absence of a crime, crime probability/density, or a parameter

216 Berk
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Crime
count Neighborhood, census tract,
or other spatial unit for which
forecasts could be needed
No dista
rth nc
−S e
ou
th

st
−We
East tance
dis

Figure 2
Distance density smoother resulting in a smoothed surface of crime counts.

of a statistical model (e.g., the rate parameter λ in Poisson models for counts). For didactic
simplicity, crime counts of binary crime events are used in the next several pages. The challenge
is to formulate the details of Equation 4 and then properly implement it with data. One must
exploit the associations in time and space.
Figure 2 is an illustration of a smoothed surface of crime fitted values that can be produced
by a detailed elaboration of Equation 4. Forecasting follows easily for any location, such as shown
by the red asterisk. One has already obtained estimates of the parameters for the procedure re-
sponsible for Figure 2. Then, at time t, the crime counts at that location, and perhaps earlier
times, along with crime counts at other locations, become predictor values. They are inserted in
the procedure, and a new fitted value for time t + 1 is obtained; this is just a new fitted value. The
fitted value at time t + 1 is the forecast. Sometimes criminogenic predictors can be used as well.
For example, if one is trying to forecast for time t + 1 homicides and attempted homicides for
location i, counts of the number of aggravated assaults and the local high school dropout rate for
location i at time t might be included for all i when the fitting procedure is fashioned. Ideally,
one can also compute a proper prediction interval at t + 1 to capture uncertainty in the forecast.
There are some new techniques that have enormous promise for properly estimating forecasting
uncertainty (Lei et al. 2018). Other recent work shows that both time and space can be smoothed
at once using kernel-density estimation over those two dimensions (Hu et al. 2018). Forecasting
accuracy may be improved.
A variety of very different methods have been used to flesh out Equation 4. An early treatment
by Groff & La Vigne (2004) provides a good review that requires updating to highlight the differ-
ences between models and algorithms. Figure 3 offers a visual summary of common forecasting
methods organized into three broad categories. The first is model-based methods in which fore-
casts depend on a subject-matter theory of crime generation or a statistical, stochastic model. The
[Link] • Criminal Justice and AI 217
 Data
analysis

Model- Synopsis- Algorithm-

based based based

Y|X X Y|X Y Y|X X

Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

e.g., e.g., e.g., e.g., e.g., e.g.,

multiple factor risk-terrain kernel-density boosting autoencoder
regression analysis modeling estimation

Figure 3
Methods of data analysis used by law enforcement and researchers for predictive policing.

second is synopsis-based methods in which forecasts are based on empirical summaries of crime in
time and space, commonly displayed on various kinds of maps. The organizing goal is explanation,
from which forecasts sometimes follow. The third is algorithm-based methods of the sort that can
be treated as artificial narrow intelligence.
Sometimes particular predictive policing tools can be placed in more than one of the three
boxes, and the literature has a penchant for calling all methods models. For example, in an article
by D’Orsogna & Perc (2015) ambitiously titled “Statistical Physics of Crime: A Review,” every
procedure reviewed is called a model perhaps because, for each, equations are prominent. Such
complications are addressed below when they matter.

Substantive and statistical models. The modeling literature distinguishes between two kinds
of models. The more familiar kind of model is a mathematical statement of processes and re-
lationships for some empirical phenomenon in at least the spirit of Equations 1 and 2. Classical
econometrics, in which one or more algebraic equations are meant to formally represent economic
theory, is the ideal case. “The science of model building consists of a set of tools, most of them
quantitative, which are used to construct and test mathematical representations of portions of the
real world” (Pindyck & Rubinfeld 1981, p. xiii). Causality typically is central (Angrist & Pischke
2009). Such models can then be used for forecasting, applying what are essentially extrapolations
(Pindyck & Rubinfeld 1981).
The econometric perspective has been common in criminology for several decades (Tarling
& Perry 1981). Equation 4 with added details becomes an algebraic theory of some crime gen-
eration processes. Models for pooled cross-section time-series data might then be appropriate
(Cameron & Trivedi 2005), and there is recent work in econometrics focusing on crime predic-
tion (Liesenfeld et al. 2017). The credibility of all such work depends on fulfilling the first- and
second-order conditions discussed above.
A model does not have to literally represent how the world works. Rather than using a sub-
stantive model, a statistical model can be employed. The statistical model must only capture
the stochastic process responsible for generating the data. There needs to be no accompany-
ing subject-matter account. Consider a very simple integrated autoregressive moving average
(ARIMA) model in time (only) with autoregressive and moving average terms but no differencing:
Yt = πYt−1 + εt + φεt−1 , 5.
where Yt is some response variable one seeks to forecast, π and φ are weights much like regression
coefficients, and ε(.) is an IID (mean-zero) disturbance just as in Equation 2. Equation 5 is called

218 Berk
 a model, although there rarely is any substantial subject-matter content. If, in principle, Yt does
not increase over time without limit (which places constraints on the weights), Equation 5 can be
used to fit the observed values of Yt . Ideally, the residuals become statistically indistinguishable
from white noise; all the structure in Yt has been removed. By that criterion, the model is correct.
Then, a one-step-ahead forecast to time t + 1, for instance, is just πYt . Proper prediction intervals
can follow.
Much richer versions of ARIMA models may be specified, but they still have little or no subject-
matter content (Box et al. 2016). There are no first- and second-order conditions as there are
for models meant to capture how nature functions. But for valid use, statistical models also have
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

assumptions that must be met. Equation 5 requires that the disturbances are really IID.
There are applications that fall between substantive models and statistical models. Sometimes a
statistical model is broadly justified by subject-matter reasoning. For example, Aldor-Noiman and
her colleagues (2016) offer a very sophisticated Bayesian, nonparametric, modeling approach for
crime data with very low counts. Especially when the spatial and temporal units are small, many
crime counts are 0 or integers of 2 or less. Most standard statistical methods will stumble.
Predictive policing forecasts are often a product of self-exciting point process models, first
exposited for crime forecasting applications by Mohler and his colleagues (Mohler et al. 2011).
Building on ideas from seismology (Helmstetter et al. 2003), the strength of earthquake after-
shocks is greater if an earthquake is of greater magnitude at the epicenter. The strength of the
aftershocks declines over time and with Euclidian distance from the epicenter. Similar notions are
applied to crimes that are likely to be serial. Insofar as one crime begets another, the chances of
that subsequent crime decline with time and distance from the ground-zero crime. This spatial-
temporal process can be combined with a time-stationary component for each location akin to a
hot spot, which can vary over space.
Point process models are a generalization of Poisson processes on a line to Poisson processes
on a plane and in practice are a statistical model for counts in time and space. There are important
assumptions. For example, the Poisson rate parameter is a deterministic function of spatial location
and time, and the background crime events are themselves independent of one another (Mohler
et al. 2011). Causal interpretations can be misleading because sequential connections between
crimes may not be directly causal but might result from a confluence of similar forcing factors (e.g.,
burglaries committed independently by different groups of individuals after school near bus stops).
Functional forms for the spatial and temporal declines usually are not specified by the statistical
model but estimated with kernel-density procedures (Mohler et al. 2011). They come with no
statistical guarantees. In practice, forecasting accuracy apparently can be better than predictions
derived solely from identified hot spots, but that is not a mathematical requirement and depends
on the spatial and temporal units of data being analyzed (Anselin et al. 2000), how distance is
measured (Rosser et al. 2017), and even the shape of the spatial units.
The application of point processes to crime forecasting has been extended in several ways.
Mohler (2014) predicts homicides and uses a marked point process model with other crime types
in given locations as leading indicators. Reinhart & Greenhouse (2018) characterize locations
with covariates thought to be related to crime. Rosser & Cheng (2016) show that direction as
well as distance can matter, and removing the role of direction appears to eliminate important
inaccuracies in the crime data. Flaxman and colleagues (2019) generalize point processes to a Cox
process model while using regression kernels in a manner that is related to kriging (Cressie 1993).
Mohler and colleagues (2015) conducted a randomized experiment in Los Angeles showing that
their approach provides more accurate forecasts of burglaries than forecasts made by Los Angeles
Police Department crime analysts using hot-spot maps.

[Link] • Criminal Justice and AI 219

 In summary, models of various kinds have to date dominated modern crime forecasting. Vari-
ants of point processes have had some demonstrable success and gained many accomplished advo-
cates. However, a close examination of the work makes clear that the models are statistical and in
much the same spirit as ARIMA approaches that include exogenous predictors (Box et al. 2016).
Statistical formulations fall in the gray area between substantive models and algorithms and, at
least from a computer science perspective, are usually not included under the rubric of artificial
narrow intelligence. One also must treat accuracy claims cautiously. Such claims rest on the per-
formance of intricate, assumption-driven models with many features of convenience or novelty
and, ultimately, all achievements are data-set dependent. No accuracy claims to date are derived
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

from mathematical necessities.

