THE FIRST STEP ACT – CONSTITUTIONALIZING PRISON RELEASE POLICIES
Harold J. Krent* and Robert Rucker**

Predictions of future violence play a central role in most systems of criminal justice.
Such assessments help determine the amount of bail imposed, the length of sentence an offender
receives after conviction, the type of prison to which he or she is assigned, and whether offenders
are released before the end date of their sentence. There are both subjective and objective
determinants of future dangerousness, including an offender’s prior record, behavior in prison,
education level, and more. To determine future dangerousness at the beginning of the last
century, we relied upon experts to analyze those factors. Indeed, legislation predicated parole on
the belief that criminal justice experts could ascertain who should be released early because they
no longer posed a threat to society.1
Given the wide latitude afforded to such predictions, criminologists grew concerned that
similarly situated offenders were not being treated alike.2 Moreover, others critiqued parole on
the ground that the rehabilitation goal was not realistic. Disenchantment with the subjectivity in
the process led Congress to end the federal parole system through the Comprehensive Crime
Control Act of 1984.3 Even when there was little chance of recidivism, inmates had to serve

*

Professor, Chicago-Kent College of Law.
3L Student, Chicago-Kent College of Law. We would like to thank Jacob Earl, also a 3L at Chicago-Kent, for his
excellent research that served as a springboard for this article.
**

1

See, e.g., Pub. L. 61-259 (1910) (legislation establishing federal parole).
For example, as a result of the subjectivity, people of color received higher bail, longer sentences, and more
restrictive prison sentences. See generally David Arnold, Will Dobbie, & Crystal Yang, Racial Bias in Bail
Decisions, 133 THE QUARTERLY J. of Economics (2018) (discussing racial bias in bail decisions); see also M. Marit
Rehavi & Sonja B. Starr, Racial Disparity in Federal Criminal Sentences, 122 J. P. ECON., 1320, 1322-23 (2014)
(discussing observable racial disparity in federal sentencing).
3
Pub. L. 98-473 (1984).
2

1

their complete sentences, shortened only modestly through good time credits. The Act
manifested Congress’s turn away from rehabilitation as a central pillar of the federal criminal
justice system.4
Criminologists outside of the parole context recently turned to more objective
assessments of future dangerousness, based not upon the judgment of experts, but rather upon a
computerized assessment of factors gleaned from field studies of large numbers of offenders.5
The goal of these efforts was to create greater uniformity and minimize the possibility of racism
in determining the length of a sentence or amount of bail. Accordingly, in most states today, the
prediction of future dangerousness turns not on evaluation of the particular offender alone, but
on assessment of whether offenders with similar characteristics in the past have re-offended. In
particular, a number of states now rely on algorithmic and Artificial Intelligence (AI) systems to
fine tune the assessment of future dangerousness.6 These states have deployed such algorithmic
tools as a means to inform various decisions in the criminal justice process, including both bail
and sentencing. Although several of the systems deployed have had a shaky start due to
questionable methodologies, they hold promise for more uniform and less biased results.7
In the First Step Act of 2018, Congress directed the Department of Justice (DOJ or the
Department) to develop a tool, subsequently called Prisoner Assessment Tool Targeting
Estimated Risk and Needs (PATTERN), to make such an algorithmic assessment of recidivism
risk based on static factors such as the nature of the underlying offense, prior substance abuse,

4

There has been a corresponding move away from subjectivity in the state systems as well, with many states
abolishing parole or predicating it on objective factors. See Kimberly Thomas & Paul Reingold, From Grace to
Grids: Rethinking Due Process Protections for Parole, J. J. CRIM. L & CRIMINOLOGY 218, 239-44 (2017).
5
Id. at 244.
6
See generally Racial Disparity in Federal Criminal Sentences, supra note 2 (discussing such usage); see also text
accompanying notes 144-49 infra (discussing the usage of these systems at the state level).
7
See, e.g., Tom Simonite, Algorithms Were Supposed to Fix the Bail System. They Haven't, WIRED (Feb. 19, 2020),
https://www.wired.com/story/algorithms-supposed-fix-bail-system-they-havent/.

2

and education level.8 Congress determined that inmates, dependent on such recidivism
assessment, be permitted to shorten their stay in prison. For instance, by pursuing vocational
courses or by electing to take classes in preventing substance abuse, offenders can now earn
credits to qualify for early release or to garner other privileges.9 Moreover, the Act facilitates
release for (almost) all offenders by awarding enhanced good time credits.10 The First Step Act
thus links the length of confinement in part to predictions of future crime as in the past, but also
attempts to parlay a prison stay into an opportunity to incentivize offenders to make adjustments
in their lives to minimize the risk of future dangerousness. A generation after Congress in
essence abandoned rehabilitation as a principal goal, rehabilitation once again has become one of
the driving forces of our federal criminal justice system. Many have praised the Act for
shortening prison stays and reintroducing rehabilitation as a goal of our penal system.11
But, commentators to date have not considered that, in revamping criminal justice
policies, the First Step Act may have constitutionalized such early release measures. Unlike in
most state systems that use algorithms as guidelines, the Act dictates that PATTERN alone
determines eligibility for early release – no discretion on the part of prison authorities is
involved. Congress’s decision to base eligibility on an algorithm accordingly raises the critical
question whether Due Process requires that individuals be permitted to show that facts not
captured by the algorithm demand an adjusted outcome. In addition, by creating a system to
encourage offenders to pursue certain opportunities in prison, Congress likely has created an

8

First Step Act of 2018, Pub. L. 115-391, 132 Stat. 5194.
Id.
10
Id.
11
See, e.g., Ames Grawert, What Is the First Step Act — And What’s Happening With It?, BRENNAN CENTER FOR J.
(June 23, 2020), https://www.brennancenter.org/our-work/research-reports/what-first-step-act-and-whats-happeningit (“The law we now know as the First Step Act accomplishes two discrete things, both aimed at making the federal
justice system fairer and more focused on rehabilitation.”).
9

3

entitlement based on liberty interests protected under the Due Process Clause. The First Step Act
tells prisoners that, if they successfully attain certain educational goals, receive psychological
counselling, etc., they will be released early. As a consequence, prison authorities will need to
ensure that prisoners who complete such programs are released early with only narrowly defined
exceptions. Finally, Congress in light of Ex Post Facto principles must respect the enhanced
good-time credit calculation in the Act for individuals who have already committed their
offenses.
In Part I, we trace Congress’s vacillations over the last century in implementing
determinate and indeterminate sentencing systems. Then, we hone in on the innovations of the
First Step Act, both in relying upon algorithms and AI to predict future dangerousness and in
setting incentives to encourage prisoners to pursue measures that will reduce further the chance
of recidivism.
In Part II, we then address the constitutional ramifications of entrenching early release
policies. First, we consider the problem endemic in all governmental benefit systems relying on
algorithms – to what extent can an individual demonstrate that, despite whatever the algorithm
dictates, data specific to the individual warrant a different outcome. The First Step Act presents
one of the first instances in which an algorithm by itself governs eligibility for a government
entitlement. Based on current jurisprudence, we conclude that prison authorities must allow
those prisoners excluded from eligibility based on the algorithm an opportunity, no matter how
truncated, to argue that the risk of recidivism determined by the algorithm needs to be adjusted
given the offender’s specific context. Relatedly, we argue that Due Process dictates that prison
authorities disclose the static inputs that underlie the findings of ineligibility under both the
statute and PATTERN.

4

Next, we analyze the dynamic features of the First Step Act by canvassing the Supreme
Court’s embrace under the Due Process Clause of an “entitlement” system under which
individuals are invited to rely upon government pledges such that the government cannot deny
those benefits without good cause. We apply that entitlement analysis to the First Step Act and
conclude that Congress’s encouragement of prisoners to pursue rehabilitative programming, such
as education, counselling, etc. to reduce recidivism, has created an entitlement. Finally, we turn
to the Ex Post Facto Clause and argue that the enhanced good time credit accumulation policy
(but not the earned credits system) in the First Step Act must be offered to all offenders who
have committed their offenses during pendency of the Act.
We conclude that such constitutionalization of release policies, though likely unintended,
should prove beneficial in striking an enforceable bargain with offenders: if the offenders take
steps to limit the chance of their own future recidivism, they can gain early release. Given the
First Step Act’s reintroduction of rehabilitative goals in the federal prison system, the application
of these constitutional requirements may further the Act’s purpose in seeking to reduce the
likelihood of inmate recidivism prior to reintroduction into society.
I
BRIEF HISTORY OF FEDERAL RELEASE POLICIES
Congress introduced indeterminate sentencing in the early twentieth century, building
upon experiments led by Zebulon Brockway in New York. As superintendent of Elmira
Reformatory, Brockway innovated in predicating early release on inmate education12 and urged

12

Thom Gehring, Zebulon Brockway of Elmira: 19th Century CE Hero, 33 J. CORRECTIONAL EDUCATION 4-5
(1982).

5

that inmates could be rehabilitated in prison.13 Under his approach, volunteer "guardians"
supervised parolees after release and submitted written reports documenting parolees’ behavior
in the community.14 Included in the early release system was a condition that the former inmate
report to the guardian each month.15
Brockway’s fundamental arguments for early release were that (1) indeterminate
sentencing would “provide a release valve for managing prison populations,” and (2) “it would
be valuable in reforming offenders because they would be earning release by demonstrating good
behavior.”16 Later in his career, he drafted New York’s Indeterminate Sentence Law, which
embodied many of his ideas and furthered these two tenets.17 Seventy-eight percent of those
released on parole under the New York system reportedly maintained “self-supported, orderly
lives.”18
At the start of the twentieth century, as rehabilitation theory gained traction, the ideas of
indeterminate sentencing and parole spread widely across jurisdictions. By 1901, twenty states
adopted parole statutes and, in 1910, Congress established the federal parole system.19 Congress
created the National Parole Board at the federal level in 1930, which set forth a uniform
system.20 Ultimately, by 1944, every state had enacted a parole system.21
A. Mechanics of Federal Parole

13

Id. at 5; Probation and Parole: History, Goals, and Decision-Making, L. LIBRARY - AM. L. AND LEGAL INFO.,
https://law.jrank.org/pages/1817/Probation-Parole-History-Goals-Decision-Making-Origins-probation-parole.html
(last visited June 4, 2021).
14
Id.
15
Id.
16
Id.
17
Gehring, supra note 12, at 4.
18
Id. at 5.
19
Probation and Parole: History, Goals, and Decision-Making, supra note 13; Pub. L. 61-259 (1910).
20
Isaac Fulwood, History of the Federal Parole System, U.S. DEP’T OF JUSTICE (May 2003),
https://www.justice.gov/sites/default/files/uspc/legacy/2009/10/07/history.pdf.
21
Probation and Parole: History, Goals, and Decision-Making, supra note 13.

