A HOOVER INSTITUTION ESSAY

National Security, Technology, and Law

Facial Recognition as a Less
Bad Option
JANE BAMBAUER Aegis Series Paper No. 2107

This essay does the unthinkable—it defends the police use of facial recognition
technology to identify suspects in crime footage or to locate individuals with
outstanding warrants. I argue that the perils that flow from facial recognition can be
mitigated through sensible limits without banning the technology, and that in any
case, the risks of facial recognition are less bad than the options police have without
its use. In other words, acknowledging the potential costs of police use of facial
recognition, I make the case that such use is still warranted.

In broad strokes, my argument goes as follows: (1) to the extent criminal justice
reformers have political capital to spend, it should be spent reducing criminal liability
and sentences for most crimes while increasing the probability that criminal conduct
will be detected so that crime rates stay low; and (2) facial recognition is a valuable
tool for increasing the probability of detection because it reduces the discretion that
police officers have as compared to other forms of surveillance. Put differently, and all
things being equal, it is more efficient and fairer for police to run a photograph through
facial recognition software to identify candidate suspects than to try to identify the
suspect using witnesses or to solve the case without using the image.

The essay unfolds in three sections. Section 1 builds a case for increasing the use of
surveillance technologies as part of a grand bargain for criminal justice reform. Those
who believe that the American criminal justice system currently relies on excessively
harsh punishment to keep crime rates low (a point I will argue for below) should be
willing to increase surveillance and criminal detection in exchange for changes in
punishment or in the substance of the criminal code. That is, given a choice among
potential criminal justice reforms, a responsible policy maker should do whatever
possible to dramatically reduce criminal sentencing and the indirect effects of
incarceration. Since this can only be accomplished by either increasing the capacity
to detect criminal conduct using surveillance tools or allowing the crime rates to
creep back up, an honest account of criminal justice reform must confront trade-offs.
Much of the political discourse ignores or wishes away the inescapable trilemma
between the probability of enforcement, the harshness of punishment, and crime
rates. If we acknowledge the trade-offs between surveillance, punishment, and
crime, then the prohibitions of new surveillance technologies have an obvious
2

practical and political cost: they will make decriminalization and sentencing reform
more difficult. To put it positively, instead of negatively, I will argue that we should
consider embracing new surveillance tools like facial recognition systems so that
they can be used as part of a strategic package of reforms that more directly benefit
communities harmed by both crime and policing.

Section 2 of this essay argues that facial recognition technologies are more valuable
than the average surveillance tool because they rely on only limited amounts of police
discretion. Facial recognition is used today in a crime-driven, rather than suspect-driven,
manner.1 Because police start with photographic or video evidence from a crime scene
or from an arrest file and use facial recognition to work out toward a suspect, the tool
has the potential to decrease bias and arbitrariness that can taint investigations when
they start with one or more identified suspects and attempt to build up to probable
cause. Facial recognition technologies can also reduce the high stakes of arrest. If
a suspect resists arrest but does not seem to present a physical danger to the police
or others, a police officer may be less likely to use force against the suspect if he has
increased confidence that a facial recognition system will find the suspect again soon.
(Likewise, the suspect himself is less likely to resist for the same reason.)

Finally, although the use of facial recognition presents real risk related to the invasion
of privacy and racially biased error, these criticisms probably fail the “Compared
to what?” test. Thus, the moral calculus of police use of facial recognition is quite
complex, notwithstanding the consensus among thought leaders that it should be
banned.

I. The Inescapable Trilemma: Crime, Detection, and Punishment

Crime is terrible, and so is its detection and punishment.2

In the United States, crime rates are quite low in historical terms. Violent crimes have
basically dropped by half since the early 1990s, and property crimes have dropped
even more dramatically.3 And although Black and Hispanic Americans are more likely
to become victims of crime than White Americans, homicide rates plummeted for
all racial groups between 1993 and 2014 and continues to be fairly low in historical
terms.4 In 1993, one out of every fifteen Black Americans was the victim of a violent
crime at some point during the year,5 but in 2018 the figure was one in eighty.6 For
those of us interested in criminal reform, this is excellent news because there is
some buffer—perhaps some room to spare—to experiment with reforms even if they
cause temporary increases in crime rates.7 On the other hand, recent spikes in crime
(particularly homicide) during 2020 and 2021 suggest this window of opportunity may
close as voters and constituents return to a state of anxiety and put pressure on the
criminal justice system to drive crime rates back down.8

Jane Bambauer  •  Facial Recognition as a Less Bad Option

 3

Economist Gary Becker famously modeled crime with a simple formula determined
by the probability of conviction and the severity of punishment.9 Becker was writing
at the height of the rational actor approach to legal design; because it was much
easier for the state to ratchet up punishment than to catch more perpetrators, his work
persuaded many politicians to manage crime through tough sentencing.10

The sparseness of Becker’s model for crime rates leaves much to be desired. There is
ample evidence that crime has a range of social and economic causes and that there
is a limit to the cold rational actor model.11 Nevertheless, there is little reason to doubt
that detection and punishment of crime are important factors that influence the
amount of crime in a given community at a given time,12 and these factors are more
directly under the control of a politically accountable mayor, police chief, or state
legislature than many of the other social and cultural determinants.