Although far less salient, there are also modeling tools applied to the predictors alone. Perhaps
the most common in crime forecasting are latent variable models of predictors formulated as a
factor analysis. That is,

X j = α0 j + α1 j U + ε j , j ∈ P. 6.

U is a latent variable, ε j is mean-zero IID disturbance term, and X j is the jth predictor from
a set of P predictors (Hastie et al. 2009). Latent variables, seen as causes of observed variables,
can be used in an otherwise conventional regression formulation. An important by-product can
be dimension reduction of the predictor space.

Empirical summaries in time and space. Figure 3 shows methods that rely on instructive sum-
maries of data that at least implicitly support forecasting (Rey et al. 2012). Commonly, maps play
a key role. Risk-terrain modeling, for instance, builds on the spatial and temporal crime patterns
(Bowers et al. 2004, Caplan et al. 2010, Johnson et al. 2007, Ratcliffe 2014). Measures of crime are
overlaid on mapping layers of criminogenic predictors. As Caplan & Kennedy (2016) describe the
approach (see also Caplan 2011), predictor variables can be binary, empirical densities, or distance
metrics from locations of concern. One might notice, for instance, that drug transactions are more
common the closer one gets to homeless encampments. The location of homeless concentrations
might, therefore, be a map overlay coded simply as a binary presence or absence.
Predictors are given weights that indicate the strength of their relationships with crime by
time and place. These weights are, by and large, based on empirical associations, although the step
from a measure of association to a weight is typically ad hoc. Statistical tests often are used as well
but with no formal rationale. When forecasts are needed, the weighted predictors are aggregated
such that higher aggregate values predict more crime. Apparently, seasonal factors frequently are
overlooked; usefully accurate prediction for the summer months may stumble during the winter
(Szkola et al. 2019). Uncertainty in the projections is not addressed in part because it presents
difficult statistical challenges.
There are interesting variations in mapping-based methods. Rosser and her colleagues (2017)
use network distance to measure the proximity between locations. Rather than using as-the-crow-
flies distance, distance is the sum of street segments required to get from one location to another.
Transit routes are also important. A next step might be to formally use Taxicab distance to measure
proximity (Malkevitch 2007), which is closely related to Manhattan distance (Munoz et al. 2009).
Mapping methods also lead to tools for working with X alone. Often, a critical question is
how to spatially aggregate values of predictors. Suppose one is trying to characterize a particular
neighborhood by the number of bars. People in that neighborhood may patronize bars in adjacent
neighborhoods as well as their own neighborhood, especially if they live near a neighborhood
boundary. Yet the farther a bar is from the neighborhood, the less likely it is that people will
make the trip. A statistical tool such as a kernel-density smoother can be used to characterize

220 Berk
 the role of that distance. Then, the total number of bars frequented by residents of a particular
neighborhood might be a weighted sum with weights derived from the kernel-density smoother:
more distance, less weight.
For this discussion, synopsis methods are ways to summarize spatial and temporal patterns in
the spirit of all summary statistics. Despite some naming conventions, there are no models; there
is no attempt to formally represent how the data were generated. One might argue that synopsis
methods are algorithms. There can be some overlap for procedures such as kernel-density
smoothers, but mapping procedures are probably best understood as visualization techniques in
which forecasting, at least to date, is an afterthought. Their main contribution is explanation, not
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

prediction.

Algorithmic methods. Figure 3 shows algorithmic methods that are likely to be considered arti-
ficial narrow intelligence. The applications to crime forecasting are very recent and relatively few.
All would likely be seen as forms of supervised learning because the training data would contain
values for predictors and one or more response variables. It is called supervised because the fitting
process capitalizes on the known values of the response. The response variables are usually crime
counts or crime events.
Supervised learning uses the same kind of data employed in crime forecasting models. But the
functions linking inputs to outputs are not specified in advance, as they largely are in modeling
approaches. The functions linking inputs to outputs are discovered by an algorithm in a highly
inductive manner that computer scientists like to call learning. That learning requires minimizing
some loss function, just as model-based methods do, but balancing bias against variance; the goal
is to come as close as possible on average over hypothetical realizations of the data to the true
relationship between inputs and outputs.
Bias refers to the difference between the expectation of the fitted values and the true response
surface. Variance refers to the variability in the fitted values over new realizations of the data.
With more complex fitted relationships between inputs and outputs, there is less bias but more
variance. With less complex fitted relationships between inputs and outputs, there is more bias
but less variance. Because ML algorithms figure out how to make these tradeoffs, they live nicely
within the rubric of artificial narrow intelligence. All this really means is that the human-imposed
structure that models force on the fitting process is replaced by computational procedures that
extract structure from the data.
There is nothing mysterious in play. Consider stepwise regression, which can be seen as an
early and clumsy form of ML. In its backward elimination implementation, stepwise regression
begins with the largest linear model that the data analyst specifies. In the first step, the procedure
(a) deletes a predictor, (b) stores how much model performance degrades (e.g., the reduction in
the adjusted R2 ), (c) puts the predictor back in the model, and (d) repeats the assessment with each
predictor in the full model. The predictor that degrades performance the least is deleted from
the model, and the evaluation process is repeated with those that remain. With two variables now
deleted, the evaluation process moves on with the predictors not yet eliminated. The winnow-
ing continues until dropping any remaining predictors degrades model performance too severely.
That smaller model is the one that is used for the data analysis. Given the full model, an algorithm,
not a human, determines the best model.
Modern ML is a far more powerful way to proceed and addresses far more than variable se-
lection. Among other accomplishments, it can construct new predictors from old ones. There is
generally nothing like a model to constrain how the algorithm proceeds, and the search for a good
fit allows for a wide variety of relationships to be explored, many of which can be highly nonlinear.

[Link] • Criminal Justice and AI 221

 f (x) = g (T )
Y

T = β0 + β T Z

Z1 Z2 ...
ZM

Zm = σ (α 0m + αTm X )
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

X1 X2 X3 X4 ... XP

Figure 4
A simple deep-learning neural network with one hidden layer composed of M nodes [adapted from Berk
(2020b) and Hastie et al. (2009)].

There are also a variety of safeguards to prevent problems before they appear or fix them after.
For example, a form of regularization is commonly used to help protect against overfitting.
There are a large number of popular ML methods: random forests, gradient boosting, support
vector machines, and others. Crime forecasting is a relatively recent application, which so far has
been dominated by various forms of neural networks, often characterized as deep learning. The
details are too lengthy to specify here and can vary depending on the kind of deep learning in play
(Goodfellow et al. 2016).
Figure 4 shows a very simple example in which deep applies only as a formality. There are x1
through x p inputs, called predictors in statistics. Each is connected to all the nodes z1 through zm
of a single hidden layer, called latent variables in statistics. The inputs are combined in a linear
manner and then transformed by an activation function, σ . The linear combinations serve as regu-
larizers, although there are several other regularization tools that are commonly used (Goodfellow
et al. 2016). Activation functions play a similar role to link functions for the generalized linear
model (e.g., in logistic regression). They introduce nonlinearities that for deep neural networks
are compounded over hidden layers. Usually, there are at least several hidden layers and often
many more. Highly nonlinear functions can be constructed.
The transformed linear combinations populate each node in the hidden layer. These trans-
formed linear combinations are then combined linearly and can be transformed by g to provide
fitted values for the output Y. Y may be numeric or categorical, which for predictive policing will
usually be crime counts or (binary) crime events. There can be more than one Y included as out-
puts for the same network (e.g., crime numbers and victim numbers). In some applications, there
are hundreds of inputs, hundreds of hidden layers, and, consequently, thousands of parameters to
be estimated. They play roles similar to the αs and βs in Figure 4. Computational challenges can
be daunting, in part because the loss function is not convex. But once the loss is minimized, one
has an algorithmic structure that can be used for forecasting when new inputs are introduced and
the output is unknown.
There are many different flavors of deep learning. Convolution neural networks (CNNs), for
example, were developed to classify images (e.g., in facial recognition) and can be used for spatial
data; a map is an image. The convolution processes can be seen as part of data preprocessing that
extracts the essential features of the image to be passed along as predictors to a neural network
(Goodfellow et al. 2016). Recurrent neural networks (RNNs) have been developed for pooled
cross-section data (Goodfellow et al. 2016) so that the temporal structure can be exploited.