6

When the United States adopted a federal parole system, the law granted individual
prisons significant flexibility and discretion in making the parole determination. Overall, the
1910 statute contained three key determinants:
(1) . . . [an inmate] whose record of conduct shows he has observed the rules of
such institution, and who has served one-third of the total of the term or terms for
which he was sentenced, may be released on parole as hereinafter provided.
(2) . . . each United States penitentiary shall constitute a board of parole for such
prison, which shall establish rules and regulations for its procedure subject to the
approval of the Attorney-General . . ..
(3) That if it shall appear to said board of parole from a report by the proper officers
of such prison or upon application by a prisoner for release on parole, that there is
a reasonable probability that such applicant will live and remain at liberty without
violating the laws, and if in the opinion of the board such release is not incompatible
with the welfare of society, then said board of parole may in its discretion authorize
the release or such applicant on parole, and he shall be allowed to go on parole
outside of said prison, and, in the discretion of the board, to return to his home,
upon such terms and conditions, including personal reports from such paroled
person . . . . (original section numbers omitted) 22
This early parole statute paralleled many of the key ideas of Brockway’s system – federal parole
rested on a subjective assessment of the inmate and imposed conditions such as post-release
reporting. In the 1930s, Congress standardized federal parole by implementing a single parole
board within the Department of Justice, but otherwise left the basic requirements of parole
intact.23 Ultimately, the only threshold requirement for federal parole was that the inmate
complete one-third of his or her sentence before becoming eligible.24
Risk assessment in various forms has long been part of the parole determination in the
federal and state criminal justice systems.25 Prison authorities originally relied mostly on clinical

22

Pub. L. 61-259 (1910).
Fulwood, History of the Federal Parole System, supra note 20, at 1.
24
Pub. L. 61-259 (1910).
25
See, e.g., Charles D. Stimson, The First Step Act’s Risk & Needs Assessment Program: A Work in Progress,
HERITAGE FOUNDATION (June 8, 2020), https://www.heritage.org/sites/default/files/2020-06/LM265_0.pdf; see also
23

7

judgments, particular those of probation officers.26 A generation ago, however, prison
authorities began to rely on more quantitative assessment of risk based on statistical modeling
drawn from evidence gleaned from prior offenders. Static factors such as type of offence
committed, history of substance abuse, and education level could be identified and then analyzed
together to gauge the likelihood of future wrongdoing. Researchers supplemented those
predictions with needs assessments that factored in dynamic factors as well, such as pursuing
vocational opportunities while incarcerated or attending programs in substance abuse prevention
to help ensure productive lives post-release.27
The Model Penal Code published in 1962 highlights the prevailing subjective approach to
parole, including risk assessment. As stated in §305.10, assessing whether an inmate was
suitable for parole turned on a series of largely subjective factors, including:
(1) a report prepared by the institutional parole staff, relating to [the prisoner's]
personality, social history and adjustment to authority, and including any
recommendations which the institutional staff may make;
(2) all official reports of his prior criminal record, including reports and records of
earlier probation and parole experiences;
(3) the pre-sentence investigation report of the sentencing court;
(4) recommendations regarding his parole made at the time of sentencing by the
judge or the prosecutor;
(5) the reports of any physical, mental and psychiatric examination of the prisoner;
(6) any relevant information which may be submitted by the prisoner, his attorney,
the victim of his crime, or by other persons;
(7) the prisoner's parole plan;

Amy B. Cyphert, Reprogramming Recidivism: The First Step Act and Algorithmic Prediction of Risk, 51 Seton Hall.
L. Rev. 331 (2020).
26
Stimson, The First Step Act’s Risk & Needs Assessment Program, supra note 25, at 5; Cyphert, Reprogramming
Recidivism, supra note 25, at 336-37.
27
Cyphert, Reprogramming Recidivism, supra note 25, at 337-38.

8

(8) such other relevant information concerning the prisoner as may reasonably be
available.28
Studies show that, despite the litany of factors noted in the Model Penal Code, there were only a
handful of factors that in practice determined whether parole authorities released an inmate.
Among these key factors were mental fitness, the severity/type of crime, history of other crimes,
sentence length, and behavior while incarcerated.29 Ultimately, however, consideration of these
factors rested on human subjectivity. Through subjective assessments of an inmate, prison
authorities could determine whether such rehabilitation occurred during incarceration, thus
warranting release on parole.
To aid in that determination at the federal level, authorities developed a tool in the 1970s
termed the Bureau Risk Assessment Verification Observation (BRAVO), to provide structure for
offender risk assessment. BRAVO initially served as a classification system for predicting
inmate misconduct during incarceration, which helped authorities assign inmates to an
appropriate security level during their incarceration.30 A revised system (BRAVO-r) added an
assessment of an inmate’s three-year recidivism rate. BRAVO-r included a detailed history of
the offender in comparison to those of others similarly situated, including factors such as age,
substance abuse, history of violence, and nature of the offense, but its details were never
released.31
B. Criticism of Discretionary Parole Systems

28

Model Penal Code § 305.10 (Am. Law Inst., Proposed Official Draft 1962).
Joel Caplan, What Factors Affect Parole: A Review of Empirical Research, 71 FED. PROB. 16, 16-17 (2007).
30
Office of the Attorney General, The First Step Act of 2018: Risk and Needs Assessment System, U.S. DEP’T OF J.,
42 (Jul. 2019).
31
See, e.g., Comment Letter to Department of Justice on PATTERN First Step Act, THE LEADERSHIP CONFERENCE
ON CIVIL AND HUMAN RIGHTS (Sept.3, 2019), https://civilrights.org/resource/comment-letter-to-department-ofjustice-on-pattern-first-step-act/ (discussing the DOJ’s refusal to provide detailed information on BRAVO and
BRAVO-r).
29

9

A generation after establishment of parole, outsiders critiqued the parole system in light
of the excessive discretion exercised by parole authorities.32 Parole rates were never uniform
among the states,33 and questions arose about the ability of parole authorities accurately to
determine which offenders were most likely to recidivate.34
Despite the discretionary nature of the parole determination, many prison authorities
apparently granted parole automatically, seemingly oblivious to the goal of making a case-bycase judgment of the likelihood of rehabilitation. Outside observers came to view parole as an
automatic step in the incarceration process, with few inmates serving a complete sentence,
regardless of the severity of the crime.35 No effort seemingly was made to determine whether
offenders had successfully rehabilitated. This led critics to allege that prison authorities granted
parole routinely to minimize the number incarcerated in their facilities.36 While the parole
system at the federal level was not automatic, federal data show that the vast majority of eligible
inmates were eventually released on parole.37
Moreover, high recidivism rates thereafter generated additional criticism and led many to
second-guess the very premise of rehabilitation. Experts questioned how rehabilitation could
ever be a rational goal given the grim existence within prison walls.38 Some even argued that

32

See, e.g., Thomas & Reingold, supra note 4; Robert W. Kastenmeier & Howard C. Eglit, Parole Release Decisionmaking: Rehabilitation, Expertise and the Demise of Mythology, 22 Am. U. L. Rev. 477. 483-84 (1973).
33
Probation and Parole Systems Marked by High Stakes, Missed Opportunities, THE PEW CHARITABLE TRUSTS, 6,
11-12 (2018).
34
Kastenmeier & Eglit, supra note 32, at 486-87.
35
See MINISTER OF SUPPLY AND SERVICES CANADA, SOME PEOPLE SAY, 1-2 (1987) (archived at
https://www.ncjrs.gov/pdffiles1/Digitization/106142NCJRS.pdf) (discussing the common critiques and criticisms of
the U.S. parole system).
36
Id.
37
Timothy Hughes & Doris James Wilson, Reentry Trends in the U.S., BUREAU OF JUSTICE STATISTICS (last updated
Dec. 3, 2020), https://www.bjs.gov/content/reentry/reentry.cfm.
38
See, e.g., Kastenmeier & Eglit, supra note 32, at 495-97.

10

stays in prison would exacerbate the likelihood of recidivism thereafter,39 given that the skill set
to survive in prison did not translate well to success upon release.40 In fact, despite the success
of the first reformers (such as Brockway), the Bureau of Justices Statistics (BJS) in 1984 found
that only approximately 60% of parolees successfully completed the terms of their release.41
C. The Federal Phaseout of Parole
In the end, the criticisms of parole and reform theory prevailed, at least at the federal
level. In 1984, Congress passed the Comprehensive Crime Control Act,42 which overhauled
administration of the federal prison system. Among its many changes, Congress explicitly
eliminated the federal parole system and adopted supervised release in its place.43 This system of
supervised release required the sentencing court to provide a period of supervised release, which
by statute was not to exceed five years, at the time of sentencing.44 Thus, Congress in essence
delinked any goal of rehabilitation from behavior in prison45—prison authorities were to base

Judge Frankel notoriously opined that “[t]he naïve faith in the present expertise of penologists and parole officials
effectively blots out some of the stark and familiar realities of prisons as they actually function. The notion that the
unrehabilitated prisoner should be denied parole because he needs more treatment is not merely unsupported; it runs
counter to considerable evidence and option concerning the effects of confinement. Taking prisons as they are, and
as they are likely to be for some time, it is powerfully arguable that their net achievement is to make their
inhabitants worse.” Marvin E. Frankel, Criminal Sentences: Law Without Order 34 (1974).
40
See, e.g., David Harding, Jeffrey Morenoff, et. al., Short- and Long-term Effects of Imprisonment on Future
Felony Convictions and Prison Admissions, PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE
UNITED STATES OF AMERICA (Oct. 2, 2017) (discussing the correlation between length of incarceration and
increased recidivism).
41
Probation and Parole 1984, BUREAU OF JUSTICE STATISTICS, 2 (Sept. 1984). While the federal phaseout
39

of parole began with the Comprehensive Crime Control Act that same year, BJS continues to track this
data and, according to its more recent analyses, success rates fell to approximately 45% in the mid-1990s
and have remained relatively stable since. Lauren Glaze and Thomas Bonczar, Probation and Parole in
the United States, 2005, BUREAU OF JUSTICE STATISTICS, 9 (Jan. 1, 2007).
42
Pub. L. No. 98-473, 98 Stat. 1976.
43
H.R. 5773 – Sentencing Reform Act of 1984, LIBRARY OF CONGRESS, https://www.congress.gov/bill/98thcongress/house-bill/5773 (last visited Dec. 1, 2020); Pub. L. 98-473 (1984).
44
18 U.S.C. § 3583.
45
The Sentencing Reform Act of 1984: Principal Features, U.S. SENTENCING COMM’N,
https://www.ussc.gov/research/research-and-publications/simplification-draft-paper-2 (last visited Dec. 3, 2020)
(stating “[the Sentencing Reform Act] instructed the [Sentencing Commission] to ensure that the guidelines reflect
the inappropriateness of using prison sentences to achieve rehabilitative goals.”).