On severity of punishment, the United States stands out with a brutal and grimly
indifferent penal system unmatched by any of our political allies. We use incarceration
intensively. In France and the United Kingdom, a criminal who punches a person
in the nose would probably be sentenced to less than six months in jail.13 The same
conduct in the United States would likely result in a sentence of about three years.14
Moreover, no outsider would mistake our prisons for institutions of rehabilitation. To
the contrary, the entire sentence is usually carried out in a facility that is punishing,
with drab quarters, humiliating toilet and bathroom facilities, and rancid food. Once
released, the negative consequences continue as the housing and labor markets
penalize criminal convicts.

All of this might be justified if the Becker model of criminal forecasting were valid,
but it is not. In fact, punishment has a U-shaped relationship to recidivism, where no
punishment and significant periods of incarceration both tend to increase the odds
that a perpetrator will recidivate. This relationship is due in part to the criminogenic
effect that the prison experience has. A recent study by Amanda Agan, Jennifer Doleac,
and Anna Harvey provides some of the most compelling proof that exposure to prison
increases recidivism rather than decreasing it.15 In terms of general deterrence, the
length of a prison sentence has swiftly diminishing marginal returns on the likelihood
that a person who has not yet committed a crime will decide to exercise restraint.16
On the other hand, crime rates can be reduced through incapacitation rather than
deterrence—that is, by imposing very long sentences on first-time convicts.

Thus, punishment reduces crime in two situations: either as part of a system with high
probability enforcement (in which case punishment can be mild) or as part of a system
that attempts to incapacitate criminals by confining them for as long as society will
tolerate (which necessitates that the punishment be harsh).

Hoover Institution  •  Stanford University

4

• • •

This brings us to the second Becker factor—the probability of detection of criminal

conduct. On this factor, the United States is in bad shape. Less than half of the violent
crimes reported to the police result in an arrest and referral for prosecution and for
property crime, the figure is under 20 percent.17 Moreover, most crime is not even
reported. Only about half of violent crimes are ever reported to the police, and about
one-third of property crimes. In other words, more than half of victim-based crimes
in the United States don’t even make it to the denominator in the clearance rates.
Together, the low likelihood of reporting and the low clearance rates means that the
probability a criminal will be prosecuted for any particular incidence of violence
is about 20 percent. (The figure for property crime is 7 percent.)18 America uses
haphazard enforcement: occasional and harsh.19

Unlike the severity of punishment, there is abundant evidence that crime rates are very
responsive to the probability of enforcement.20 There is also some weaker evidence
that the swiftness of enforcement—the “celerity”—makes a difference. The detection,
identification, and arrest of suspects requires surveillance, of course, and the privacy
intrusions from surveillance constitute the third leg of the inescapable trilemma.

II. Surveillance as the Least Bad Option

Let’s step back and observe the three legs of the miserable trilemma: surveillance,
punishment, and crime. Holding all else constant,21 a policy goal to reduce any one of
these will require tolerance of an increase in at least one of the others. So, at the outset,
the trilemma illuminates why we should develop a tolerance (perhaps even a desire) to
increase surveillance.

Tolerance for surveillance serves two purposes. First, and most importantly, it will
be necessary if we want to reduce punishment without a spike in crime rate. This
is especially relevant for reformers and lawmakers interested in rapid decarceration
and sentencing reductions. Since rising crime rates have already taxed the public’s
tolerance for crime, the only responsible and politically palatable course is to increase the
chance of detection and enforcement of serious crimes, which will require either more
intensive use of surveillance technologies that are already available or the development
of new surveillance technologies. For the purposes of this section, I will treat both
of those options (more use of extant surveillance systems or development of new
surveillance tech) under the single concept of increased surveillance.

The clash between American values in privacy and security is most pronounced at the
early stages of investigation, before police have probable cause to arrest a particular
individual. Norms and the Fourth Amendment offer quite a bit of latitude for

Jane Bambauer  •  Facial Recognition as a Less Bad Option

 5

investigation after the police have reached the probable cause threshold. At that point,
the police can conduct full-blown searches and collect communications information
(within the bounds and limitations of the warrant requirement, of course). But before
probable cause is established, police must build cases using information that is acquired
outside the scope of the technical definition of a “search.” It is the pre–probable cause
stage of an investigation where advances in surveillance technologies are most likely to
improve detection. But this is also the stage of investigation that is most embroiled in a
battle over the scope and future of the Fourth Amendment.

Facial recognition technologies sit right at the center of this clash. With the possible
exception of drones, facial recognition is the technology most reviled by progressives
and civil libertarians alike.22 Yet it seems to me that legal bans on facial recognition
will be a Pyrrhic victory because poor criminal detection will leave the harsh
punishment equilibrium in place.

Progressive concerns over the inhumanity and inequity in mass incarceration and
libertarian concerns about unchecked state power would benefit more directly from
decriminalization and from greatly reducing sentences for the crimes that are left
on the books. I have argued that legislators or courts should set ceilings that are
significantly lower than average sentences today in order to constrain and properly
calibrate the state’s imposition of punishment on the unlucky minority of criminal
actors who are caught.23 That is, if the state is constrained by law or constitutional
interpretation from detaining or imprisoning individuals at all based on minor
infractions, or from levying long sentences for anything other than the most serious
and violent offenses, then the threat of state surveillance is reduced.