222 Berk
 Deep output RNN
Yt–1 Yt Yt+1

Vanilla RNN
H H H
Yt–1 Yt Yt+1

Ht–1 Ht Ht+1 Ht–1 Ht Ht+1

Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

X t–1 Xt Xt+1 X t–1 Xt Xt+1

Stacked RNN
Yt–1 Yt Yt+1

Deep transition RNN

Ht–1 Ht Ht+1
Yt–1 Yt Yt+1

Ht–1 H Ht H Ht+1 Ht–1 Ht Ht+1

X t–1 Xt Xt+1 X t–1 Xt Xt+1

Figure 5
Examples of deep recurring neural network (RNN) structures. Red ellipses represent additional hidden layers beyond those in the
vanilla RNNs.

Generative adversarial networks (GNNs) can be used as target hardeners for algorithms
vulnerable to hacking (Krohn 2019).
Figure 5, drawing heavily on work by Pascanu and colleagues (2014), shows four different
RNN structures that differ in complexity and structure. Each vertical slice has a structure ab-
stracted from Figure 4. There is a set of inputs (Xt ), a single hidden layer (Ht ), and a single output
(Yt ). RNNs allow for one such structure at each time point. Conventionally, they share the as-
sumption that dependence over time in the response Yt is solely a result of associations between
the hidden layers over time. Hidden Markov models have the same rationale (Zucchini et al. 2016).
In Figure 5, the hidden layers in red ellipses introduce additional nonlinearities at particular
points in the network. A greater number of hidden layers generally implies more complex fitted
values, and their locations reveal where the additional complexity is needed. One can think of
RNNs as nonlinear versions of analysis techniques for cross-sections over time.
Recent papers introduce powerful deep-learning procedures to crime forecasting (Wang et al.
2017a,b). They use the same kinds of data as employed by the more complex modeling approaches
discussed earlier but are able to inductively allow for a large number of nonlinear relationships.
Zhang & Cheng (2020) developed a variant of deep learning that exploits a network distance met-
ric rather than Euclidian distance as an elaboration of RNNs for forecasting crime. Zhuang and
colleagues (Zhuang et al. 2017) used a creative extension of RNNs that includes a spatial compo-
nent to forecast crime hot spots defined by calls for service. The output predicted is categorical
(i.e., hot spot or not), not a crime count. They claim that their approach forecasts more accurately
than other deep-learning methods and traditional ML classifiers.
Finally, just as for other predictive policing methods, there are procedures for predictors
alone. Some are a rebranding of statistical tools developed more than 50 years ago, typically

[Link] • Criminal Justice and AI 223

 as unsupervised learning. Clustering and multidimensional scaling are examples. Autoencoder
(Goodfellow et al. 2016) is a new tool that can be seen as a nonlinear version of principal com-
ponent analysis. It too can be used for dimension reduction but also for a number of novel tasks
like denoising data. For example, all handwritten offense reports should have a date that a police
officer fills in. But different police officers will often use different date conventions (e.g., 1/6/20
versus 01/06/2020) and sometimes have unreliable penmanship. An autoencoder can be given a
large number of inputs that are the dates as written and also the outputs correctly determined by
humans. From these data, the algorithm can learn how to reliably go from the handwritten dates
to the correct dates so that in future, handwritten dates can be automatically and accurately read
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

by a machine or recoded to a consistent format for data processing.

Criticisms of Predictive Policing

Consider again Figure 1. Much of the pushback on predictive police is less about artificial intel-
ligence and more about the activities surrounding it. Ratcliffe & McCullagh (2001) focused on
the dissemination and utilization of crime forecasting information. They discovered that the road
from forecasts to implementation can be rocky. A lot depends on how the information is transmit-
ted to patrol officers and how it conforms to existing police officer priorities. Accurate forecasts
alone are no guarantee of improved policing performance.
Bennett Moses & Chan (2018) write that police agencies are purchasing predictive policing
software when evidence of forecasting accuracy is spotty and that, among many other things, al-
gorithms may misdirect oversight concerns from police officers and their superiors to technical
details of some forecasting algorithm. In addition, the data on which algorithms are trained can
be badly flawed because crimes reported is not the same as crimes committed and reported crimes
may be mischaracterized. There are also concerns that disadvantaged neighborhoods will be over-
policed, which can be seen as a form of racial bias.
Kaufmann and colleagues (2019) seek to go beyond the data and forecasting methods to the
different ways in which the empirical patterns are translated into larger formulations about the
nature of crime and the societal systemic responses. For example, a focus on crime forecasts leads
to a societal emphasis on preemptive policing that can become an indictment of neighborhoods
that are already disadvantaged. In his important book The Rise of Big-Data Policing: Surveillance,
Race, and the Future of Law Enforcement, Ferguson (2017) considers a wider range of surveillance
practices in the context of US constitutional protections and the growing role of the surveillance
state.
There are several responses to the concerns raised. First, much of the uneasiness can be legit-
imate, but typically has little to do with artificial intelligence itself. In that sense, it is beyond the
scope of our discussion.
Second, the baseline for forecasting accuracy is current practice, not perfection. The guiding
question is whether predictive policing leads to improvements in accuracy beyond current proce-
dures. There have been some successes, but so far we have no general conclusions one way or the
other.
Third, it is surely true that much of the information used for algorithm training is shaped by
historical patterns. But does that information lead to substantial racial bias in the forecasts made?
Police patrols have long been substantially directed by calls for service. These come dispropor-
tionately from disadvantaged neighborhoods. It follows that police can be more densely present in
some neighborhoods than others. In turn, for some neighborhoods, this can lead to more official
reports of crimes, more apprehensions, and more arrests. But those are not necessarily a product of
racial animus or racially motivated administrative practices. Police merely are concentrated where

224 Berk
 they are summoned. Furthermore, because clearance rates are generally lower in disadvantaged
neighborhoods, rap sheets for many of the neighborhood residents may actually underestimate
criminal activities.
Finally, what about the crime victims? There is no debate that crime victims and crime perpe-
trators disproportionately live in disadvantaged neighborhoods. For example, in Philadelphia last
year, there were 361 homicide victims. Most were African-Americans from low-income parts of
the city. The same was true of most perpetrators. If one cares about crime victims, law enforcement
is surely part of the way cities should respond even though that can mean arresting perpetrators
from disadvantaged neighborhoods. Yet most critiques of predictive policing give crime victims
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

no voice.

RISK ASSESSMENT
Because predictive policing and risk assessments are different, so are the data they use, the fore-
casting procedures they employ, and the jurisprudential issues raised. Predictive policing forecasts
crimes by place and time. Risk assessment forecasts crimes to be committed by particular individ-
uals without much regard to where or when. An actuarial approach naturally follows. The kinds
of people who will commit crimes in the future are like those who have done so in the past. A
key challenge, therefore, is to identify features of past offenders that are strongly associated with
criminal behavior. This is little different from forecasting which individuals are likely to suffer
from heart disease, which applicants for a job are unlikely to perform well, or which applicants
for a mortgage are prone to default on the loan. Often overlooked, however, is that the alternative
to a failure is a success; when individuals are not classified as high risk, they can be automatically
classified as low risk. Good things can follow. For example, at an arraignment, an individual may
be diverted into a drug treatment program.
In the United States, criminal justice risk assessment has a very long history. Parole decisions
were in part informed by risk starting in the 1920s (Burgess 1928), and the use of artificial narrow
intelligence in law enforcement has been an important component of risk assessment long be-
fore it was introduced into predictive policing. For example, Berk and colleagues (2005), building
on work done for the California Department of Corrections in the 1990s, used ML to develop
intimate partner violence forecasting tools for the Los Angeles County Sheriff’s Department.
Risk assessment is now commonly employed to inform a variety of criminal justice decisions:
release decisions at arraignment, post-conviction sentencing decisions, assignments to security
levels in prison, release decisions by parole boards, and the intensity and nature of supervision on
probation and parole (Berk 2018). There are a few jurisdictions in which police use risk assessments
and other approaches to identify particular individuals at high risk of being shot or being a shooter
(Hollywood et al. 2019).
Over time, quantitative risk assessment has gone from simple scales computed using craft lore
and bivariate associations (e.g., between age and crime) to various kinds of traditional regression
analysis (e.g., logistic regression) to modern ML. The main rationale for these changes has been
demonstrable improvements in forecasting accuracy (Berk & Bleich 2013). The early scales were
not only crude by today’s standards but also not properly evaluated (Ohlin & Duncan 1949, Reiss
1951). It was extremely difficult to determine how well they forecasted. Regression models clearly
have been an improvement and are still commonly used, but they are flawed by modern statistical
standards. For example, because the response variable in logistic regression is the logarithm of the
odds of some binary outcome (e.g., rearrest), a regression coefficient shows the average change in
the logit of Y for a unit change in a given predictor, adjusted for its linear dependence on all other
predictors. Yet criminal justice officials want to forecast the chances (i.e., probability) of rearrest.
Altering the coefficient for the probability of a particular outcome means employing a nonlinear

[Link] • Criminal Justice and AI 225

 transformation from which there is no longer a single coefficient to interpret. It is, therefore, not
clear how to properly construct interpretable weights to assign the predictors when a forecast is
needed.
Recalling Figure 1, risk assessment methods as well as predictive policing methods should be
distinguished from the activities surrounding them. With this done, ML as a form of artificial
narrow intelligence is proving to be superior to regression models along several dimensions when
conducting criminal justice risk assessments.