11

release decisions on an inmate’s completion of a percentage of the court-imposed sentence,
reflecting a return to a more “determinate” sentencing structure. The Comprehensive Crime
Control Act marked the formal end of rehabilitation as a goal of the federal prison system.46
D. The First Step Act
The First Step Act of 2018 modified the early release structure in three principal ways.
First, Congress directed that good-time credits be recalculated for all offenders, thus shortening
their likely stay behind prison walls. Section 102(b) of the Act provides that inmates can now
earn up to 54 days for each year of the sentence imposed by the court,47 instead of for each year
of actual time served. Basing the award on the sentence length permits the offender to earn
approximately an additional week of credit per year. Perhaps because of prior adoption of harsh
penalties or due to overcrowding in prisons, Congress embraced a new calculation of good time
credits so as to facilitate earlier release of all inmates, except those serving life imprisonment or
a sentence of less than one year.
Second, Congress jettisoned the pre-1984 subjective determination for determining
inmates’ risk of future dangerousness and replaced it with an objective assessment of whether
offenders with similar characteristics in the past had committed offenses upon release. To that
end, Congress directed the Attorney General to devise a system to assess objectively the
likelihood of recidivism for offenders entering the system.48 Congress specified that the tool

46

Although the Comprehensive Crime Control Act of 1984 abolished parole for new inmates, the federal system has
been in the midst of a decades-long phase out. The U.S. Parole Commission and the system of parole itself has been
extended by statute multiple times since 1984, primarily to serve those inmates who were grandfathered into parole
(i.e., those sentenced prior to the enactment of the Comprehensive Crime Control Act). The most recent extension
occurred in 2018, which extended the life of the federal parole system through 2021. H.R.6896 - United States
Parole Commission Extension Act of 2018, LIBRARY OF CONGRESS, https://www.congress.gov/bill/115thcongress/house-bill/6896 (last visited Dec. 1, 2020).
47
18 U.S.C. § 3624(b)(1).
48
Nathan James, The First Step Act of 2018: An Overview, CONG. RSCH. SERV., R45558, 1 (2019).

12

separate offenders into categories of “minimum, low, medium, or high risk of recidivism,”49 and
that only those with a modest likelihood of recidivism could be released before the end of their
sentences: Congress provided that the algorithm be used to “determine whether a prisoner is
ready to transfer to prelease custody.”50 For perhaps the first time, a legislature provided that
eligibility for early release would be governed solely by administration of an algorithm.
To effectuate that directive, DOJ developed an assessment tool called the Prisoner
Assessment Tool Targeting Estimated Risk and Needs (PATTERN).51 In meeting the deadline
set by Congress for the development of this system for evaluating recidivism risk, the
Department of Justice acknowledged the limitation of the available data and time for validation
of the tool.52 Working within these constraints, the Department of Justice developed the initial
version of PATTERN using seven years of federal prison data and designed the tool to be an
effective predictor of recidivism over the initial three-year period after release of an inmate.53
Notably, the First Step Act requires re-validation of the tool annually to add to the initial dataset,
estimate its predictive performance, and make modifications as necessary.54
As in other contexts, PATTERN’s AI-like approach promises efficiency, objectivity, and
consistency.55 Essentially, artificial intelligence seeks to transform decisions that were

49

18 U.S.C. § 3632(a)(1).
18 U.S.C. 3632(a)(6).
51
Office of the Attorney General, The First Step Act of 2018: Risk and Needs Assessment System, U.S. DEP’T OF
JUSTICE (Jul. 2019).
52
Id. at 70.
53
Id. at 84.
54
Id. at 84-85.
55
While there is ongoing debate as to the exact dividing line between algorithms and Artificial Intelligence, we
accept the general proposition that AI is “the ability of a machine to perceive and respond to its environment
independently and perform tasks that would typically require human intelligence and decision-making processes, but
without direct human intervention.” Christopher Rigano, Using Artificial Intelligence to Address Criminal Justice
Needs, 280 NIJ Journal (2019), https://www.ncjrs.gov/pdffiles1/nij/252038.pdf. Regardless of PATTERN’s current
application, Dr. Mir Emad Mousavi has eloquently explained the relationship between an algorithm and AI as
equivalent to “the relationship between ‘cars and flying cars.’” Kaya Ismail, AI vs. Algorithms: What's the
Difference?, CMS Wire (Oct. 26, 2018), https://www.cmswire.com/information-management/ai-vs-algorithms50

13

previously subjective into those capable of objective resolution based on data and inputs to
various algorithms.56 To that end, PATTERN implemented a classification system which, as
noted, sought to classify inmates based on an objective assessment of the potential for
recidivism. As discussed below, prison authorities are to use that classification system in
providing for early release.
The initial release of PATTERN started a 180-day statutory clock for all federal inmates
to be assessed for recidivism risk, with each inmate to be assigned a level of high, medium, or
low risk.57 By statute, the risk level of each federal inmate is then required to be re-assessed on a
bi-annual basis, with new inmates to be initially assessed at the time of intake58 -- a formidable
task given the well over 150,000 individuals in federal custody.59
As with prior tools to predict the risk of recidivism, PATTERN considers static factors
that cannot be changed by the inmate. These factors are:
1. The age of the inmate at the time of assessment.
2. Whether the crime of conviction that resulted in the current incarceration was
violent.
3. Whether the inmate is identified as a sex offender under the definition used by
the Sex Offender Registration and Notification Act.
4. A criminal history score based on BRAVO.60

whats-the-difference/). AI tools almost inherently rely upon algorithms. While PATTERN is by no means currently
a fully autonomous AI or machine-learning algorithm today, its algorithm nonetheless provides the foundation for
greater application as a more AI-like tool, including for example, automatic updating independent of human
intervention. See also Cary Coglianese and Lavie Ben Dor, AI in Adjudication and Administration, BROOKLYN L.
REVIEW (unpublished 2020) (discussing the current limitations of tools such as PATTERN).
56
See generally, Driving Impact at Scale from Automation and AI, MCKINSEY (Feb. 2019),
https://www.mckinsey.com/~/media/McKinsey/Business%20Functions/McKinsey%20Digital/Our%20Insights/Driv
ing%20impact%20at%20scale%20from%20automation%20and%20AI/Driving-impact-at-scale-from-automationand-AI.ashx (discussing the promises of AI).
57
The First Step Act of 2018: Risk and Needs Assessment System, supra note 30, at 71.
58
Id.
59
John Gramlich, Under Trump, the Federal Prison Population Continued its Recent Decline, PEW RESEARCH
CENTER (Feb. 17, 2021), https://www.pewresearch.org/fact-tank/2021/02/17/under-trump-the-federal-prisonpopulation-continued-its-recent-decline/.
60
This score is based on BRAVO which is discussed supra p. 9; See Office of the Attorney General, The First Step
Act of 2018: Risk and Needs Assessment System UPDATE, U.S. DEP’T OF JUSTICE, 7 (Jan. 2020).

14

The current revision of the PATTERN algorithm also includes the following dynamic factors,
which are weighed differently in the male and female populations:
1. A count of the total number of infractions, resulting in a guilty finding, that the
inmate incurred during the current incarceration.61
2. A count of the serious and violent infractions, resulting in a conviction, that the
inmate has incurred during the current incarceration.62 A serious or violent
infraction is defined to be in the top two severity levels of the Bureau of Prison’s
Inmate Discipline Program.63 This includes, but is not limited to, such infractions
as homicide, assault, escape, and fighting.64
3. A score associated with the amount of time that the inmate has been infraction
free during their current period of incarceration.65
4. A score associated with the amount of time that the inmate has spent free of
serious and violent infractions during their current period of incarceration.66
5. A measure of the number of qualifying programs which have been completed by
the inmate. These range from educational and vocational programs to drug
treatment and parenting programs.67
6. Participation by the inmate in work programming during their current
incarceration.
7. A need-based factor for the inmate’s participation in a drug treatment program
while incarcerated.
8. The inmate’s compliance with financial responsibility. For example, their
willingness to use income earned during incarceration as payment toward victim
restitution.
9. The inmate’s history of violence, factoring in the elapsed time since the violent
behavior.68
10. The inmate’s history of escapes, factoring in the elapsed time since the escape.
This score is based on BRAVO.69
11. An education score based on the inmate’s completion of High School education
or a GED. This score is based on BRAVO.70

61

The First Step Act of 2018: Risk and Needs Assessment System, supra note 45, at 53.
Id.
63
Id.
64
Id.
65
The First Step Act of 2018: Risk and Needs Assessment System UPDATE, supra note 45, at 37.
66
Id.
67
Id.; see also First Step Act Approved Programs Guide, DEP’T OF JUSTICE (Oct. 2020).
68
The First Step Act of 2018: Risk and Needs Assessment System, supra note 30 at 45-49 (This measure is borrowed
from the Bureau Risk and Verification Observation (BRAVO), which is the bureau of prison’s current classification
system for predicting serious misconduct and assigning inmates to the appropriate security level during their
incarceration. This assessment was created in the 1970s and has been updated periodically).
69
Id.
70
Id.
62

15

Third, in addition to enhancing good time credits and calling for an AI assessment of risk
recidivism, the Act implemented a system of “earned” credits. The Act directs the Bureau of
Prisons to offer inmates “recidivism reduction programs,” with most prisoners who successfully
complete recidivism reduction programming eligible to earn up to 10 days of time credits for
every 30 days of program participation.71 Minimum and low-risk prisoners under PATTERN
who successfully complete recidivism reduction or productive activities and whose assessed risk
of recidivism has not increased over two consecutive assessments earn an additional five days of
time credits for every 30 days of successful participation.72 DOJ directed that inmates who have
a medium to high risk of recidivism receive priority for recidivism reduction programs, while
those with minimum or low risk be afforded opportunities through work and other programs to
earn credits. Inmates released early are to serve out their sentences in home confinement or on
supervised release, at least until 85% of the original sentence is served.73
The First Step Act, however, precludes early release for a significant swathe of offenders,
including those committing particularly violent offenses, sexual offenses, kidnapping, terrorist
activities and others,74 presumably on the ground that early release would unjustifiably
jeopardize citizen safety.75 Nonetheless, all such inmates can participate and earn other “credits,’
such as greater phone privileges, email access, visitation rights and commissary purchases.
In constructing the classification system, DOJ focused on the “average rate” of
recidivism for the Bureau of Prisons population and set cutoff points based on various deviations
from that base rate. The final score reflects a combination of scores based on (1) the likelihood

71

First Step Act of 2018, Pub. L. 115-391 § 101(d)(4).
18 U.S.C. § 3624(b).
73
Id.
74
18 U.S.C. § 3632(d)(4)(D).
75
See The First Step Act of 2018: Risk and Needs Assessment System UPDATE, supra note 45, at 19.
72

16

of recidivism in general and (2) the likelihood of violent recidivism.76 The Department set the
highest risk category roughly 2/3 above the base rate for general recidivism and just over twice
the base rate for violent recidivism, with the minimal risk category set at half the base rate for
general recidivism and 1/3 the base rate for violent recidivism.77 DOJ explained that it chose
those cutoffs to maximize “the number of inmates eligible to earn early release time credits . . .
while also considering public safety and the risk of recidivism.”78
Those inmates who have accrued sufficient time credits and fall into either the minimal or
low risk category based on the PATTERN score are then eligible for early release. As noted,
inmates within the medium and high risk categories can, however, participate in the recidivism
reduction/incentive programs to lower the score calculated under PATTERN and qualify for a
lower risk classification.79 Then, once the inmate has accrued sufficient time credits, the inmate
can become eligible for early release.
According to the DOJ, its deployment of PATTERN allows for “99 percent of
offenders”80 potentially to qualify for early release. In other words, application of PATTERN
makes only a subset of inmates ineligible for early release. This is by virtue of the fact that an
inmate’s score based on PATTERN’s static factors alone can be so high as to preclude a
sufficient classification change through the dynamic factors.