Political pressure to ban surveillance tools might make sense as a second-best solution if
decriminalization and reduced sentencing is politically infeasible, but the risk is that the
strategy can lock out the first-best solution—the low-penalty/high-detection solution.
Indeed, as murder rates have risen over the last year and a half, the decriminalization
and police reform movements are already more politically controversial than they
were a couple years ago. If crime rates continue to rise while detection is capped or
suppressed through new legal constraints on technology, politically accountable
decision makers are likely to continue to rely on incapacitation (i.e., very long prison
sentences) to manage crime.

The second reason to tolerate surveillance more than the other options is that
surveillance is likely to be a necessary component for alternatives to criminal justice
systems anyway. If we expect the government to handle social problems related to
drugs, mental health, or gangs using public health models of intervention, then those,
too, will depend on surveillance to gauge risk, monitor compliance, and create positive
feedback loops for individuals in nonpunitive treatment programs. Indeed, in the

Hoover Institution  •  Stanford University

6

case of health crises, Americans sometimes want more surveillance if the data might
improve public health.24

To summarize, the least bad approach to criminal justice would involve dramatic
reductions in the civil or criminal penalties for misconduct combined with
a greatly increased probability of detection. And the best alternatives to a
criminal justice system are also likely to require surveillance. On these assumptions,
and assuming that the risks of abuse of new surveillance tools can be effectively
managed (which requires its own commitment to surveilling and policing the police),
there is reason to embrace effective surveillance as part of a first-best or least bad
approach to managing crime.

What is much less clear at this point is how we could possibly achieve the first-best
solution given the legal and practical constraints we have today. The US Constitution
places minimal constraints on crime rates and sentencing length, but it places significant
restrictions on surveillance. That is, there is no explicit constitutional duty for the
government to keep crime below a particular threshold. The crime leg of the tripod
can be as large as circumstances, local government, and the voting public allow it to
get. Police rarely have an affirmative duty to respond to crime.25 The punishment leg
is also effectively unconstrained: statutory and constitutional limits on the severity of
punishment are so lax that they create no meaningful constraint. The surveillance leg
is the only one limited by constitutional and statutory rules, and police often operate
up to the bounds of the legal limits. When law enforcement does find a way to expand
that leg of the tripod—by adopting a new surveillance technology—the public is often
quick to react and demand legal restrictions on its use. Thus, for the foreseeable future,
the reformers have only two legs to work with—they can reduce punishment and
incarceration (causing a rise in crime rates), or they can reduce crime (by increasing
punishment).

So far, I have discussed the benefits of facial recognition as a generic tool of surveillance
and detection. But I will go further. Even among possible surveillance tools, facial
recognition has some advantages over traditional policing. Facial recognition is no
worse and often better than other investigation tools when it comes to criminal justice
problems related to privacy, use of force, and bias.

III. Facial Recognition as a Least Bad Form of Surveillance

Police must rely on some form of information-gathering to build an initial case against
a suspect or to locate a person with an outstanding warrant. Whatever they do to
gather incriminating information will constitute some form of surveillance and will
raise the possibility of an intrusive privacy invasion.

Jane Bambauer  •  Facial Recognition as a Less Bad Option

 7

In typical accounts of law enforcement surveillance, observations of clear evidence

of crime are not in themselves socially harmful. If a police officer observes a person
committing a crime, that observation is generally thought to implicate no legally
recognized privacy interest.26 However, whatever surveillance is used will usually
collect more information than strictly relevant and necessary for the investigation
or prosecution of a crime. When a police officer patrols a public street or enters a
home with consent or pursuant to a warrant, whatever evidence of criminality is
discovered there is accompanied with the observation of lots of other irrelevant
information related to perfectly legal behavior. The revelation of those licit details
is the privacy harm. That information can be used later to harass or embarrass an
individual, or to pursue a personally or politically motivated investigation. Even if
negative consequences never occur, the loss of control over information related to
private (and legal) conduct constitutes a dignitary loss.

Facial recognition technologies present privacy risks, but they are modest compared
to other forms of investigation (at least, as the technology is typically used today).27
In the use cases that are most likely to proliferate across police departments, facial
recognition is used when police have already collected photographic or video evidence
from the scene of the crime, or where the police have already sought and received a
warrant based on probable cause from other sources and are pursuing the final step of
identifying or locating a suspect. This differs from investigations that involve tailing
a suspect for a period of time or talking to confidential sources because the amount of
extraneous information gathered by the police is limited. Police will not observe the
inside of a suspect’s car or home and often will not even know his movement patterns.
Other than identity, little is revealed by facial recognition technologies per se. They are
privacy-preserving compared to other forms of information-gathering.