 Regression models must meet a variety of formal assumptions that usually are implausible
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

in practice (Freedman 2009). ML methods are not constrained in this manner (Hastie et al.
2009).
 ML forecasts as accurately as conventional regression models when inputs are related to
outputs in a simple manner but substantially better when inputs are related to outputs in a
complex manner, the predictors have distributions that can map well to that complexity, and
the ML procedures are properly tuned (Berk 2018, Berk & Bleich 2013, Berk & Sorenson
2020). This point is elaborated on below.
 Criminal justice decisions based on risk assessment are typically categorical (e.g., detain at
an arraignment or not). This means that in the end, assessments of risk should be categorical
(e.g., high risk or low risk). For regression models, this is typically done, if at all, in an ad
hoc manner. Some ML procedures automatically forecast outcome classes in a thoroughly
principled manner and most others can be easily altered to do so (Berk 2018).
 Policymakers are of late less concerned about binary outcomes, e.g., recidivate or not, and
more concerned with forecasting with three or even four outcome classes. In the case of
domestic violence, for instance, at an arraignment one usefully can forecast one of three
outcomes: no new reported domestic violence incidents, reported incidents not involving
injuries, or reported incidents in which injuries result (Berk & Sorenson 2016). ML can
easily accommodate outcomes with more than two categories (Berk 2018).
 Regression models treat the costs of false positives as the same as the costs of false negatives.
ML procedures can easily incorporate asymmetric costs of forecasting errors. This can make
forecasts more responsive to stakeholder needs (Berk 2020b).
 There is substantial sophistication and improvements in how to address forecasting unfair-
ness when ML forecasting methods are used (Berk et al. 2018, 2020a). Regression models
have yet to effectively address issues of racial bias.

Risk Assessment Machine Learning Methods

Perhaps the two most important and effective ML methods used for criminal justice risk assess-
ment are random forests (Breiman 2001a) and stochastic gradient boosting (Friedman 2002). They
share a strategy of constructing an ensemble of fitted values from classification or regression trees
(Breiman et al. 1984) and then combining them in a sensible fashion for purposes of regularization.
Figure 6 is meant to convey the general idea.
The top row represents M passes over the data, each time addressing the conditional distribu-
tion of Y at time t given a set of predictors X also at time t. The setting might be a post-conviction
sentencing decision. Y could be an arrest for a violent crime and X might be predictors such as age,
prior record, and the current crime of conviction. The asterisks are meant to convey that one is not
working with the original data in the form it was generated. Rather, one is working with a random
sample with replacement, sometimes called a bootstrap sample, or with reweighted observations
as described earlier. The M samples being analyzed can differ in other ways as well.

226 Berk
 Ŷt*|X*1t Ŷt*|X*2t ... Ŷt*|X*Mt

Ŷt+1
** |Xt+1
**

Figure 6
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Stylized ensemble methods for machine learning.

Here are the algorithmic steps for random forests (Berk 2020b). Let N be the number of ob-
servations in the training data and assume for now that the response variable is binary.

1. Take a random sample of size N with replacement from the data.

2. Take a random sample without replacement of the predictors.
3. Construct the first recursive partition for a classification tree of the data as usual.
4. Repeat Step 2 for each subsequent split until the tree is as large as desired. Often this leads
to one observation in each terminal node. Do not prune. Compute each terminal node
proportion as usual.
5. Drop the out-of-bag (OOB) data down the tree. The OOB data were not used to grow
the tree because they were by chance excluded when the random sample of the data was
drawn with replacement. Save the class assigned to each OOB observation along with each
observation’s predictor values. The OOB data help reduce overfitting because they were not
used for training,
6. Repeat steps 1–5 many times (e.g., 500).
7. Using only the class assigned to each observation when that observation is OOB, count the
number of times over different trees that the observation is classified in one category and
the number of times over different trees it is classified in the other category.
8. Assign each case to an outcome category by a majority vote over the set of trees when that
case is OOB. Thus, if, over a large number of trees, 51% of the time or more a given case is
classified as a 1, then that becomes its assigned class. If, over a large number of trees, 51%
of the time or more a given case is classified as a 0, then that becomes its assigned class.
9. If forecasts are needed, the new IID cases from the same joint probability distribution are
dropped down each tree and depending on the terminal node in which a case falls, a class
is assigned as usual. There is then a vote over different trees through which the winning
class becomes the forecasted class. For numeric response variables or if forecasted condi-
tional proportions are desired, one simply averages the terminal node conditional means or
conditional proportions for each case. These averages serve as forecasts.

Stochastic gradient boosting can also be sequentially summarized using a binary response vari-
able here for simplicity (Berk 2020b).

1. There is a binary response variable in the training data coded as 1 or 0. The algorithm is
initialized with some constant such as the overall proportion of 1s. This constant serves as
the fitted values from which residuals are obtained by subtraction in the usual way. The
residuals are then appended to the training data as a new variable.
2. A random sample of the data is drawn without replacement. One might, for example, sample
half the data. Formally, the sampling is not necessary, but it seems to help performance.

[Link] • Criminal Justice and AI 227

 3. Because the residuals are numeric, a regression tree, not a classification tree, is applied to
the sample with the residuals as the response. Another set of fitted values is obtained and
used to update the prior set of fitted values. From these, a new set of residuals is obtained
and appended. In effect, the residuals serve as weights such that observations that are less
well fit in one pass through the data are made more important for the next pass through the
data.
4. Another random sample is taken from the training data, and the fitting process is repeated.
As before, the sampling is not formally necessary.
5. The entire cycle is repeated many times: (a) fitted values, (b) residuals, (c) sampling, and
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

(d) a regression tree. In the end, the fitted values from each pass through the data have
been combined in a linear fashion. For classification, these aggregated fitted values can be
interpreted as probabilities.
6. Commonly, observations with ŷi > 0.5 are assigned a 1 and observations with ŷi ≤ 0.5 are
assigned a 0. Numeric response variables are also permitted. The steps are largely the same
but without an assigned class.
One central point is that weighting, which is so fundamental to boosting, occurs implicitly
through the residuals from each pass. Larger positive or negative residuals imply that for those
observations, the fitted values perform less well. As each regression tree attempts to maximize its
fit, the tree responds more to the observations that in the previous iteration had larger positive or
negative residuals.
Ensemble procedures can forecast very well. One important reason is that each decision tree
can produce a very complex set of fitted values that is regularized by combining the fitted values
over different trees. One can arrive at a good balance between the bias and the variance. Other
reasons are discussed elsewhere (Berk 2020b).

Accuracy, Fairness, and Transparency

Almost regardless of the type of artificial narrow intelligence or the applications, there has been
a range of concerns about matters that go well beyond the technology. Questions about accuracy
address the algorithmic fitting and forecasting errors. Fairness involves treating similarly situated
groups similarly. Transparency refers to the accessibility of ML technology to stakeholders.