76

Id. at 51-52.
Id. at 50-52.
78
Id. at 51.
79
Id. at 58.
80
Id. at 57. For a critique that DOJ’s estimate is wildly optimistic and discussion that certain inmates will be
perpetually classified as ineligible for early released under PATTERN by virtue of the static factors alone, regardless
of what remains subject to the control of the inmate under the dynamic factors, see Statement of Sarah Anderson,
FREEDOM WORKS (Sept. 11, 2019), https://fw-d7-freedomworksorg.s3.amazonaws.com/09_11_2019_FreedomWorks_Anderson_DOJ.pdf.
77

17

Through the First Step Act, Congress thus sought to leverage the prison stay into an
opportunity to work with inmates to facilitate their reentry into society. In doing so, the First
Step Act has reintroduced rehabilitation as a central goal of our federal criminal justice system.
II
CONSTITUTIONAL RAMIFICATIONS
The First Step Act may well usher in a salutary change in the nation’s approach to
criminal justice, dialing back the clock to focus once again on potential rehabilitation while
an offender is incarcerated. Its implementation, however, raises three sets of constitutional
issues. First, conditioning liberty on the basis of an algorithm’s prediction of recidivism
resembles the use of a conclusive presumption, which should entitle the offenders to
present evidence that the prediction is not accurate as applied to their own circumstances.
Second, offering offenders early release predicated on satisfactory completion of education
and treatment programs creates a liberty interest, which the prison authorities must then
respect. Third, at least part of the First Step Act must be seen to amend the governing
sentencing structure and, thus, consistent with the Ex Post Facto Clause, cannot be altered
for those who thereafter commit federal offenses.
A. Due Process
1. Right to Individual Determination of Future Dangerousness
Congress need not grant early release at all, or it could grant parole across the board to all
offenders who serve three-quarters of their sentences. Indeed, Congress in the First Step Act
determined that those convicted of delineated crimes such as kidnapping and murder, no matter

18

the circumstances, are ineligible for early release.81 Offenders therefore should be able to
challenge whether they in fact committed a listed crime,82 but not introduce evidence that,
despite committing the crime, they pose little risk of future dangerousness. Inmates have no
right to a due process hearing when Congress itself determines when release is appropriate or
leaves that determination to the unfettered discretion of prison authorities.83 For years, Congress
and state legislatures routinely left the release decision to “the judgment” of parole authorities.
However, once Congress determines that release turns on evaluation of an objective
factor – here, the risk of future dangerousness – then Due Process is triggered. The Supreme
Court has held that, if the government creates an expectation that an individual will be granted
parole absent misconduct or a finding of future dangerousness, the individual is entitled to a
hearing to determine whether the statutory condition has been satisfied. In Greenholtz v. Inmates
of Nebraska Penal and Correctional Complex84 the Court stated that statutory and regulatory
language can create an “expectancy of release . . . entitled to some measure of constitutional
protection.”85 The Court continued that “the most common manner in which a State creates a
liberty interest is by establishing ‘substantive predicates’ to govern official decision-making . . .
and, further, by mandating the outcome to be reached upon a finding that the relevant criteria
have been met.”86 Conversely, if a statute confers unbridled discretion upon decisionmakers to

81

See text accompanying notes 74-75 supra.
See text accompanying notes 156-63 infra.
83
Greenholtz v. Inmates of Nebraska Penal and Correctional Complex, 442 U.S. 1, 7 (1979).
84
Greenholtz, 442 U.S. at 7-9.
85
Id. at 12. The Court in Board of Pardons v. Allen, 482 U.S. 369 (1987), similarly held that mandatory language in
statutes and regulations governing parole could create a liberty interest.
86
Greenholtz, 442 U.S. at 12-14.
82

19

determine whether to grant parole, no liberty interest is created, and the Due Process Clause is
not triggered.87
In deciding whether a decision to revoke parole similarly triggered a protected liberty
interest, the Court in Morrissey v. Brewer88 explained that “[t]he parolee has been released from
prison based on an evaluation that he shows reasonable promise of being able to return to society
and function as a responsible, self-reliant person. Subject to the conditions of his parole, he can
be gainfully employed and is free to be with family and friends and to form the other enduring
attachments of normal life . . . The parolee has relied on at least an implicit promise that parole
will be revoked only if he fails to live up to the parole conditions.”89
Moreover, in Wolff v. McDonnell, the Supreme Court applied the liberty interest rationale
to a Nebraska statute affording inmates the ability to earn good time credits, which by statute
could only be revoked for misconduct.90 The Court held that the statute created a liberty interest
through its conferral of a benefit, the deprivation of which would constitute a significant loss to
the inmate.91 At the required hearing, inmates can present evidence that they had not engaged in
misconduct and therefore that their good-time credits should be restored. As the Court held,
credits already earned could not be revoked by means not delineated in the statute as the inmates
were entitled to the application of the law as written.92

See Kentucky Dep’t of Corrections v. Thompson, 490 U.S. 454, 462 (1989) (holding that prison administrators’
guidelines for who could visit prisoners was discretionary).
88
Morrissey v. Brewer, 408 U.S. 471 (1972).
89
Id. at 482.
90
Wolff v. McDonnell, 418 U.S. 539, 558 (1974).
91
Id.
92
Id.
87

20

The Court since has scaled back its liberty interest analysis in the prison context. In
Sandin v. Conner93 the Court held that only restrictions of liberty that result in “atypical and
significant hardship on the inmate in relation to the ordinary incidents of prison life” trigger due
process scrutiny.94 In Sandin, itself, the Court held that regulatory restrictions on when an
inmate could be placed in thirty days’ segregation for misconduct did not give rise to a liberty
interest given that prison authorities not infrequently need to place inmates in segregation for a
variety of reasons. Although prisoners now cannot claim a liberty interest in being guaranteed a
particular exercise period or nutritious meals, irrespective of mandatory statutory or regulatory
language,95 Sandin should not affect prisoner claims focused on the liberty at stake with earned
credits. Credits equate to earlier release from prison confinement, implicating the most basic
understanding of “liberty.”96
Given Congress’s decision to permit early release only for those with a minimal or low
risk of recidivism, can DOJ through PATTERN exclude inmates from eligibility without holding
any type of hearing?97 Courts have held that statutory or regulatory creation of “fixed eligibility
criteria” give rise to entitlements, even if another component of the broader regulatory scheme
remains discretionary. For instance, applicants for public housing may enjoy an entitlement
based on existence of objective regulatory criteria such as financial need, even if the availability

93

Sandin v. Conner, 515 U.S. 472 (1995).
Id. at 484.
95
The Court in Sandin noted that such mandatory language may in fact benefit prisoners by permitting prison
authorities to impose uniform goals without thereby constitutionalizing procedures. Prison authorities have reacted
to the liberty interest line of cases by modifying mandatory language in regulations to make administration of
prisons more discretionary. See, e.g., Harold J. Krent, Reviewing Agency Action for Inconsistency with Prior Rules
and Regulations, 72 CHI-KENT L. REV. 1187, 1207-10 (1997).
96
Courts on occasion have applied the Sandin limitation expansively. See, e.g., Chappell v. Mandeville, 706 F.3d
1052 (9th Cir. 2013) (six-day confinement on contraband watch during which inmate was shackled, closely
monitored, and had no mattress did not state a Due Process claim).
97
See also Prisoner Assessment Tool Targeting Estimated Risk and Needs (PATTERN) Interactive Tool, URBAN
INSTITUTE (Sept. 4, 2019) https://apps.urban.org/features/risk-assessment/ (explaining that those in high risk
category of PATTERN are not eligible for early release).
94

21

of housing is not guaranteed.98 PATTERN therefore represents one of the first instances in
which a legislature has decided that eligibility for early release be governed exclusively by an
algorithm.
Although the algorithmic determination may be far more accurate than prior subjective
assessments as to future dangerousness, PATTERN may nevertheless omit key variables. For
instance, an individual’s marriage after committing an offence might make it less likely that
recidivism will occur, as can a severe illness, or in an extreme case, a sex offender may have had
an operation to blunt his or her sexual drive.99 Alternatively, the material prospects of an
offender’s family may improve to the point where there is less risk of recidivism upon release. 100
Must the decisionmaker factor in those developments, which evidently are not captured by
PATTERN, when determining eligibility for early release?101
a) Right to Present Individualized Information in other Contexts
The Wisconsin Supreme Court in State v. Loomis102 considered a comparable Due
Process challenge to the use of an algorithm to predict dangerousness, which arose out of the
trial court’s refusal to permit probation for Mr. Loomis based in large part on an algorithmic
score. Loomis had pleaded guilty to operating a vehicle without the owner’s permission and

98

Ressler v. Pierce, 692 F.2d 1212 (9th Cir. 1982) (recognizing entitlement in public housing context); See also
Kapps v. Wing, 404 F.3d 105, 114 (2d Cir. 2004) (holding that, by defining eligibility criteria for utility benefits
should federal aid be forthcoming, New York had created an entitlement).
99
See, e.g., K.S. Kendler, S. L. Löhm, et. al., The Role of Marriage in Criminal Recidivism: a Longitudinal and Corelative Analysis, 26 EPIDEMIOLOGY & PSYCHIATRIC SCI. (2017) (discussing the relationship between marriage and
reduction in recidivism). For a discussion of chemical castration, see Alan Blinder, What to Know about the
Alabama Chemical Castration Law, NEW YORK TIMES (June 11, 2019),
https://www.nytimes.com/2019/06/11/us/politics/chemical-castration.html.
100
See, e.g., Kristy Holfreter, Michael Reisig, & Merry Morash, Poverty, State Capital, and Recidivism Among
Women Offenders, 3 Criminology & Public Policy, 185-208 (2008) (discussing the relationship between poverty
status and recidivism).
101
Over time, the algorithm might include such factors in its assessment.
102
State v. Loomis, 881 N.W.2d 749 (Wisc. 2016).