Of course, licit details will be revealed anytime facial recognition falsely identifies a
suspect who is then subjected to an arrest or probable cause-based search. Much like
the drug-sniffing dog,28 facial recognition is far from infallible, and the false match error
will lead to privacy invasions. But no investigation tool is free from error, and facial
recognition outperforms the accuracy rates of eyewitnesses and PC-based warranted
searches by a large margin.29 The same is true for racial differences in error rates:
while some facial recognition technologies are more likely to produce false matches
for photographs of Black faces,30 the gap in false match error is likely to be reduced
over time, and in any event may already be less bad than the difference in false match
error from human systems of suspect identification.31 Moreover, unlike traditional
policing methods, facial recognition technology can be calibrated to only produce a
match when the risk of a false match is below a certain threshold (ensuring equal false
positive rates across race).32

Hoover Institution  •  Stanford University

8

Facial recognition surveillance also differs in important respects from suspect-driven

investigations. In suspect-driven investigations, police have developed suspicion
(or a hunch) around a particular individual and focus their observations on the
suspect in order to develop a case. Suspect-driven investigations are propelled by the
theories of police officers and proceed at their discretion. By contrast, police have less
control over the results of facial recognition investigations that stem from evidence
at a crime scene.33 If facial recognition identifies a wealthy or politically connected
individual as the suspect of a crime, it will be much more difficult for police to fail
to pursue that lead than in a case where the police use informants or witnesses as the
main source of identification. Like geofencing techniques (where GPS data is used to
identify who was at the scene of a crime at the appropriate time), police cannot exert
control over which individuals will wind up within the scope of suspicion. Techniques
that involve starting from the facts of a crime and working toward an identity are
sometimes criticized for failing to limit the number of people who wind up in the
ambit of potential suspicion,34 but at a conceptual level, it’s hard to fault investigative
techniques that begin from the facts of a crime.

None of this is meant to suggest that facial recognition is free from bias or abuse. First,
police will decide which crime scene images should be subjected to facial recognition.
They will have to decide, for example, whether to pursue arrest and prosecution
of violent or destructive rioters at a Black Lives Matter protest or at a pro-Trump
rally; this decision is subject to justified criticism if the standards between the two
differ. Likewise, in the case of locating identified individuals with outstanding
search warrants, police departments will decide which public and online spaces to
monitor for potential matches. Still, these types of enforcement decisions about which
crimes to investigate and where to look for outstanding suspects will have to be made
regardless of the form of surveillance used to carry out the investigation.

In fact, there is reason to believe that such an effective technology as facial recognition
would reduce, rather than increase, the chance that police will focus enforcement on
underprivileged communities. The role of discretion in decision-making related to
which types of crimes to prosecute or which parts of a town to monitor are themselves
a product of limited capacity to detect crime. A police department’s limited resources
give it the power and excuse to pick and choose between different law enforcement
missions and different locations of scrutiny. If cheap, privacy-preserving surveillance
tools were readily available, there would be no reason to use it for some crime
footage and not for others, or to scan for people with outstanding arrest warrants in
some neighborhoods and not others. The decision to concentrate the tool on minority
neighborhoods would be highly suspect (even if the neighborhoods have higher crime
rates) if there is no significant cost to deploying them everywhere, or in locations
where more people congregate. Thus, while a marginally useful surveillance tool
might exacerbate racial disparities in criminal investigation and enforcement rates,

Jane Bambauer  •  Facial Recognition as a Less Bad Option

 9

a cheap and highly efficacious criminal detection tool can reduce them (by increasing
detection across the board).

Facial recognition can also help reduce the stakes during police encounters and
seizures. Police will not need to use force to stop a suspect who is resisting arrest and
attempting to flee if they are confident that the suspect will be located again very
soon using facial recognition.35 (And for the same reasons, the suspect is also less likely
to flee.)

Finally, tools like facial recognition can be calibrated to report an alert only for
particular types of crime. If a police department sets up facial recognition at a
crowded sports stadium, they could program the app to alert on individuals who have
outstanding warrants for only a subset of serious crimes. If there are doubts that the
police would constrain their use in this way, courts or lawmakers could restrict use to
solving serious crimes rather than banning the technology outright.36

Some fear the government will begin to combine facial recognition with a network of
surveillance camera footage to record and store the location and movement history
of law-abiding individuals in identified form indefinitely, and for unlimited purposes.
But Fourth Amendment cases like Carpenter v. United States and possibly United States
v. Jones seem to effectively foreclose that scenario.37 The city of Baltimore, for example,
has already had a police program dismantled on Fourth Amendment grounds because
the city was keeping aerial surveillance footage of the whole city for up to six months.38
And law enforcement use of automated license plate reading technologies have been
approved by courts only when their use is limited to short amounts of time or to cars
that are not owned by the target.39 Persistent monitoring of movement using facial
recognition would presumably be struck down for the same reasons. In any case, the
network of surveillance cameras and indefinite storage of their footage rather than
facial recognition technology would be the most proximate cause of risk if a stockpile
of this sort of data were to be created.40

Conclusion: The Least Bad among Bad Options

As a society, we value freedom from government surveillance as well as freedom from
crime victimization. Since law enforcement operates in the unavoidable clash between
these two goals, all options are bad.

Reducing harsh criminal penalties and keeping crime rates low are, or should be, the
greatest goals for American policy makers who want to reform the criminal justice system.
Keeping surveillance capacities stuck at twentieth-century standards is the least important
criminal justice reform goal. In other words, surveillance is the more manageable option
in an unavoidable trilemma.

Hoover Institution  •  Stanford University

10

Among surveillance tools, compared to surveillance that gathers and retains licit
details about individuals or that relies on the discretion of police, investigations that
use facial recognition to match a known crime to an unidentified suspect is the best
among bad options.