Accuracy. Accuracy is usually treated in reverse: How often does the ML algorithm get it wrong?
For criminal justice risk assessment, there are three popular definitions of error. The first is the
overall fraction of cases for which the actual outcome class differs from the assigned outcome
class. Given the known outcome classes in the training and/or test data, what proportion of the
time does the algorithm identify them incorrectly? Sometimes this is interpreted as the overall
probability of a classification error.
An important problem with the overall measure of error is that the costs of forecasting errors
usually differ depending on the outcome class. One way to address this problem is to use the
available data to compute an error rate separately for each observed outcome class. For example,
given the subset of individuals who actually failed on parole, what proportion of the time does the
algorithm not identify them correctly? Likewise, given the subset of individuals who actually did
not fail on parole, what proportion of the time does the algorithm not identify them correctly?
Sometimes these proportions are treated as if they are probabilities.
If failure on parole is a negative, the first is called the false-positive rate. If success on parole is
a positive, the second is called the false-negative rate. The logic applies to more than two outcome

228 Berk
 classes, but these naming conventions become unclear. If one outcome class is called a positive and
another outcome class is called a negative, what names are attached to the other outcome classes?
In practice, an individual’s outcome class is not known when a forecast is required (or there
would be no need for a forecast). The first three definitions condition on the known outcome
class and are used to measure algorithmic performance. In practice, more interest targets fore-
casting accuracy. Using test data, one conditions on the forecasted outcome class and ascertains
the proportion of times the forecasted outcome class is wrong. Typically, forecasting error is com-
puted separately for each outcome class, which is often interpreted as probabilities.
Because most risk algorithms minimize a loss function, anything that constrains that optimiza-
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

tion reduces accuracy (Kearns et al. 2018a). There is a mathematically required price. Constraints
to improve fairness (see below) are a common instance. For example, at arraignments, more of-
fenders are detained who pose no risk to public safety, and more offenders are released who pose
a threat to public safety.
Practitioners are sometimes flummoxed when ML risk accuracy is on occasion little different
from that of conventional regression models. However, any conclusions about relative perfor-
mance need to be based on a solid understanding of how ML risk forecasts are made.

 Sometimes the high-risk offenders and the low-risk offenders are patently obvious. Most risk
assessments, even just a casual reading of a presentencing report, agree on which is which.
The real challenge is accurately classifying offenders in the middle. Recent work (Berk 2017)
suggests that for parole board decisions, ML earns its keep by correctly classifying the equiv-
ocal cases. Inmates prone to violence and those likely to become model citizens are apparent
to all. In other words, most any of the more common risk tools would perform reasonably
well over the full range of inmates, and that is often how performance comparisons are made.
Narrowly crafted accuracy assessments can be much more instructive.
 Suppose an ML procedure estimates with a probability of 0.95 that a given offender will be
rearrested. A linear model estimates that probability as 0.55. These are very different prob-
abilities of risk. Yet both classify the offender as high risk because both probabilities exceed
a threshold of 0.50. Classification, therefore, masks important performance differences, and,
along with a forecast, the numerical values that determined the forecast should be provided.
For example, measures like the vote from random forests should be coupled with each fore-
casted outcome class (even though those votes are not probabilities). The larger the fraction
of the vote is for a winning outcome class, the more reliable the forecast; out-of-sample per-
formance is generally better. Such performance differences will not be apparent from the
forecasted class alone.
 When the true response surface is relatively simple, more complex sets of fitted values usu-
ally do not improve performance much. When the true response surface is complex, one
needs complex fitted values. But a complex set of fitted values depends substantially on the
quality of the predictors. If they are too weak to capture complex surfaces, ML may not per-
form better than conventional regression. For example, if a binary outcome has a strongly
unbalanced distribution (e.g., the probability of a rearrest for a violent crime is only 0.05),
having at least some relevant predictors with long right or left tails can dramatically help.
 In part because the primary outputs from ML procedures are sets of fitted values, and there
is no pretense of a formally correct model, the bias–variance tradeoff becomes essential.
When the true response surface is simple, the bias–variance tradeoff matters little. There
is no need to trade variance for bias because making the fitted values more complex is not
typically helpful. A variety of regression-like procedures perform equally well. When the

[Link] • Criminal Justice and AI 229

 true response surface is complex, the bias–variance tradeoff can be usefully exploited, which
is precisely what effective ML does and conventional regression does not.
 Finally, performance depends on how an ML algorithm is tuned. Often comparable perfor-
mance with conventional procedures comes from faulty tuning. Relying on the default values
of the tuning parameters, for instance, is often a serious error that cripples risk algorithms. It
is also unwise to rely exclusively on automated fitting criteria such as cross-validation. Many
ML risk algorithms are sample-size dependent, and the folds required by cross-validation
necessarily reduce the effective sample size. Accuracy can suffer.
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Fairness. The analysis of fairness for criminal justice risk assessment is much more developed than
the analysis of fairness for predictive policing. Many different academic disciplines have weighed
in (Berk et al. 2018, Corbett-Davies & Goel 2018, Goel et al. 2019, Hamilton 2016, Huq 2019,
Kearns et al. 2018b, Kleinberg et al. 2017, Mayson 2019). In criminal justice settings, fairness
represents equal treatment across groups. There is a related approach in which fairness is defined
as equal treatment across people, which has been used in other kinds of applications such as privacy
(Dwork & Roth 2014).
There are often direct links between common measures of accuracy and different definitions
of fairness.

 In equality of outcome, each different group is represented in equal proportions for some
outcome. For example, the same proportion of men and women are detained at an arraign-
ment; e.g., 26% of the men and 26% of the women are detained.
 In equality of opportunity, for each outcome class, every group is classified with the same ac-
curacy. For two outcomes, this means, for example, the same false-positive and false-negative
rates for blacks and whites. The classifier works equally well for both races.
 In equal prediction accuracy, every group’s outcome class is forecasted with equal accuracy.
For example, individuals of Hispanic descent and individuals of Chinese descent have fore-
casts of rearrest that are correct 65% of the time and forecasts of no rearrest that are correct
72% of the time.
 In equality of treatment, the cost ratio of false positives to false negatives is the same for ev-
ery group. For instance, there are three false positives for every false negative for American
citizens born in the United States and American citizens born in Somalia. Equal treatment
is implicated because the cost ratio indicates how the algorithm is trading one kind of clas-
sification error for another. In this 3 to 1 illustration, false negatives are three times more
costly than false positives; thus, the algorithm works harder to avoid them.

Unfortunately, you can’t have it all. Not only is accuracy affected by efforts to improve fairness,
but there are often tradeoffs between different kinds of fairness. In particular, if “the base rates
differ across groups, any [risk] instrument that satisfies predictive parity at a given threshold. . .must
have imbalanced false positive or false negative errors rates at that threshold” (Chouldechova 2017,
p. 5; emphasis in original).
A base rate is the distribution of the outcome classes. For example, if the proportion of men
who fail to appear (FTA) when required to return to court is 0.65, the FTA base rate is 0.65. If
the proportion of women who fail to appear when required to return to court is 0.55, the FTA
base rate is 0.55. The threshold refers to some cutoff on a risk score such that those offenders
who are above that score have a particular forecasted outcome class. An estimated probability of
failing a drug test greater than 0.75 might be used to forecast failing a drug test on parole. The
term predictive parity means the same as equal forecasting accuracy.

230 Berk
 The message is that given different base rates, there is an inevitable tradeoff between equal
false-positive or equal false-negative rates on one hand and equal forecasting accuracy on the
other. More generally, the several kinds of accuracy and several kinds of fairness require a far
more nuanced discussion than commonly undertaken. Much too often, stakeholders just talk past
one another.
Transparency. ML algorithms are often characterized as a black box. The manner in which inputs
are related to outcome is obscure. At the very least, black-box algorithms can make stakeholders
uneasy.
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

There are several ways to constructively respond. First, the human brain is a black box too.
For example, when a magistrate determines that some individuals are to be detained and others
not, the reasons usually are not apparent. Even if at sentencing a judge’s reasoning is explained,
the content can be very abbreviated, subject to public relations concerns, especially if in that ju-
risdiction judges are elected, and shaped by unconscious bias (Rachlinski et al. 2009). Even more
difficult to comprehend are the decisions by parole boards that are made behind closed doors. The
transparency bar is very low, and algorithms can compete.
Second, there are ways to construct algorithms that are easier to understand. For example,
Zeng and colleagues (2017) develop a linear approach that uses only binary predictors and integer
weights. Each unique value for numeric variables has its own binary variable coded 1 or 0 so that
nonlinear relationships can be computed, and the different units of the original numeric variables
no longer complicate interpretations. The integer weights are also easier to interpret than con-
ventional regression coefficients. A penalized discrete optimization approach is used that seeks
to minimize the number of classification errors and, in the spirit of the lasso (Tibshirani 1996),
produces a sparse set of predictors.
Third, there is a cottage industry of algorithms meant to complement ML procedures that
extract useful information, often using visualizations, about otherwise obscure processes. For ex-
ample, there are several ways in which the contributions to the fitted values can be calculated and
displayed. One simple approach computes the decline in accuracy when each predictor in turn is
precluded from affecting the fitted values (Breiman 2001a). Another approach displays how each
predictor is related to the response holding all other predictors constant without the use of the
standard covariance adjustments employed by linear regression (Hastie et al. 2009).
Fourth, one can often provide answers to specific questions about what is going on under the
hood. For example, one can take a single case and obtain the fitted class. Then for the same case,
the value of some predictor such as age is changed. One can learn whether the fitted class would
have been meaningfully different if the individual were 10 years older, with everything else about
the individual fixed. In this manner, the way the algorithm uses age can be easily demonstrated.
Finally, unlike the human brain, there is code that details exactly how risk is computed. If the
software is not proprietary, the code can be studied. Most stakeholders will not have the skill or
inclination to take on this task, but many universities and business establishments have individuals
who can lend a hand.
In short, ML risk assessment has a public relations problem. The hype surrounding artificial
intelligence encourages science fiction parallels that are very misleading. Unbridled enthusiasm
from some ML proponents reinforces the narrative. Private firms that sell ML software capitalize
on both through slick advertising and well-crafted sales pitches. It should be no surprise that many
stakeholders feel overmatched.