22

attempting to flee from a traffic officer. In a post-conviction proceeding, an expert witness
testified for Loomis that reliance on the tool at issue, Correctional Offender Management
Profiling for Alternative Sanctions (COMPAS), posed a significant risk of overpredicting risk
and therefore biased the sentencing court’s judgment.103 The Wisconsin Supreme Court upheld
the sentence, reasoning that (1) the trial court relied on more than the algorithm in rejecting
probation, and (2) the information upon which the algorithm was based was in the public
domain, subject to Loomis’ review. Nevertheless, in dicta, the Court stated that reliance on the
algorithm could not be “the determinative factor in deciding whether an offender can be
supported safely and effectively in the community.”104
The First Step Act’s reliance on PATTERN to determine eligibility for early release
brings the Loomis Court’s dicta into sharp focus. Although Congress need not provide for early
release, if it determines that release will only be permitted for those who statistically are unlikely
to commit a new offense, must it then permit individualized consideration despite what the
algorithm concludes? The algorithm resembles a conclusive (or irrebuttable) presumption.
Several analogies are informative.
First, in cases such as Cleveland Board of Education v. LaFleur,105 the Supreme Court
held that, when significant liberty interests are at stake, the government cannot condition that
liberty on conclusive presumptions. There, the school board had long followed a rule that only
“fit” teachers could serve in the classroom, and also had determined that all pregnant teachers
were no longer “fit” several months before the end of their term of pregnancy. The Court held

103

COMPAS is a computerized tool developed in the private sector and is generated from a 137-item questionnaire
filled out by prison authorities. For a brief discussion of COMPAS, see infra text accompanying notes .
104
Loomis, 881 N.W.2d at 769.
105
Cleveland Board of Education v. LaFleur, 414 U.S. 632 (1974).

23

that such a presumption was unconstitutional, and that the school district had to consider
individualized medical determinations rather than relying on the objective determination of
fitness tied to a specific moment in the term of pregnancy.106 The government’s interest in
“speed and efficiency” did not outweigh the right to individualized consideration.107 Although
the Court has since signaled that the irrebuttable presumption approach cannot be extended
wholesale into the social welfare state,108 the Court has never disavowed its core, namely that
when significant liberty is at stake, the government cannot cutoff the right of those involved to
present information about their own circumstances when liberty is predicated on regulatory
factors such as being “fit” for work.109
Second, consider an offender’s constitutional right to present mitigating factors in a
capital punishment case. The Supreme Court has held that an offender enjoys the right to present
any relevant mitigating evidence before the sentence is imposed. In Lockett v. Ohio110 the Court
stated that “the sentencer in all but the rarest kind of capital case cannot be precluded from
considering as a mitigating factor any aspect of a defendant’s character and circumstances of an
offense that the defendant proffers as a basis for a sentence less than death.”111 Although the
Court pegged its analysis on the Eighth Amendment, it later clarified that Due Process requires

106

See id. at 644-46. See also Vlandis v. Kline, 412 U.S. 441 (1973) (holding unconstitutional an irrebuttable
presumption that those students applying to Connecticut universities from out of state remained out of state residents
throughout their college years).
107
LaFleur, 414 U.S. at 646-47.
108
See Weinberger v. Salfi, 422 U.S. 749, 777 (1975) (“the question raised is not whether a statutory provision
precisely filters out those, and only those, who are in the factual position which generated the congressional concern
reflected in the statute …. The question is whether Congress, its concern having been reasonably aroused by the
possibility of an abuse which it legitimately desired to avoid, could rationally have concluded both that a particular
limitation or qualification would protect against its occurrence, and that the expense and other difficulties of
individual determinations justified the inherent imprecision of a prophylactic rule”).
109
See also United States v. Klein, 80 U.S. 128 (1871) (holding unconstitutional congressional determination that
accepting a pardon after the Civil War conclusively demonstrated disloyalty).
110
Lockett v. Ohio, 438 U.S. 586, 604 (1978).
111
Id.

24

individualized consideration as well.112 Lower courts soon thereafter extended the requirements
of individualized consideration to other settings, such as an individual’s right to present
mitigating information before being remanded back to prison from parole. If the decision to
revoke parole is not automatic113 but rather requires assessment of future dangerousness, Due
Process mandates individualized consideration. In John v. U.S. Parole Commission,114 for
example, the Ninth Circuit held that Due Process affords parole offenders the right to introduce
information that might bear upon the decision to reincarcerate the individual. Given that the
parole authority must “[make] a prediction as to the ability of the individual to live in society
without committing antisocial acts,” the individual “is entitled to identify circumstances in
mitigation of his violation so that he might demonstrate to the Commission that parole revocation
was an inappropriate disposition.”115 The liberty interest in staying out of prison dictated
individualized consideration before remand back to prison. To the extent that an algorithm cuts
off that opportunity, Due Process may be violated.116
The irrebuttable presumption and the right to present mitigating evidence contexts are
specialized applications of more basic administrative law doctrine recognizing that, when
government decisionmakers make individualized decisions pursuant to objective legislative
factors, they can rely upon generalized information but must permit some room for

112

See Skipper v. South Carolina, 470 U.S. 1, 8 (1986).
See Pickens v. Butler, 814 F.2d 237, 240 (5th Cir. 1987) (revocation automatic upon commission of a felony,
therefore no individualized consideration required).
114
John v. U.S. Parole Commission, 122 F.3d 1278, 1282 (9th Cir. 1997).
115
Id. at 1282 (citation omitted); see also Caton v. Smith, 486 F.2d 733, 735-36 (7th Cir. 1973) (similarly holding
right to present mitigating evidence); Preston v. Piggman, 496 F.270, 274 (6th Cir. 1974) (same); Williams v.
Johnson, 171 F.3d 300, 307 (5th Cir. 1999) (same); cf. Kelly v. Parole Board, 2017 Mich. App. LEXIS 1240 *36
(affirming right of inmate to see file before parole hearing).
116
Courts have long mandated that offenders be permitted to introduce a wide range of individuated information
before sentence imposed. See, e.g., Williams v. New York, 337 U.S. 241, 247 (1949) (“And modern concepts
individualizing punishment have made it all the more necessary that a sentencing judge not be denied an opportunity
to obtain pertinent information . . . . The belief no longer prevails that every offense in a like legal category calls for
an identical punishment without regard to the past life and habits of a particular offender.”).
113

25

individualized consideration. In Heckler v. Campbell,117 at stake was the Department of Health
and Human Services’ creation and use of a grid to determine whether an applicant for disability
could perform any gainful work in the economy. Instead of determining whether such work was
available based on an in-depth analysis of the claimant’s medical history, residual capacity, and
the demands of particular jobs in the economy, the agency conducted a rulemaking to establish,
based on countless prior determinations, whether individuals with particular ages, education
level, and medical limitations could find jobs.118 Based on that rulemaking, the grid resembled
an unsophisticated pre-AI framework based not explicitly on the characteristics of the individual
claimant, but rather on data culled from countless similarly placed claimants.119
In upholding use of the grid, the Court noted that, as with the use of PATTERN, relying
on the grid would enhance uniformity and objectivity.120 Nonetheless, the Court explained that
the agency under the system must “assess each claimant’s individual abilities . . . on the basis of
evidence adduced at a hearing. We note that the regulations afford claimant ample opportunities
to present evidence relating to their own abilities and offer evidence that the guidelines do not
apply to them.”121 When Congress directs that the administrative determination turns on factors
specific to an individual, an algorithm – just like any other rule – cannot preclude all opportunity
to present individualized information. In accordance with Heckler v. Campbell, the question in

117

Heckler v. Campbell, 461 U.S. 458 (1983).
Id. at 461-62.
119
Id.
120
Id. at 468; see also Am. Hosp. v. NLRB, 499 U.S. 616, 622 (1991) (“the decisionmaker has the authority to rely
on rulemaking to resolve certain issues of general applicability unless Congress clearly expresses and intent to
withhold that authority).”
121
Campbell, 461 U.S. 458, 467. (The ability of agencies to craft rules of general applicability plainly limits
individuals’ ability to present information to protect their entitlements.). See also Daniel J. Rodriguez, Whither the
Neutral Agency? Rethinking Bias in Regulatory Administration, 69 BUFF. L. REV. 375, 427-28 (2021).
118

26

considering the Department’s use of PATTERN is whether that requirement of individualized
consideration, despite the algorithm, is constitutionally required.122
b) Eligibility Determination under PATTERN
There is little question but that a substantial liberty interest is at stake in determining
eligibility for early release.123 Congress itself determined that all federal inmates, with the
exception of those committing one of a long list of offenses, would be eligible for early
release.124 To that end, it called for an objective determination of recidivism risk for eligible
inmates through PATTERN. Congress directed DOJ to create classifications of high, medium,
low and minimal risk,125 with those in the medium and high category having the opportunity to
lower their classification through the dynamic factors. DOJ, in turn, implemented Congress’s
delegation to preclude early release for those with an exceptionally high risk of recidivism, who
mathematically could not qualify for early release even if they were able to pursue the recidivism
reduction programs. Thus, based solely on “static” factors under PATTERN, DOJ has
determined that a subset of the inmate population is not eligible for early release. The group
excluded by virtue of PATTERN should, however, enjoy a limited right under Due Process to
provide information unique to them as to why the exclusion overstates their risk for recidivism.
For an analogy, assume that individuals with a property interest in government
employment are discharged based on metrics culled by the employer from an algorithm
indicating that the employees’ work was substandard. Even if the employer’s algorithm violates

122

For an examination of the use of AI generally in the federal government, see generally David F. Egstrom, Daniel
E. Ho, Catherine M. Sharkey, Mariano-Florentino Cuellar, Government by Algorithm – Artificial Intelligence in
Federal Administrative Agencies (ACUS Report 2020), https://www.acus.gov/report/government-algorithmartificial-intelligence-federal-administrative-agencies.
123
Congress has recognized a liberty interest in a pre-parole program. Young v. Harper, 520 U.S. 142, 143 (1997).
124
18 U.S.C. § 3632(d)(4)(D).
125
Id. at § 101.