ACKNOWLEDGMENTS

I am very grateful to the excellent feedback on early drafts from Barry Friedman,
Jack Goldsmith, Farhang Heydari, Elizabeth Joh, Orin Kerr, Jennifer Lynch,
Christopher Slobogin, Mark Verstraete, Andrew Woods, and the participants of
the Hoover Institution roundtable on “The Private Market and Public Surveillance.”
I am also indebted to the support of the University of Arizona College of Law
library staff who helped me track down statistics and provided other research
support.

NOTES
1 ​See Shirin Ghaffary, How to Avoid a Dystopian Future of Facial Recognition in Law Enforcement, Vox
(Dec. 10, 2019), https://­w ww​.­vox​.­com​/­recode​/­2019​/­12​/­10​/­20996085​/­ai​-­facial​-­recognition​-­police​-­law​
-­enforcement​-­regulation (describing current law enforcement uses).

2 ​See, e.g., David A. Anderson, The Aggregate Burden of Crime, 42 J.L. & Econ. 611 (1999); Aaron Chalfin &
Justin McCrary, Are U.S. Cities Under-Policed? Theory and Evidence (National Bureau of Economic Research.
Working Paper No. 18815, 2013); Kathryn E. McCollister et al., The Cost of Crime to Society: New Crime-
Specific Estimates for Policy and Program Evaluation, 108 Drug & Alcohol Depend. 98 (2010).

3 ​John Gramlich, What the Data Says (and Doesn’t Say) about Crime in the United States, Pew Rsch. Ctr.
(Nov. 20, 2020), https://­w ww​.­pewresearch​.­org​/­fact​-­tank​/­2020​/­11​/­20​/­facts​-­about​- ­crime​-­in​-­the​-­u​-­s
[hereinafter Gramlich 2020]; Rachel E. Morgan & Barbara A. Oudekerk, U.S. Dept. of Just., NCJ-253043,
Criminal Victimization, 2018 (2019). Although crimes of all sorts (and particularly murder) have
skyrocketed during the COVID-19 pandemic, the pandemic-related stress on social and economic
­well-being make the recent data difficult to interpret. Compare Paul G. Cassell, Explaining the Recent
Homicide Spikes in U.S. Cities: The “Minneapolis Effect” and the Decline in Proactive Policing, 33 Fed.
Sent’g Rep. 83 (2020), with Jeffrey Fagan & Daniel Richman, Understanding Recent Spikes and Longer Trends
in American Murders, 117 Colum. L. Rev. 1235 (2017), and German Lopez, The Rise in Murders in the US,
Explained, Vox (Dec. 2, 2020), https://­w ww​.­vox​.­com​/­2020​/­8​/­3​/­21334149​/­murders​- ­crime​-­shootings​-­protests​
-­riots​-­trump​-­biden.

4 ​Federal Bureau of Investigation, Crime Data Explorer, https://­crime​-­data​-­explorer​.­app​.­cloud​.­gov​/­pages​

/­explorer​/­crime​/­shr (last visited Oct. 20, 2021) (focusing on “Expanded Homicide Data” from 1985 to 2020).

5 ​Craig A. Perkins et al., U.S. Dept. of Just., NCJ-151657, Criminal Victimization in the United States,
1993, at v (1996).

6 ​U.S. Dept. of Just., supra note 3, at 16.

7 ​I should note, though, that there may not be much political appetite even when crime rates are dropping
since the United States, even in its lowest crime period, is still far more crime-ridden than other developed

Jane Bambauer  •  Facial Recognition as a Less Bad Option

 11

nations. For example, 5.4 out of every 100,000 Americans were killed by homicide in 2016, whereas in
France the rate was 1.4 out of every 100,000. Victims of Intentional Homicide, 1990–2018, United Nations
Off. on Drugs & Crime, https://­dataunodc​.­un​.­org ​/­content​/­data​/­homicide​/­homicide​-­rate.

8 ​See Jeff Asher, Murder Rose by Almost 30% in 2020. It’s Rising at a Slower Rate in 2021, N.Y. Times
(Sept. 22, 2021), https://­w ww​.­nytimes​.­com​/­2021​/­0 9​/­2 2​/­upshot​/­murder​-­r ise ​-­2020​.­html; Domenico
Montanaro, Rising Violent Crime Is Likely to Present a Political Challenge for Democrats in 2022, NPR
(July 22, 2021), https://­www​.­npr​.­org​/­2021​/­07​/­22​/­1018996709​/­rising​-­violent​-­crime​-­is​-­likely​-­to​-­present​-­a​
-­political​-­challenge ​-­for​-­democrats​-­in.

9 ​Gary S. Becker, Crime and Punishment: An Economic Approach, 76 J. Polit. Econ. 169 (1968); see also
A. Mitchell Polinsky & Steven Shavell, The Theory of Public Enforcement of Law, in Handbook of Law &
Economics 421 (2007).