FUTURE DEVELOPMENTS
Over the next decade, likely developments are easy to anticipate. Although the algorithms used
for predictive policing and criminal justice risk assessment will continue to improve, the pace of
[Link] • Criminal Justice and AI 231
 change will slow. The major algorithmic advances probably have been delivered already. What re-
mains for the software is fine-tuning, although new risk outcomes will be predicted. For example,
there is already work in progress to forecast whether an apprehension will lead to a shooting of
either the suspect or any of the arresting officers. The increasing use of employment assessment
algorithms in the private sector will migrate widely to police recruitment, promotion, and disci-
plinary actions. Beyond predictive policing and criminal justice risk assessment, some previously
fanciful kinds of algorithmic applications will grow in importance and experience substantial per-
formance improvements, particularly within robotic devices, as military applications lead the way
(Kessel 2019). There are already piloted drone devices that can search an abandoned building and
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

demonstrations that self-guided quadrotor swarms will soon be able to do the same much more
quickly (Ackerman 2017).
Computer processing power will continue to grow rapidly. Concurrently, these developments
will be provided in smaller housings so that, for instance, hand-held devices will have the ca-
pabilities of today’s large desktop machines. Significant progress is already being made building
computer code into the design of computer chips. The role of the software middle man is fading.
Artificial narrow intelligence will soon be mobile, powerful, and part of the routine equipment
used by police officers on patrol. A wide variety of policing tasks will be affected beyond those
addressed in this review.
Databases will become much larger, more accurate, and dramatically richer. Existing data sets
will be merged so that, for instance, detailed measures of performance on parole will be combined
with data on employment, health, and social media use as well as with similar information on
family and friends. Widespread surveillance data will also be readily available through devices like
CCTV, cell phones, drones, automobile black boxes, police body cameras, and ankle bracelets. The
accuracy and fairness of forecasts could be dramatically improved. Several Scandinavian countries
are already far down this path; for instance, it seems quite possible to forecast with a useful level
of accuracy whether a newborn infant will be arrested for a felony by the time he or she is 20 years
old (Lyngstad et al. 2017).
As with all new technology, there are major risks attached to the significant benefits. Privacy
concerns immediately come to mind. Misuse of the forecasts is also a potential hazard. For ex-
ample, might forecasts of malfeasance be sold to health insurance firms or prospective employ-
ers or used for extortion? More generally, might the capacities developed for law enforcement be
ported to other applications? Suppression of political dissent is an obvious example that seems well
underway in some countries. None of these concerns, and others closely related, are new. Safe-
guards are needed. In the United States, Europe, and elsewhere, political institutions are beginning
to respond. Citizen groups are forming and many NGOs are engaged. There is even counter-
surveillance technology being developed that individuals can obtain. For example, eyeglasses are
now available that diffuse the light from a CCTV camera to prevent facial recognition. But how all
this will play out is unclear, and substantial worry is healthy. “We are playing cat and mouse. . .and
that always ends poorly for the mouse” (Woodrow Hartzog, quoted in Hill 2020, p. 3).
All these developments will be implemented using artificial narrow intelligence applied to ex-
tensions of activities that are already being undertaken or that have been proposed. There is no
need to consider science fiction scenarios. The real world is challenging enough.

SUMMARY AND CONCLUSIONS

Predictive policing and criminal justice risk assessment can be seen as examples of artificial narrow
intelligence that do better what is already being done. Current practice is a very long way from

232 Berk
 artificial general intelligence, let alone artificial superintelligence. Despite some media portrayals,
the singularity is nowhere in sight.
Algorithms are very different from models. Models are mathematical expressions about how
the world works. Algorithms are methods for computation. Concerns that may be central for one
may be irrelevant, or even misleading, for the other. One powerful example is that there is no
such thing as mean function misspecification for algorithms. Another powerful example is that
for algorithms, explanation takes a distant back seat to prediction; accurate forecasting does not
depend on subject-matter understanding.
Concerns about algorithms in criminal justice settings must allow that algorithms by them-
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

selves are rarely the problem. Accuracy and fairness stem far more from the data and how they
are managed. And there can be no material benefit or harm until algorithm-informed decisions
are made and actions are taken. Furthermore, the performance benchmark is not perfection but
current practice. The goal of predictive policing and criminal justice risk assessment is just to do
better.
Predictive policing forecasts crimes in place and time. To date, models have dominated practice
and performance, especially models based on point processes and their extensions. Algorithms
recently have made an appearance, typically as some form of deep learning. There are no definitive,
defensible claims of superior accuracy for any model, algorithm, or synopsis method because all
such claims depend on the setting and the data; all results are setting and data-set specific. Such
dependence is almost universal in organized forecasting competitions as well. There are no formal
mathematical results that preordain a winner. Although the balance of evidence seems to favor
predictive policing compared to existing policing practices, the winning margin is often small and
limited to certain kinds of crimes. Finally, as emphasized earlier, there are no guarantees that police
officers will use the forecasts as intended or even have tactics that can capitalize on them.
Criminal justice risk assessments forecast which individuals will re-offend among those already
being supervised or processed by the criminal justice system. Many legacy risk tools dating back
decades are still in use, but ML classifiers are gaining proponents. There are well understood for-
mal reasons why ML forecasts of risk will be at least as accurate as the model-based competitors
and, with complex relationships between inputs and outputs, can perform better. The use of reg-
ularized sets of complex fitted values offers a superior tradeoff between the bias and the variance.
At the same time, all the more powerful ML procedures offer comparable accuracy because they
are operating with much the same underlying principles. Just as for predictive policing, however,
there are no guarantees that a risk assessment will be properly used or, even if it is, that effective
practices can be implemented.
The mind of a criminal justice decision-maker is a black box, and, relative to that baseline, ML
risk assessments can be more transparent. Relative accuracy is more difficult to ascertain because
there are several kinds. Nevertheless, research in cognitive science, as well as criminal justice,
strongly favors quantitative risk assessment. There are also several different kinds of fairness and
inevitable tradeoffs between them coupled with inevitable tensions between fairness and accu-
racy. You can’t have it all. There can be no technical fix. Legal, political, and legislative processes
must strike an acceptable balance between competing priorities, which is, in any case, the way we
currently do things.
Finally, the future will bring further improvements in predictive policing and criminal justice
risk assessment. The power of these tools will grow dramatically with new computer technology
and far better data sets. This will lead to spillovers into applications outside of criminal justice,
many of which may not be benign. There will also be an influx of high-tech tools from military
applications, and some will pose significant risks ([Link] All this
will be substantially driven by algorithms that embody artificial narrow intelligence.

[Link] • Criminal Justice and AI 233

 DISCLOSURE STATEMENT
The author is not aware of any affiliations, memberships, funding, or financial holdings that might
be perceived as affecting the objectivity of this review.

ACKNOWLEDGMENTS
A special thanks goes to Penn Cary Coglianese, Michael Kearns, and Aaron Roth for helping to
introduce me to many of the issues discussed and to Penn colleagues Andreas Buja, Eric Tchetgen
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Tchetgen, Ed George, Linda Zhao, and Arun Kuchibhotla for hours of technical instruction.