27

no regulation or collective bargaining provision, the employees should nonetheless be able to
introduce information that the algorithm for unique reasons did not capture the realities of their
employment efforts.126 When an entitlement is at stake, across the board rules cannot cut off all
right to present individualized information relevant to the entitlement determination.
Thus, no matter that PATTERN limits the magnitude of error in determining the risk of
future dangerousness, the liberty interest implicated militates for at least an informal opportunity
for those whose liberty is at stake to present information that may well be unique to them. An
algorithm cannot fully determine whether an entitlement should be granted. Viewed through this
lens, although AI may end up playing a revolutionary role in making governmental decisions
more accurately and efficiently, individuals whose liberty interests are affected should be entitled
to present individualized information not captured by the algorithm. Thus, those inmates
excluded from the possibility of earned credits by virtue of their PATTERN score should be able
to present individualized information to the decisionmaker that may, in the future, unlock the
doors to the prison before expiration of the original sentence.
Inmates might argue with some force not only that factors that the algorithm omitted
caused them to be assessed for too great a recidivism risk, but also that factors were weighed in
such a fashion that unfairly prejudiced them. But, general rules may always cause some
unfairness in particular cases – the virtue of a general rule is to limit discretion and try to ensure
that likes are treated alike. The very function of AI and algorithms (especially machine-learning

126

In Houston Fed. of Teachers Local 2415 v. Houston Independent School District, 251 F. Supp. 3d 1168, 1178-81
(S. D. Tex. 2017), the court considered a Due Process challenge to use of an algorithm to assess teacher
effectiveness. Although the court stated that it was likely to rule that Due Process required disclosure of the
algorithm itself to enable the teachers to attack it more effectively, it could as well have concluded that Due Process
required that the teacher be able to demonstrate that the algorithm’s assessment of effectiveness was not applicable
given that the algorithm had not captured the teacher’s unique situation.

28

algorithms) is to improve accuracy over time as more factors are identified and previously
identified factors are re-weighed. Thus, hearings should be reserved for those inmates who argue
that the algorithm did not take into account their unique situations, not that the algorithm failed
to weigh the factors appropriately in their situation.
c) Informal Hearing Requirements
To be sure, to hold hearings for the approximately 150,000 inmates in federal custody
each year would impose a hardship upon the BOP.127 The expense and distraction would be
substantial. But, given that PATTERN to date has only excluded a small fraction of inmates
from ever being considered for early release, the administrative burden should be manageable.128
And, given that inmates would be hard pressed to present unique information as in the examples
sketched earlier129 to persuade the prison decisionmakers that the algorithm’s predictions do not
apply in their situation, required hearings would be few and far between. Indeed, too many
exceptions would introduce the very subjectivity that the algorithm in part was designed to
avoid.
Moreover, the hearings themselves could be quite informal – inmates would have the
right to bring information to the attention of the authorities – no witnesses would be needed. The
Supreme Court in Wilkinson v. Austin130 held that such an informal hearing was sufficient in the
prison context to permit inmates to contest placement in Ohio’s Supermax prison. The Court
determined that Ohio had created a liberty interest in providing that only those inmates posing a

127

Statistics, FEDERAL BUREAU OF PRISONS, https://www.bop.gov/about/statistics/population_statistics.jsp (last
visited July 1, 2021).
128
If we assume that DOJ is correct in stating that 99% of inmates would be eligible, only approximately 1500
inmates would fall into the excluded category.
129
See text accompanying notes 99-101 supra.
130
Wilkinson v. Austin, 545 U.S. 209 (2005).

29

significant risk of danger should be housed in the prison, which greatly restricted individual
freedoms. As the Court explained, prison authorities assigned all Ohio inmates entering the
system a numerical security classification from level 1 through level 5, with 1 the lowest security
risk and 5 the highest.131 Somewhat as with PATTERN, officials based the initial security
classification on numerous static factors (e.g., the nature of the underlying offense, criminal
history, or gang affiliation), with Level 5 inmates, based on those factors or misconduct in
prison, assigned to the Supermax prison. The Court readily concluded that liberty was so
constricted in the Supermax prison that the Sandin hurdle was no obstacle to finding a liberty
interest.132 Nonetheless, in light of the prison context, “informal, nonadversary procedures”
were all that was required to challenge the classification decision.133 The Court explained that
“[r]equiring prison officials to provide a brief summary of the factual basis for the classification
review and allowing the inmate a rebuttal opportunity” satisfied Due Process.134
Supreme Court decisions in other contexts as well have held that Due Process can be
satisfied by an informal hearing. Perhaps most notably, the Court in Goss v. Lopez135 held that,
before school authorities could suspend a student, only an “informal give and take between
student and disciplinarian”136 was required under the Due Process Clause. Due Process
guaranteed the student “the opportunity to characterize his conduct to put it in what he deems the
proper context”137 to reduce the risk of error. In contexts as varied as in Wilkinson and Goss,

131

Id. at 215.
Id. at 223-24.
133
Id. at 228-29.
134
Id. at 226. The Court in Greenholtz earlier stated that, even when states created liberty interests, informal parole
hearings were constitutionally sufficient. According to the Court, the key was for prison authorities to afford an
opportunity to be heard and. If parole were denied, to provide the inmate a statement of the reasons why parole was
denied -- “The Constitution,” the Court held, “does not require more.” 442 U.S. at 16.
135
Goss v. Lopez, 419 U.S. 564 (1975).
136
Id. at 584.
137
Id. Furthermore, Mathews v. Eldridge, 424 U.S. 319 (1976), permits such an informal process when the risk of
error is low and the burden on the government great, particularly in contexts in which greater reliance is placed on
132

30

therefore, the Court has upheld informal mechanisms to allow individuals affected to in essence
tell their side of the story before an adverse action is taken. Accordingly, inmates challenging
exclusion from early release should be entitled to notice, an opportunity to confer with an outside
advisor or attorney, and an opportunity to present information that his or her situation was not
contemplated by the algorithm.138
Due Process, however, would not likely protect the vast majority of inmates who can
pursue the recidivism reducing programs provided in the First Step Act in order to lower their
recidivism category and thereby make early release more likely. Due Process protects the
eligibility decision, but not the pace at which eligible inmates can receive credits. BOP reserves
sufficient discretion in offering programs and in determining which inmates get priority to defeat
any settled expectation under Wolff and Sandin.139 Neither the First Step Act nor current Bureau
of Prisons guidance requires provision of any specific recidivism reduction programming to
inmates based on their PATTERN scores.140 Thus, the classification system – whether the
inmate is placed in high, medium, low, or the minimal category – does not by itself trigger a

the specialized knowledge of government decisionmakers. Cf. Latif v. Holder, 28 F. Supp.3d 1134 (D. Ore. 2014)
(approving of informal mechanism to challenge placement on no-fly list).
138
As noted by the Mathews court, “[p]rocedural due process rules are shaped by the risk of error inherent in the
truthfinding process as applied to the generality of cases, not the rare exceptions.” Id. at 344. Thus, while we find
that an informal hearing is likely to be sufficient, BOP could afford due process by providing an informal hearing
conducted by a variety of means. For example, BOP could establish a standardized form that allows an inmate to
provide documentation challenging the factual inputs used in PATTERN. The warden would be responsible for
reviewing the form to determine whether the inmate provided facts that warrant changes to the inmate’s assessment.
Alternatively, BOP maintains the Administrative Remedy Program, which could be utilized to address PATTERNrelated issues. 28 C.F.R. § 542.10. The Program provides a comprehensive process for inmates “to seek formal
review of an issue relating to any aspect of his/her own confinement,” and provides for multiple layers of appeals.
Id. However, this would likely require some modification to afford for at least a brief appearance before the warden
to comport with the interests at stake.
139
First Step Act of 2018, Pub. L. 115-391, 132 Stat. 5194.
140
At this time, DOJ does not appear to utilize PATTERN in such a manner and the lack of available programming
is one of the common critiques of the First Step Act. See, e.g., Comment Letter to Department of Justice on
PATTERN First Step Act, THE LEADERSHIP CONFERENCE ON CIVIL AND HUMAN RIGHTS,
https://civilrights.org/resource/comment-letter-to-department-of-justice-on-pattern-first-step-act/. Should the
recidivism reduction programs be implemented more formulaically, it is possible that similar Due Process
protections should be extended to all inmates, not just to those currently deemed ineligible by virtue of PATTERN.

31

hearing right at this time, no matter how important. As the Supreme Court stated in Moody v.
Daggett:
We have rejected the notion that every state action carrying adverse consequences for
prison inmates automatically activates a due process right. In Meachum v. Fano, 427 U.S.
215 (1976), for example, no due process protections were required upon the discretionary
transfer of state prisoners to a substantially less agreeable prison, even where that transfer
visited a 'grievous loss' upon the inmate. The same is true of prisoner classification and
eligibility for rehabilitative programs in the federal system.141
Even in situations in which an inmate has been identified as suitable for recidivism reduction
programs that have the potential to result in the award of earned credits, at most the current
system merely affords an opportunity to be placed on a “waitlist” for such programs.
Accordingly, although we believe that inmates should have the right to present
individualized information to determine “eligibility,” prevailing Due Process doctrine cannot be
extended beyond.142 In short, those deemed ineligible by PATTERN should have a limited right
under Due Process to present individualized information, but Due Process does not mandate that
prison authorities consider individualized information bearing on which classification – high,
medium, low, or minimal – inmates should be placed because DOJ has not guaranteed any
inmate specific opportunities to earn credits towards early release. Although the distinction
between “eligibility” for early release and placement in a category from which one can work to
obtain early release may seem artificial, the Supreme Court’s decisions in Sandin and in Moody
v. Daggett signal discomfort with opening up the courts to Due Process suits when the prospect
of early release is uncertain.

141

Moody v. Daggett, 429 U.S. 78, 88 n.9 (1976).
Nonetheless, they might be able to utilize the grievance system that the BOP by regulation has provided for all
inmates See 28 C.F.R. 542(B); see also BOP LEGAL Guide,
https://www.bop.gov/resources/pdfs/legal_guide_march_2019.pdf.
142

32

d) Other Possible Challenges to DOJ’s Algorithmic Determination
Inmates might argue in addition that, as in Loomis, the PATTERN algorithm is unfair or
that it has not been constructed properly. The danger of conditioning liberty on probabilistic data
in sentencing is highlighted by the evidence-based sentencing adopted by over twenty states.143
Those states use predictive analytics to predict the likelihood of recidivism, which affects both
bail determination and sentences. The algorithms used rely on factors such as the offender’s
socioeconomic status and level of education. Predicating liberty on statistics, particularly when
the statistics derive from individual characteristics beyond the offender’s control, departs from
fundamental notions of moral dessert. And the risk of using factors that disproportionately
disadvantage individuals based on race or poverty is high.
For one specific example, ProPublica released a study of risk assessment under the
Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) algorithm
used in a number of state court systems – including Wisconsin as discussed in Loomis -- to
predict the risk of recidivism. It investigated application of COMPAS to 7,000 people in
Broward County, Florida for the purpose of determining whether to release those individuals on
bail.144 According to the subsequent report, the data revealed that race played a substantial factor
in the recidivism projection, which then led to greater jail time for African Americans who
committed similar offenses to whites.145 ProPublica tentatively concluded that the questions
Florida law enforcement authorities asked about socio-economic status and demographic

143

Sonja B. Starr, Evidence-Based Sentencing and the Scientific Rationalization of Discrimination, 66 STAN. L.
REV. 803, 805 (2014).
144
Julia Angwin, Jeff Larson, et. al., Machine Bias, PROPUBLICA (May 23, 2016),
https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing. For a more recent
critique of ProPublica’s study, see Cynthia Rudine, Caroline Wang & Beau Coker, The Age of Secrecy and
Unfairness in Recidivism Prediction, HDSR (Mar. 31, 2021) https://hdsr.mitpress.mit.edu/pub/7z10o269/release/4.
145
Angwin, Machine Bias, supra note 122.