See Cass R. Sunstein et al., Do People Want Optimal Deterrence?, 29 J. Legal Stud. 237 (2000).
10 ​

11 ​See, e.g., Cesare Lombroso, Criminal Man (Mary Gibson & Nicole Rafter, trans., 2006) (1876) (discussing
genetic theories of crime); S. J. Schoenthaler & I. D. Bier, The Effect of Vitamin-Mineral Supplementation
on Juvenile Delinquency among American Schoolchildren: A Randomized, Double-Blind Placebo-Controlled
Trial, 6 J. Alt. & Complementary Med. 7 (2000) (discussing malnutrition as a factor in crime); Civic Research
Institute, The Science, Treatment, and Prevention of Antisocial Behaviors (Diana H. Fishbein ed., 1999)
(reviewing evidence of the impact of alcoholism, drug use, sexual abuse, cognitive and genetic factors,
and family/gender role factors); Clifford R. Shaw & Henry D. McKay, Juvenile Delinquency and
Urban Areas (1942) (exploring the effect of weakened or disorganized social institutions on crime and
including the roots of what would become the “broken windows” theory).

See Exec. Off. of the President, Economic Perspectives on Incarceration and the Criminal Justice
12 ​
System 36–40 (2016) (citing to the empirical literature finding that increased incarceration reduces crime,
but less effectively than equivalent increased spending on police); Andrew von Hirsch, Doing Justice:
The Choice of Punishments 61– 65 (1976); Raymond Paternoster, The Deterrent Effect of the Perceived
Certainty and Severity of Punishment: A Review of the Evidence and Issues, 42 Just. Q. 173 (1987).

13 ​House of Commons Library, CBP 7218, Comparative Prison Sentences in the EU, 2010 (2015),
https://­commonslibrary​.­parliament​.­uk​/­research​-­briefings​/­cbp​-­7218.

14 ​U.S. Sent’g Comm’n, Table 15, Sentence Imposed by Type of Crime, Fiscal Year 2020, in Sourcebook of
Federal Sentencing Statistics (2020), https://­w ww​.­ussc​.­gov​/­research. Note, though, that the differences
for nonviolent offenses like theft appear to be smaller (<6 months in the United Kingdom compared to a
median of eight months in the United States). It should also be noted that the US data is drawn from 2020
while the European data relates to 2010; however, data from the United States in 2010 leads to similar
results. U.S. Sent’g Comm’n, Table 13, Sentence Length in Each Primary Offense Category, Fiscal Year 2010,
in Sourcebook of Federal Sentencing Statistics (2010), https://­w ww​.­ussc​.­gov​/­research.

15 ​Amanda Y. Agan et al., Misdemeanor Prosecution (National Bureau of Economic Research. Working Paper
No. 28600, 2021).

Von Hirsch, supra note 12, at 61–65.

16 ​

17 ​John Gramlich, Most Violent and Property Crimes in the U.S. Go Unsolved, Pew Rsch. Ctr. (Mar. 1, 2017),
https://­w ww​.­pewresearch​.­org ​/­fact​-­tank​/­2017​/­03​/­01​/­most​-­violent​-­and​-­property​-­crimes​-­in​-­the​-­u​-­s​-­go​
-­unsolved [hereinafter Gramlich 2017]; Gramlich 2020, supra note 3. These statistics deal with clearance
rates that measure the annual number of reported cases that result in an arrest and referral for prosecution
but do not take into account reported cases that are cleared in subsequent years.

18 ​See Gramlich 2017, supra note 17; Gramlich 2020, supra note 3.

Hoover Institution  •  Stanford University

12

19 ​This critique, it should be noted, dates back to the eighteenth-century work of Jeremy Bentham and
Cesare Beccaria. See Raymond Paternoster, How Much Do We Really Know about Criminal Deterrence?,
100 J. Crim. L. & Criminology 765, 767–73 (2010).

20 ​Reviews of empirical literature consistently find that harsh sentences cannot be justified on deterrence
grounds because while crime rates are highly sensitive to the probability of enforcement, the severity of
punishment has no consistent effect. See, e.g., Aaron Chalfin & Justin McCrary, Criminal Deterrence: A Review of
the Literature, 55 J. Econ. Literature 5, 13–15, 23–29 (2017); Steven N. Durlauf & Daniel S. Nagin, Imprisonment
and Crime: Can Both Be Reduced?, 10 Criminology & Pub. Pol’y 13, 17 (2011); Daniel S. Nagin, Deterrence in
the Twenty-First Century, 42 Crime & Just. 199 (2013); Daniel S. Nagin, Deterrence: A Review of the Evidence
by a Criminologist for Economists, 5 Ann. Rev. Econ. 83 (2013); Jeffrey Grogger, Certainty vs. Severity of
Punishment, 29 Econ. Inquiry 297 (1991).

21 ​Even if we do not hold all else constant—if, for example, social disruption causes crime to increase, or if
nutritional improvements cause crime to decrease—unless the change causes a dramatic shift in crime, the
trade-offs between the three factors in the trilemma will still have to be made under conditions similar to
those that we have today.

22 ​ See, e.g., Press Release, Ban Dangerous Facial Recognition Technology That Amplifies Racist Policing,
Amnesty International (Jan. 26, 2021), https://­w ww​.­amnesty​.­org​/­en​/­latest​/­press​-­release​/­2021​/­01​
/­ban​-­dangerous​-­facial​-­recognition​-­technology​-­that​-­amplifies​-­racist​-­policing (“Facial recognition risks
being weaponized by law enforcement against marginalized communities around the world . . . ​[ T]his
invasive technology turns our identities against us and undermines human rights.”); The Fight to Stop Face
Recognition Technology, Am. Civ. Liberties Union (July 15, 2021), https://­w ww​.­aclu​.­org ​/­news​/­topic​
/­s topping​-­face ​-­recognition​-­surveillance/ (“Face recognition surveillance presents an unprecedented
threat to our privacy and civil liberties.”); Press Release, Pressley, Clarke, Tlaib Reintroduce Bill to Ban
Facial Recognition Technology in Public Housing, U.S. Congresswoman Ayanna Pressley (July 7, 2021),
https://­pressley​.­house​.­gov​/­media​/­press​-­releases​/­pressley​-­clarke​-­tlaib​-­reintroduce​-­bill​-­ban​-­facial​
-­recognition​-­technology​-­public.