LITERATURE CITED
Ackerman E. 2017. This autonomous swarm doesn’t need GPS. IEEE Spectrum, Dec. 27. [Link]
[Link]/automaton/robotics/drones/this-autonomous-quadrotor-swarm-doesnt-need-gps
Aldor-Noiman S, Brown LD, Fox EB, Stine RA. 2016. Spatio-temporal low count processes with applications
to violent crime events. Stat. Sin. 26:1587–610
Angrist JD, Pischke J-S. 2009. Most Harmless Econometrics. Princeton, NJ: Princeton Univ. Press
Angwin A, Larson J, Mattui S, Kirchner L. 2016. Machine bias. ProPublica, May 23. [Link]
org/article/machine-bias-risk-assessments-in-criminal-sentencing
Anselin L, Cohen J, Cook D, Goor W, Tita G. 2000. Spatial analyses of crime. In Criminal Justice 2000:
Measurement and Analysis of Crime and Justice, ed. R Kaminsk, N La Vigne, pp. 213–62. Washington, DC:
Natl. Inst. Justice
Arute F, Arya K, Babbuch R, Bacon D, Bardin JC, et al. 2019. Quantum supremacy using a programmable
superconducting processor. Nature 575:505–10
Bennett Moses L, Chan J. 2018. Algorithmic prediction in policing: assumptions, evaluation, and accountabil-
ity. Polic. Soc. 28(7):806–22
Berk RA. 2017. An impact assessment of machine learning risk forecasts on parole board decisions and recidi-
vism. J. Exp. Criminol. 13:193–216
Berk RA. 2018. Machine Learning Forecasts of Risk in Criminal Justice Settings. Cham, Switz.: Springer
Berk RA. 2020a. Almost politically acceptable criminal justice risk assessment. Criminol. Public Policy. In press
Berk RA. 2020b. Statistical Learning from a Regression Perspective. Cham, Switz.: Springer. 3rd ed.
Berk RA, Bleich J. 2013. Statistical procedures for forecasting criminal behavior: a comparative assessment.
J. Criminol. Public Policy 12(3):515–44
Berk RA, Brown L, Buja A, George E, Pitkin E, et al. 2014. Misspecified mean function regression: making
good use of regression models that are wrong. Sociol. Methods Res. 43:433–51
Berk RA, Buja A, Brown L, George E, Kuchibhotla AK, et al. 2019. Assumption lean regression. Am. Stat.
[Link]
Berk RA, Heidari H, Jabbari S, Kearns M, Roth A. 2018. Fairness in criminal justice risk assessment: the state
of the art. Sociol. Methods Res. [Link]
Berk RA, Sorenson SB. 2016. Forecasting domestic violence: a machine learning approach to help inform
arraignment decisions. J. Empir. Legal Stud. 13(1):95–115
Berk RA, Sorenson SB. 2020. An algorithmic approach to forecasting rare violent events: an illustration based
on intimate partner violence perpetration. Criminol. Public Policy 19(1):213–33
Berk RA, Sorenson SB, He Y. 2005. Developing a practical forecasting screener for domestic violence. Eval.
Rev. 29(4):358–82
Bowers KJ, Johnson SD, Pease K. 2004. Prospective hot-spotting: the future of crime mapping? Br. J. Criminol.
44:541–658
Box GEP, Jenkins GW, Reinsel GC, Ljung GM. 2016. Time Series Analysis: Forecasting and Control. Hoboken,
NJ: Wiley
Braga AA, Papachristos AV, Hureau DH. 2014. The effects of hot spots policing on crime: an updated sys-
tematic review and meta-analysis. Justice Q. 31(4):633–63

234 Berk
 Breiman L. 2001a. Random forests. Mach. Learn. 45:5–32
Breiman L. 2001b. Statistical modeling: two cultures (with discussion). Stat. Sci. 16:199–231
Breiman L, Friedman JH, Olshen RA, Stone CJ. 1984. Classification and Regression Trees. Monterey, CA:
Wadsworth Press
Burgess EM. 1928. Factors determining success or failure on parole. In The Working of the Indeterminate Sen-
tence Law and the Parole System in Illinois, ed. AA Bruce, AJ Harno, EW Burgess, EW Landesco, pp. 205–49.
Springfield, IL: State Board Parole
Cameron AC, Trivedi PK. 2005. Macroeconometrics: Methods and Applications. New York: Cambridge Univ. Press
Caplan JM. 2011. Mapping the spatial influence of crime correlates: a comparison of operational schemes and
implications for crime analysis and criminal justice practice. J. Policy Dev. Res. 13(3):57–83
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Caplan JM, Kennedy LW, Miller J. 2010. Risk terrain modeling: brokering criminological theory and GIS
methods for crime forecasting. Justice Q. 28(2):360–81
Caplan JM, Kennedy LW. 2016. Risk Terrain Modeling: Crime Prediction and Risk Reduction. Berkeley: Univ.
Calif. Press
Cellen-Jones R. 2014. Stephen Hawking warns artificial intelligence could end mankind. BBC News: Technology,
Dec. 2. [Link]
Chinoy S. 2019. The racist history behind facial recognition. New York Times, July 10. [Link]
com/2019/07/10/opinion/[Link]
Chouldechova A. 2017. Fair prediction with disparate impact: a study of bias in recidivism prediction instru-
ments. arXiv:1610.075254v1 [[Link]]
Coglianese C, Lehr D. 2017. Regulating by robot: administrative decision making in the machine-learning
era. Georget. Law J. 105:1147–223
Corbett-Davies S, Goel S. 2018. The measure and mismeasure of fairness: a critical review of fair machine learning.
Paper presented at the 35th International Conference on Machine Learning, Stockholm
Cressie NAC. 1993. Statistics for Spatial Data. Hoboken, NJ: Wiley
Dinis-Oliveira RJ. 2017. Metabolic profile of flunitrazepam: clinical and forensic toxicological aspects. Drug
Metab. Lett. 11(1):14–20
Domingos P. 2015. The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our
World. New York: Basic Books
D’Orsogna MR, Perc M. 2015. Statistical physics of crime: a review. Phys. Life Rev. 12:1–21
Dwork C, Roth A. 2014. The algorithmic foundations of differential privacy. Found. Trends Theor. Comput. Sci.
9(3–4):211–407
Ferguson AG. 2017. The Rise of Big-Data Policing: Surveillance, Race, and the Future of Law Enforcement. New
York: NYU Press
Flaxman S, Chirico M, Pereira P, Loeffler C. 2019. Scalable high-resolution forecasting of sparse spatiotempo-
ral events with kernel methods: a winning solution to the NIJ “Real-Time Crime Forecasting Solution.”
Ann. Appl. Stat. 13(4):2564–85
Freedman DA. 2009. Statistical Models: Theories and Practice. Cambridge, UK: Cambridge Univ. Press. 2nd ed.
Friedman JS. 2002. Stochastic gradient boosting. Comput. Stat. Data Anal. 38:367–78
Goel S, Shroff R, Skeem JL, Slobogin C. 2019. The accuracy, equity, and jurisprudence of criminal
risk assessment. Risk-Resil. Res. In press. [Link]
journal-articles/files/manuscript_the_accuracy_equity_and_jurisprudence_of _criminal_risk_
assessment_1.2.19_.pdf
Goodfellow I, Bengio Y, Courville A. 2016. Deep Learning. Cambridge, MA: MIT Press
Grahm O, Dowe D. 2019. The Turing test. In The Stanford Encyclopedia of Philosophy, ed. EZ Zalta. Stanford,
CA: CSLI
Groff ER, La Vigne NG. 2004. Forecasting the future of predictive crime mapping. Crime Prev. Stud. 13:29–57
Hamilton M. 2016. Risk-needs assessment: constitutional and ethical challenges. Am. Crim. Law Rev.
52(2):231–92
Hastie T, Tibshirani R, Friedman J. 2009. The Elements of Statistical Learning: Data Mining, Inference, and
Prediction. New York: Springer. 2nd ed.
Helmstetter A, Ouillon G, Sornette D. 2003. Are aftershocks for large California earthquakes diffusing?
J. Geophys. Res. 108(B10):2483

[Link] • Criminal Justice and AI 235

 Hill K. 2020. Activate this ‘bracelet of silence,’ and Alexa can’t eavesdrop. New York Times, Febr. 16, Sect. BU,
p. 3
Hollywood JS, McKay KN, Woods D, Agneil D. 2019. Real-Time Crime Centers in Chicago. Santa Monica, CA:
RAND
Hu Y, Wang F, Guin C, Zhu H. 2018. A spatial-temporal kernel density estimation framework for predictive
crime hotspot mapping and evaluation. Appl. Geogr. 9:89–97
Huq AZ. 2019. Racial equality in algorithmic criminal justice. Duke Law J. 68(6):1043–134
Johnson SD, Birks DJ, McLaughlin L, Bowers KJ, Pease K. 2007. Prospective crime mapping in operational context.
Home Off. Rep. 19/07, Res. Dev. Stat. Dir., London
Kaufmann M, Egbert S, Leese M. 2019. Predictive policing and the politics of patterns. Br. J. Crimiol. 59:674–
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