33

conditions, such as whether a parent had been in jail or the number of people known to have used
illegal drugs, played a substantial role in the bond decisions, which led to a significant
disproportionate impact on offenders of color.146
In another study of COMPAS, researchers found that African Americans were nearly
twice as likely as whites to be categorized as high risk but not actually re-offend.147 At the same
time, the algorithm disproportionately categorized white people as being lower risk when followup studies documented that they later re-offended.148 Furthermore, aside from questions of race,
the algorithms may not predict the risk of recidivism well. An independent study from
Dartmouth College found that COMPAS149 is no better at predicting an individual’s risk of
recidivism than random non-expert volunteers.150 These examples demonstrate how algorithms
assessing future dangerousness are driven by the underlying data, which may be linked to race or
other classifications.151

146

Id.
Ed Yong, A Popular Algorithm is No Better at Predicting Crimes than Random People, THE ATLANTIC (Jan. 17,
2018), https://www.theatlantic.com/technology/archive/2018/01/equivant-compas-algorithm/550646/.
148
Id.
149
Complicating the issue, in contrast to PATTERN, the factors underlying COMPAS have not been released to the
public. For a discussion of this issue, see Jason Tashea, Courts Are Using AI to Sentence Criminals. That Must Stop
Now, WIRED (April 17, 2017), https://www.wired.com/2017/04/courts-using-ai-sentence-criminals-must-stop-now.
150
See Julia Dressel & Henry Farid, The Accuracy, Fairness, and Limits of Predicting Recidivism, 17 SCIENCE
ADVANCES (Jan. 17, 2018).
151
The implementation of algorithmic and AI tools in other fields also provide illuminating lessons on how such
tools can have negative, although unintended, consequences. For example, consider Amazon’s use of AI to help
with recruiting and hiring. The team behind the project created a learning AI that was trained by reviewing patterns
present in the company’s recruitment over the preceding ten-year period. One factor the team failed to consider,
however, was the gender gap in the company’s workforce. Through its training, i.e., the algorithm’s review of
previous employment decision data, the AI determined that men were more favorable to hire than women and thus
eliminated qualified applicants from the pool in part based on gender. Jeffrey Dastin, Amazon Scraps Secret AI
Recruiting Tool that Showed Bias Against Women, REUTERS (Oct. 10, 2018), https://www.reuters.com/article/usamazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-womenidUSKCN1MK08G.
147

Similarly, Optum Healthcare developed an AI to help hospitals predict patient care needs in order to drive the
allocation of resources. A study conducted shortly after implementation of the AI determined that it underestimated
healthcare needs of black patients. Conversely, it overestimated the needs of white patients, resulting in the
misallocation of scarce and critical medical resources. After release of the study, Optum released a statement

34

The decision to create categories by algorithm rests in the government’s hands and
should be subject to challenge only to the extent that governmental rules generally can be
challenged, whether under the Administrative Procedures Act (APA) or other mechanism.
Indeed, courts in the past have scrutinized BOP rules to ensure conformance with the APA, with
the Ninth Circuit in Crickon v. Thomas152 invalidating a rule excluding certain categories of
prisoners from early release on the ground that “[t]he BOP offered absolutely no rationale for its
decision to use the inmate’s criminal history as a surrogate for early release ineligibility.”153 In
addition to possible APA challenges,154 any inmate claims of unconstitutionality are cognizable,
as they were in Loomis.155 But, inmates cannot thereafter challenge administration of the

expressly cautioning against the removal of human oversight over AI tools, stating: “[p]redictive algorithms that
power these tools should be continually reviewed and refined, and supplemented by information such as socioeconomic data, to help clinicians make the best-informed care decisions for each patient . . . As we advise our
customers, these tools should never be viewed as a substitute for a doctor’s expertise and knowledge of their
patients’ individual needs.” Carolyn Johnson, Racial Bias in a Medical Algorithm Favors White Patients Over
Sicker Black Patients, WASHINGTON POST (Oct. 24, 2019, 2:00 PM,
https://www.washingtonpost.com/health/2019/10/24/racial-bias-medical-algorithm-favors-white-patients-oversicker-black-patients; see also Charlotte Jee, A Biased Medical Algorithm Favored White People for Health-care
Programs, MIT TECH. REVIEW, https://www.technologyreview.com/2019/10/25/132184/a-biased-medicalalgorithm-favored-white-people-for-healthcare-programs/.
152
Crickon v. Thomas, 579 F.3d 978 (9th Cir. 2009).
153
Id. at 984. See also Elwood v. Jeter, 386 F.3d 842 (8 th Cir. 2004) (invalidating BOP rule limiting early release);
Goldings v. Winn, 383 F.3d 17 (1st Cir. 2004) (same); Arrington v Daniels, 516 F.3d 1106, 1113 (9th Cir. 2008)
(invalidating similar exclusion from early release under the APA); cf. Handley v. Chapman, 587 F.3d 273 (5 th Cir.
2009) (reviewing BOP rule under APA but disagreeing with Arrington Court on the merits).
154
Others have argued that, at least in the sentencing context, those convicted should have a right to understand the
methodologies that are guiding imposition of their final sentences, even if no discrimination is involved. Michael
Brenner, Jeannie Suk Gerson, Michael Haler, Matthew Lin, Amil Merchant, Richard Millett, Suproteem Sarkar &
Drew Wagner, Constitutional Dimensions of Predictive Algorithms in Criminal Justice, 55 HARV. CIV.-RTS-CIV.
LIB. LAW REV. 267, 283-87 (2020). Other than in a broad challenge under the APA, however, we do not believe
that prison authorities need explain the algorithm’s methodology to each prisoner.
155
From its inception, PATTERN’s factors and algorithm have been based on statistical models, which leverage
historical federal prison data, and DOJ has been mindful to avoid any disproportionate impact based on race. In
response to stakeholder concerns that certain factors in the algorithm serve as proxies for race, the Department
removed two static factors; first, the offender’s age at first arrest; and second, whether the inmate voluntarily
surrendered. Although each factor appears neutral on its face, critics argued that reliance on those factors would
plainly result in racially disproportionate stays in prison. The Department noted that the reduction in algorithmic
accuracy was outweighed by risk of actual or perceived bias that these factors posed. PATTERN should survive any
Equal Protection Clause challenge based on race.
PATTERN’s algorithm applies a separate and distinct model for men and for women and as a result has been subject
to allegations of unconstitutionality. Citing statistical evidence and gender specific pathways to crime, the
Department determined that applying the same algorithm to men and women would yield unfairly elevated results

35

algorithm. At the end of the day, the key for inmates should be their ability to introduce
information that their unique situations fall outside of PATTERN, that such information
indicates a low likelihood of future dangerousness, and thus that they should be eligible for early
release.
2. Right to Consider Inputs used to Determine Statutory and PATTERN Eligibility
Offenders will only rarely be able to persuade prison authorities of the uniqueness of their
situation. In most contexts, the results of the algorithm as to future dangerousness will govern.
In light of the liberty interest created, however, offenders who are excluded by statute or by
operation of PATTERN from earning credits should enjoy a limited right to challenge the data
used in reaching that statutory or regulatory exclusion. First, with respect to the statutory
exclusion, an inmate in an extraordinary case may have reason to believe that he or she did not
commit the disqualifying offenses listed in the Act. For instance, perhaps there was a dispute as
to whether his or her offense qualifies as a sexual offense that precludes “earned” credits, or
whether a conviction for false imprisonment should be tantamount to one for “kidnapping.”156
Prison authorities may base assessment of historical factors on data in the pre-sentence report,
but errors may persist and, in some contexts, the report may have been withheld from the
offender.157 Arguably, in such rare contexts, liberty interest analysis should allow inmates a

for women, inaccurately reflecting their recidivism risk. To be constitutional, a government policy which expressly
discriminates based on gender must be shown—by an exceedingly persuasive justification—to serve important
governmental objectives and to employ a means which is substantially related to those objectives. As it relates to
PATTERN, the Department of Justice is on strong ground that explicit gender differences in PATTERN serve the
important government objective of ensuring that female offenders not be incarcerated longer than their predicted risk
of recidivism.
156
18 U.S.C. § 3632(d)(4)(D) (providing list of convictions making a prisoner “ineligible to receive [earned]
credits”).
157
See Danielle Kehl, Priscilla Guo & Samuel Kessler, Algorithms in the Criminal Justice System: Assessing the
Use of Risk Assessments in Sentencing, HARVARD L. SCHOOL, 15 (2017) (also discussing how the use of algorithms
in the sentencing context can implicate constitutional concerns such as the “right to review and challenge the
evidence used to determine guilt and punishment” independent of a statutorily created liberty interest).

36

right to request disclosure of the inputs that led to the statutory disqualification, and an informal
opportunity to present information that the input was incorrect. Eligibility for earned credits
creates a liberty interest, so an informal hearing should be required if there is a dispute about the
disqualifying offenses under the First Step Act.158
Similarly, for those deemed ineligible by virtue of PATTERN, disclosure of the inputs
should be constitutionally required. Given that PATTERN’s assessment of risk initially is based
on the listed static factors, prison authorities should disclose those factors upon request, and the
inmate should be afforded a limited opportunity to challenge an error – whether the nature of a
prior offense or education level.159 Required disclosure represents a key check to ensure against
errors in algorithmic decision-making.
Congress previously directed prison authorities to disclose files pertaining to an
upcoming parole decision to allow the inmate a chance to respond before the decision is
finalized: “At least thirty days prior to any parole determination proceedings, the prisoner shall
be provided . . . reasonable access to a report or other document to be used by the Commission in
making its determination.” 160 The Supreme Court as well has noted the importance, if not
requirement, that parole authorities share information with the inmate that is in his or her file.161
Thus, even though the PATTERN tool limits the issues to be resolved at a hearing, Due Process
should extend to disclosure of the inputs upon which the decision is based.

158

See also Cary Coglianese & David Lehr, Transparency and Algorithmic Governance, 71 Admin. L. Rev. 1, 41
(2019) (similarly arguing that Due Process protects the right of individuals to gain access to the inputs).
159
The inmate need not, however, be entitled to receive any confidential information upon which a static factor is
based. See, e.g., Harrison v. Shaffer, 2019 U.S. Dist. LEXIS 236681 (N.D. Cal. 2019) (canvassing cases upholding
government’s right not to disclose confidential information.).
160
18 U.S.C. 4208(b).
161
Swartout v. Cooke, 562 U.S. 216, 220 (2011). See also Tasker v. Moh, 165 W. Va. 55, 65-66 (1980)
(“Permitting the prisoner to access his own file is one means of accomplishing the [Due Process required] notice.”).