23 ​Jane Bambauer & Andrea Roth, From Damage Caps to Decarceration: Extending Tort Law Safeguards to
Criminal Sentencing, 101 B.U. L. Rev. (forthcoming 2021).

24 ​Cason Schmit et al., U.S. Privacy Laws Go against Public Preferences and Impede Public Health and
Research: Survey Study, 23 J. Med. Internet Rsch. (2021).

25 ​See, e.g., DeShaney v. Winnebago County, 489 U.S. 189 (1989) (no constitutional duty to protect a
young child from domestic abuse); Castle Rock v. Gonzales, 545 U.S. 748 (2005) (same); Riss v. City of
New York, 22 N.Y.2d 579 (N.Y. 1968) (denying tort recovery to the victim of an attack who had requested
police assistance multiple times). However, some state Riot Act statutes allow state and local government
to be sued for failure to respond and protect others in the course of a riot. See Susan Glassberg, Liability for
Urban Riot Damage, 1971 Urban L. J. 193 (1971).

26 ​Unless, of course, the substance of the criminal law is flawed or the penalties are too harsh. The implicit
threat from surveillance is that it can lead to the enforcement of civil or criminal laws that many believe
are substantively flawed (e.g., drug or immigration enforcement) or that are overpenalized (e.g., most
nonviolent crimes) is a concern of high importance to me. However, faults in the substance or punishment
of criminal law exist regardless of the type of surveillance used and are better addressed through
decriminalization and reduction in sentences than through haphazard reductions in detection.

27 ​By contrast, programs that use facial recognition as part of a mass surveillance or suspect-driven
investigation do not have the benefits that I describe here. As I describe below, such programs would face
a constitutional challenge in the wake of Carpenter v. United States. Perhaps more concerning is the risk
that a surveillance tool could be used by police officers in violation of internal protocols in order to harass

Jane Bambauer  •  Facial Recognition as a Less Bad Option

 13

or exploit another person. The rogue use of police equipment is always a risk, but in the case of facial
recognition software, the technology is readily available to private actors anyway. See Kashmir Hill, Activists
Turn Facial Recognition Tools against the Police, N.Y. Times (Aug. 1, 2021), https://­w ww​.­nytimes​.­com​/­2020​
/­10​/­21​/­technology​/­facial​-­recognition​-­police​.­html. Thus, an officer’s access to official law enforcement
systems may not greatly increase the risk that a rogue agent will make inappropriate use of the tool since
he may be able to use the same technology outside of work. That said, the legitimacy of the use of this and
any surveillance system depends on law enforcement’s willingness to hold its own agents to high standards
and to avoid corruption and abuse.

28 ​Illinois v. Caballes, 543 U.S. 405, 411 (2005) (Souter, J., dissenting) (“The infallible dog, however, is a
creature of legal fiction.”).

29 ​False match error rates for facial recognition algorithms are now under 1 percent in ideal
conditions and under 10 percent when used in the field, and one facial recognition vendor recommends
law enforcement use a threshold of 95 percent confidence. William Crumpler, How Accurate Are Facial
Recognition Systems—and Why Does It Matter?, Ctr. for Strat. & Int’l Stud. (Apr. 14, 2020), https://­w ww​
.­c sis​.­org ​/­blogs​/­technology​-­policy​-­blog ​/­how​-­accurate​-­are​-­f acial​-­recognition​-­s ystems​-­%E2%80%93​
-­and​-­why​- ­does​-­it​-­matter. By comparison, eyewitness identification during a lineup has error rates of
20 percent or more. Gary L. Wells & John W. Turtle, Eyewitness Identification: The Importance of Lineup Models,
99 Psych. Bull. 320, 323 (1986).

30 ​Tom Simonite, The Best Algorithms Struggle to Recognize Black Faces Equally, Wired (July 22, 2019),
https://­w ww​.­wired​.­com​/­story​/­best​-­algorithms​-­struggle​-­recognize​-­black​-­faces​-­equally.

31 ​Race bias compared to eyewitness identification.

32 ​Crumpler, supra note 29; Simonite, supra note 30. The Wired article claims that changing the threshold
in this way might violate the equal protection clause, id., but given that the chance of a false match would
be equalized, I have my doubts. Setting the false match rate to be equal is equivalent to ensuring that
“probable cause” for Black suspects means the same thing it does for whites.

33 ​I have called these “crime-out” investigations. Jane Bambauer, Other People’s Papers, 94 Tex. L. Rev. 205
(2015). Christopher Slobogin calls these “event-driven” investigations. Christopher Slobogin, Government
Data Mining and the Fourth Amendment, 75 U. Chi. L. Rev. 317, 331, 338–40 (2008); Christopher Slobogin,
Policing, Databases, and Surveillance, 18 Criminology, Crim. Just., L. & Soc. 70 (2017).