92
Kearns M, Neel S, Roth A, Wu Z. 2018a. An empirical study of rich subgroup fairness for machine learning.
arXiv:1808.08166v1 [[Link]]
Kearns M, Neel S, Roth A, Wu Z. 2018b. Preventing fairness gerrymandering: auditing and learning subgroup
fairness. arXiv 1711.05144v4 [[Link]]
Kearns M, Roth A. 2019. The Ethical Algorithm: The Science of Socially Aware Algorithm Design. Oxford, UK:
Oxford Univ. Press
Kessel JM. 2019. Killer robots aren’t regulated. Yet. New York Times, Dec. 13. [Link]
2019/12/13/technology/[Link]
Kleinberg J, Mullainathan S, Raghavan M. 2017. Inherent tradeoffs in the fair determination of risk scores. Paper
presented at the 8th Innovations in Theoretical Computer Science Conference, Berkeley
Krohn J. 2019. Deep Learning Illustrated: A Visual, Interactive Guide to Artificial Intelligence. Boston: Addison-
Wesley
Kroll JA, Huey J, Barocas S, Felten EW. 2017. Accountable algorithms. Univ. Pa. Law Rev. 165(3):633–705
Lei J, G’Sell M, Rinaldo A, Tibshirani RJ, Wasserman L. 2018. Distribution-free predictive inference for
regression. J. Am. Stat. Assoc. 113(523):1094–111
Liesenfeld R, Richard J-F, Vogler J. 2017. Likelihood-based inference and prediction in spatio-temporal panel
count models of urban crimes. J. Appl. Econom. 32:500–620
Lyngstad T, Skardhammar T, Berk R. 2017. Predicting future crime at birth. Paper presented at the 73rd Annual
Meeting of the American Society for Criminology, Philadelphia
Malkevitch J. 2007. Taxi! American Mathematical Society, October. [Link]
feature-column/fcarc-taxi
Mayson SG. 2019. Bias in, bias out. Yale Law J. 128:2218–300
Metz C. 2019. Google claims a quantum breakthrough that could change computing. New York Times,
Oct. 23. [Link]
Mitchell M. 2019a. Artificial Intelligence: A Guide for Thinking Humans. New York: Farrar, Straus and Giroux
Mitchell M. 2019b. We shouldn’t be scared of “superintelligent A.I.” New York Times, Oct. 31. [Link]
[Link]/2019/10/31/opinion/[Link]
Mohler G. 2014. Marked point process hotspot maps for homicide and gun crime prediction in Chicago. Int.
J. Forecast. 30(3):491–97
Mohler GO, Short MB, Brantingham PJ, Schoenberg FP, Tita GE. 2011. Self-exciting point process modeling
of crime. J. Am. Stat. Assoc. 493:100–8
Mohler GO, Short MB, Malinowski S, Johnson M, Tita GE, et al. 2015. Randomized controlled field trials of
predictive policing. J. Am. Stat. Assoc. 110(512):1399–411
Mullainathan S. 2019. Biased algorithms are easier to fix than biased people. New York Times, Dec. 6. https://
[Link]/2019/12/06/business/[Link]
Munoz D, Bouchereau F, Vargas C, Enriquez R. 2009. Heuristic approaches to the position location problem.
In Position Location Techniques and Applications, pp. 103–52. Cambridge, MA: Academic
Ohlin LE, Duncan OD. 1949. The efficiency of prediction in criminology. Am. J. Sociol. 54:441–52
Pande V. 2018. Artificial intelligence’s ‘black box’ is nothing to fear. New York Times, Jan. 25. [Link]
[Link]/2018/01/25/opinion/[Link]
Pascanu R, Gulcehere C, Cho K, Begio Y. 2014. How to construct deep recurrent neural networks.
arXiv:13126026v5 [[Link]]

236 Berk
 Perry WL, McInnis B, Price CC, Smith SC, Hollywood JS. 2013. Predictive Policing: The Role of Crime Fore-
casting in Law Enforcement and Operations. Santa Monica, CA: RAND
Pindyck RS, Rubinfeld DL. 1981. Econometric Models and Economic Forecasts. New York: McGraw Hill. 2nd ed.
Poole D, Mackworth A, Goebel R. 1998. Computational Intelligence: A Logic Approach. Oxford, UK: Oxford
Univ. Press
Rachlinski JJ, Johnson SL, Wistrich AJ, Guthrie C. 2009. Does unconscious racial bias affect trial judges? Notre
Dame Law Rev. 84(3):1195–252
Ratcliffe JH. 2014. The hotspot matrix: a framework for spatio-temporal targeting of crime reduction. Police
Pract. Res. 5(1):5–23
Downloaded from [Link]. Guest (guest) IP: [Link] On: Tue, 01 Oct 2024 [Link]

Ratcliffe JH, McCullagh MJ. 2001. Chasing ghosts? Police perception of high crime areas. Br. J. Criminol.
41:330–41
Reinhart A, Greenhouse J. 2018. Self-exciting point processes with spatial covariates: modelling the dynamics
of crime. J. R. Stat. Soc. C 67(5):1305–29
Reiss AJ. 1951. The accuracy, efficiency, and validity of a prediction instrument. Am. J. Sociol. 17:268–74
Rey SJ, Mack EA, Koschinsky J. 2012. Exploratory space-time analysis of burglary patterns. J. Quant. Criminol.
28:509–31
Rosser G, Cheng T. 2016. Improving the robustness and accuracy of crime prediction with a self-exciting
point process through isotropic triggering. Appl. Spat. Anal. 12:5–25
Rosser G, Davies T, Bowers K, Johnson SD, Cheng T. 2017. Predictive crime mapping: arbitrary grids or
street networks? J. Quant. Criminol. 33:569–94
Scharre P. 2018. Army of None: Autonomous Weapons and the Future of War. New York: Norton
Smith CS. 2019. Dealing with bias in artificial intelligence: three women with extensive experience in A.I.
spoke on the topic and how to confront it. New York Times, Novemb. 19. [Link]
2019/11/19/technology/[Link]
Spielkamp M. 2017. Inspecting algorithms for bias. MIT Technology Review, Jan. 12. [Link]
[Link]/s/607955/inspecting-algorithms-for-bias
Starr S. 2014. Sentencing by the numbers. New York Times, Aug. 10. [Link]
magazine/[Link]
Szkola J, Piza EL, Drawve G. 2019. Risk terrain modeling: seasonality and predictive validity. Justice Q. http://
[Link]/10.1080/07418825.2019.1630472
Tarling R, Perry JA. 1981. Statistical methods in criminological prediction. In Prediction in Criminology, ed. DP
Farrington, R Tarling, pp. 210–30. Albany, NY: SUNY Press
Tibshirani R. 1996. Regression shrinkage and selection via the lasso. J. R. Stat. Soc. Ser. B 58(1):267–88
Turing A. 1950. Computing machinery and intelligence. Mind 59(237):433–60
Wang B, Yin P, Bertozzi AL, Brantingham PJ, Osher SJ, Xin J. 2017a. Deep learning for real-time crime
forecasting and its ternarization. arXiv:1711.08833v1 [[Link]]
Wang B, Zhang D, Zhang D, Brantingham PJ, Bettozzi AL. 2017b. Deep learning for real time crime fore-
casting. arXiv:1707.03340v1 [[Link]]
Weisburd D, Groff ER, Yang S-M. 2012. The Criminology of Place. Oxford, UK: Oxford Univ. Press
Weisburd D, Mastrofski SD, Greenspan R, Willis JJ. 2004. The growth of Compstat in American policing. Rep.,
Police Found., Washington, DC
Zeng J, Ustun B, Rudin C. 2017. Interpretable classification models for recidivism prediction. J. R. Stat. Soc.
Ser. A 180(3):689–722
Zhang Y, Cheng T. 2020. Graph deep learning model for network-based predictive hotspot mapping of sparse
spatio-temporal events. Comput. Environ. Urban Syst. 79:101403
Zhuang Y, Almeida M, Morabito M, Ding W. 2017. Crime hot spot forecasting: a recurrent model with spatial
and temporal information. Paper presented at the 8th IEEE International Conference on Big Knowledge,
Hefei, China
Zucchini W, MacDonald IL, Langrock R. 2016. Hidden Markov Models for Time Series: An Introduction Using
R. Boca Raton, FL: CRC Press. 2nd ed.

[Link] • Criminal Justice and AI 237