34 ​See, e.g., Tim Cushing, Government’s “Reverse” Warrant Rejected by Two Consecutive Federal Judges,
Techdirt (Sept. 8, 2020), https://­w ww​.­techdirt​.­com​/­articles​/­20200902​/­11565245234​/­governments​-­reverse​
-­warrant​-­rejected​-­t wo​-­consecutive​-­federal​-­judges​.­shtml.

35 ​This is all the more true if police, too, are under greater surveillance through body-worn cameras.
See Morgan C. Williams Jr. et al., Body-Worn Cameras in Policing: Benefits and Costs (National Bureau of
Economic Research. Working Paper No. 28622, 2021).

36 ​Jeffrey Bellin has suggested that Fourth Amendment jurisprudence should mediate rules related to
searches based on the severity of a crime. Jeffrey Bellin, Crime-Severity Distinctions and the Fourth Amendment:
Reassessing Reasonableness in a Changing World, 97 Iowa L. Rev. 1 (2011).

37 ​Carpenter v. United States, 138 S. Ct. 2206 (2018); United States v. Jones, 565 U.S. 400, 418 (2012)
(Alito, J., concurring). These cases suggest tracking individuals’ whereabouts for long periods of time is
an intrusion on reasonable expectations of privacy under Katz v. United States and would not ordinarily be
permissible without a warrant. Note that it is the geolocation information that is most critical. In Maryland
v. King, the Court has suggested that analyzing data solely for the purpose of identification is not intrusive
enough on its own to constitute a violation of reasonable privacy expectations (though the application to
individuals who have not been arrested is not entirely clear). 133 S. Ct. 1958, 1979–80 (2013).

Hoover Institution  •  Stanford University

14

38 ​Leaders of a Beautiful Struggle v. Balt. Police Dep’t, 2 F.4th 330 (4th Cir. 2021) (en banc).

See U.S. v. Yang, 958 F.3d 851 (9th Cir. 2020); Commonwealth v. McCarthy, 142 N.E.3d 1090 (Mass. 2020);
39 ​
Commonwealth v. McCarthy: Massachusetts Supreme Judicial Court Holds That Use of Automated License Plate
Readers May Constitute a Search, 134 Harv. L. Rev. 2887 (2021) (case comment).

40 ​A much more likely scenario, and one which I would defend (but not here), is a system where footage
is stored for a period without identification and used in combination with facial recognition only after
probable cause has been established with respect to a targeted suspect.

Jane Bambauer  •  Facial Recognition as a Less Bad Option

 15

The publisher has made this work available under a Creative Commons Attribution-NoDerivs 4.0 International license.
To view a copy of this license, visit https://­creativecommons​.­org​/­licenses​/­by​-­nd​/­4​.­0.

The views expressed in this essay are entirely those of the author and do not necessarily reflect the views of the staff,
officers, or Board of Overseers of the Hoover Institution.

hoover​.­org

Copyright © 2021 by the Board of Trustees of the Leland Stanford Junior University

27 ​26 ​25 ​24 ​23 ​22 ​21   7 ​6 ​5 ​4 ​3 ​2 ​1

The preferred citation for this publication is Jane Bambauer, Facial Recognition as a Less Bad Option, Hoover Working
Group on National Security, Technology, and Law, Aegis Series Paper No. 2107 (November 4, 2021), available at
https://­w ww​.­lawfareblog​.­com​/­facial​-­recognition​-­less​-­bad​-­option.

Hoover Institution  •  Stanford University

 The Jean Perkins Foundation Working Group on
National Security, Technology, and Law
The Jean Perkins Foundation Working Group on National
Security, Technology, and Law brings together national
and international specialists with broad interdisciplinary
expertise to analyze how technology affects national security
About the Author and national security law and how governments can use
that technology to defend themselves, consistent with
constitutional values and the rule of law.
James E. Rogers College of Law

The group focuses on a broad range of interests, from

surveillance to counterterrorism to the dramatic
University of Arizona

impact that rapid technological change—digitalization,

computerization, miniaturization, and automaticity—are
having on national security and national security law. Topics
include cybersecurity, the rise of drones and autonomous
weapons systems, and the need for—and dangers of—state
JANE BAMBAUER
surveillance. The group’s output will also be published on
Jane Bambauer is a the Lawfare blog, which covers the merits of the underlying
professor of law at the legal and policy debates of actions taken or contemplated
University of Arizona. to protect the nation and the nation’s laws and legal
Her research assesses the institutions.
social costs and benefits
of Big Data and questions Jack Goldsmith is the chair of the National Security,
the wisdom of many Technology, and Law Working Group.
well-intentioned privacy
For more information about this Hoover Institution Working
laws. Her work has been
Group, visit us online at http://­w ww​.­hoover​.­org​/­research​-­teams​
published in leading
/­national​-­security​-­technology​-­law​-­working​-­group.
legal journals.

Hoover Institution, Stanford University Hoover Institution in Washington

434 Galvez Mall 1399 New York Avenue NW, Suite 500
Stanford, CA 94305-6003 Washington, DC 20005
650-723-1754 202-760-3200
Author Name  •  Essay Title