37

Consider the analogous decision by the Kansas appellate court in State v. Walls.162 Walls
pleaded no contest to a low-level felony, but his assessment score indicated that he was a high
risk-high needs candidate for probation. Based largely on the assessment, the trial court placed
him in a restrictive probation setting. On appeal, the appellate court reversed, holding in part
that the trial court committed reversible error in failing to disclose the data upon which the
assessment was based. Given that the court could not have imposed the restrictive conditions in
the absence of the assessment, it was incumbent upon the court to disclose the data to permit the
defendant to challenge the data used in the assessment, even though he could not challenge how
the assessment was structured.163 Accordingly, prison authorities should have the duty, upon
request, to furnish inmates the relevant data upon which the statutory or PATTERN exclusion is
based.
3. Right to “Earned” Credits
The earned credit system in PATTERN also has created a limited liberty interest under
the Morrissey and Wolff analysis. The First Step Act provides that credits can be earned by
inmates for successful completion of qualified educational and treatment programs. The
congressional language is clear: Section 101 provides that “a prisoner shall earn 10 days of time
credits for every 30 days of successful participation” and “shall earn an additional 5 days of time
credits for every successful participation in evidence-based recidivism reduction
programming.”164 Moreover, Congress directed the Bureau of Prisons to develop guidelines for
the reduction of awards in case of “violations,”165 and provided that any reduction “shall require

162

2017 Kan. App. Unpub. LEXIS 487 (June 23, 2017).
Id. at *13.
164
Id. at § 101.
165
18 U.S.C. § 3632(6).
163

38

written notice to the prisoner” and “a procedure to restore time credits that a prisoner lost as a
result of a rule violation.”166
Indeed, Congress directed that inmates determined to be “at minimum or low risk of
recidivating,” pursuant to their last two assessments, shall be released”167 once they have earned
sufficient credits into a halfway house or similar program unless the “warden of the prison finds
by clear and convincing evidence [based on conduct in prison] that the prisoner should not be
released.”168 Through PATTERN, the Department of Justice plainly wanted to alter inmate
behavior to reduce the risk of recidivism. By offering early release and other benefits to
participants in programming aimed to reduce recidivism after incarceration, the government
plainly hoped to elicit participation.169
Contrast the First Step Act to the facially similar program operated by the Bureau of
Prisons a generation earlier to encourage nonviolent inmates to pursue drug treatment programs
while incarcerated. The key statutory provision stated “[t]he period a prisoner convicted of a
nonviolent offense remains in custody after successfully completing a traditional program may
be reduced by the Bureau of Prisons.170 The Supreme Court in Lopez v. Davis171 stressed that, in
light of such permissive language, “when an eligible prisoner successfully completes drug
treatment, the Bureau thus has the authority, but not the duty, to reduce the term of

166

Id.
18 U.S.C. § 3632(d)(4)(C).
168
Id.
169
Indeed, psychological theory also supports the premise that those observed likely will alter their behavior in ways
that they know the observer favors. One such theory is that of the Hawthorne effect. For a brief discussion of the
Hawthorne effect, see, e.g., Abraham Zaleznik, The Hawthorne Effect, HARVARD BUSINESS SCHOOL,
https://www.library.hbs.edu/hc/hawthorne/09.html#nine (last visited Jan. 19, 2021).
170
18 U.S.C. § 3621(e)(2)(B) (emphasis added).
171
Lopez v. Davis, 531 U.S. 230 (2001).
167

39

punishment.”172 In contrast, the directive in the First Step Act is clear – the prisoner “shall earn”
the credits, and if that prisoner has previously been classified as “low” or “minimal” risk, the
prisoner must be released once sufficient credits have been earned.173
Thus, in light of Morrissey, inmates must be afforded a hearing if the government refuses
to award the credits after completion of the program or attempts thereafter to revoke them.174
Even if Congress decides to scrap the dynamic factors, inmates who relied on those factors to
obtain the education and treatment previously thought necessary to reduce the risk of recidivism
would be entitled to the early release they otherwise would have earned. Although courts have
yet to recognize a liberty interest in the earned credit system,175 the First Step Act invites inmates
to rely on the dynamic traits to reap the benefits of their action. The Due Process Clause
demands as much.
The government might argue that no liberty interest can be created until the inmate
receives the credits. To be sure, the Supreme Court has not squarely decided whether the Due
Process Clause applies to applicants for government entitlements such as benefit programs, as
opposed to individuals already receiving benefits who are at risk of losing them. In other words,
the Court has yet to resolve whether applicants for Social Security Disability or public housing
have a right to a hearing under the Due Process Clause if their applications are denied. Lower
courts, however, have reasoned that Due Process applies in such contexts.176 Whatever the result
in other administrative contexts, Due Process analysis seems particularly appropriate to apply in
Id. at 241. See also Morales v. Francis, 2005 U.S. Dist. LEXIS 26253 (S.D. Tex) (rejecting challenge to BOP’s
limitation if inmates’ ability to serve out sentences in half-way house in light of similar permissive language).
173
First Step Act of 2018, Pub. L. 115-391 § 101.
174
See generally Mathews v. Eldridge, 424 U.S., 319 (1976) (discussing the requirement of an impartial hearing
when due process protections are found based on a statutorily created interest).
175
See, e.g., Allen v. Hendrix, 2019 U.S. Dist. LEXIS 22791 (E.D. Ark.); Wren v. Watson, 2020 U.S. Dist. LEXIS
21282 (S.D. Ind.).
176
See, e.g., Cushman v. Shinesky, 576 F.3d 129 (Fed. Cir. 2009); Kapps v. Wing, 404 F.3d 105 (2d Cir. 2005).
172

40

the prison context when the opportunity to earn credits to gain early release is at stake. Under
the First Step Act, those offenders hoping to earn credits pursue the vocational and treatment
programs in reliance on the government’s pledge, and that reliance should remove any doubt but
that Due Process protects the inmate’s interest when he or she completes the educational or
treatment programs.
The government conceivably could argue as well that no liberty interest is created
because the First Act conditions earned release credits on “successful”177 completion of certain
education and treatment programs raises a closer question. On the one hand, to the extent that
“successful” refers to technical completion of an educational or treatment program, then a liberty
interest is created. Inmates know that, through completion, credits are earned. The statutory
language strongly suggests that “successful” incorporates an objective standard as to whether the
inmate attended a treatment program, held down a prison job, or pursued vocational training, not
how well they performed such tasks. Should prison authorities deem that the inmate’s efforts are
not successful, some brief hearing under Mathews v. Eldridge178 would need be convened to
determine if the credits should be awarded.179 On the other hand, if prison authorities retain the
ability subjectively to assess whether the offender’s participation in the program was
“successful,” then the liberty interest only arises after the official deems the offender’s
participation “successful.” In light of Congress’s intent to persuade offenders to pursue the
recidivism reduction measures, “successful” seems more of a technical requirement. In context,

177

18 U.S.C. § 3632(d)(4).
Mathews v. Eldridge, 424 U.S. 319 (1976).
179
As addressed in note supra, BOP could create a process that affords for a brief appearance before the warden,
who would then be responsible for determining whether credits should be awarded based on factual evidence. We
set forth that Due Process requires some form of an informal hearing, but BOP could meet the requirements of Due
Process in this regard through a variety of means.
178

41

therefore, “successful” construed within the earned credit system suggests that an objective
standard is contemplated, triggering a right to a hearing if the credits are denied.
In short, the First Step Act creates a liberty interest in creating an “earned credit system.”
Once earned, such credits cannot be withheld or rescinded absent a hearing. Although the earned
credits under the First Step Act do not reduce the length of the sentence per se, they entitle the
inmate to release from the prison setting to home confinement or other type of supervised
release. Such release roughly equates to release on parole, which the Greenholtz and Allen
Courts held subject to liberty interest analysis.180
In contrast, the more violent offenders who can only earn greater telephone privileges and
the like might not surmount the Sandin hurdle. Those privileges, while incredibly important to
the inmate, are not currently protected under the Due Process Clause in line with Sandin, as the
denial of such benefits would not constitute the requisite atypical hardship.181
B. Ex Post Facto Issues
To the extent that the enhanced ability to accumulate good time (not earned) credits under
the First Step Act effectively modifies the sentence imposed on the offender, the Ex Post Facto
Clause182 prevents Congress from changing the system retroactively. The Clause prevents the
federal (and state) government from increasing punishment after an offender commits the
covered offense. Just as the Due Process Clause limits the government’s ability to deny release
to particular inmates without, at times, affording a hearing, so the Ex Post Facto Clause prevents
the government from retroactively changing release policies built into the sentence retroactively.

180

See also note 123 supra.
Sandin v. Conner, 515 U.S. 472 (1995).
182
U.S. CONST. art. I, § 9-10.
181

42

The Supreme Court addressed an analogous situation in Weaver v. Graham,183 holding
that a state’s retroactive change in its good time policy violates the Ex Post Facto Clause as long
as the change “constricts the inmate’s opportunity to earn early release.”184 Similarly, in
Greenfield v. Scafati,185 the Court prohibited Massachusetts from retrospectively altering its good
time credit system so that any inmate violating parole would be unable to accumulate good-time
credits upon return to prison.186 When Congress in the First Step Act increased the number of
good time credits offenders could earn, it locked in that benefit until Congress alters the
program. Straightforward application of Weaver precludes the government from making it more
difficult for offenders to earn good time credits after they have committed the underlying
offense.187
Whether the government could eliminate the “earned” credit system for offenders who
have already offended presents a closer question. On the one hand, the earned credit system,
unlike the good-time credits, may not be considered part of the sentence given that the credits
turn on dynamic factors. Inmates earn credits toward release based on what they do after
reaching prison. Lower courts have held that the Ex Post Facto Clause does not protect inmates
from changes in discretionary application of factors bearing on early release. For instance, the
Eighth Circuit in Ellis v. Norris188 permitted repeal of a statute allowing for discretionary awards
of good-time credits. The fact that the change left the inmates worse off did not persuade the

183

Weaver v. Graham, 450 U.S. 24, 25 (1981).
Id. at 35-36.
185
Greenfield v. Scafati, 277 F. Supp. 644 (D. Mass. 1967), aff’d per curiam, 390 U.S. 713 (1968).
186
Id. at 644-45.
187
The government need not apply enhanced good time credits retroactively, but it has so specified in the Act.
188
Ellis v. Norris, 232 F.3d 619 (8th Cir. 2000).
184

43

court. Similarly, the Fourth Circuit in Burnette v. Fahey,189 held that Virginia’s change from a
discretionary parole system based on risk assessment to one based principally on the seriousness
of the original offense did not violate the Ex Post Facto Clause, even though some inmates
undoubtedly would be treated less favorably.190
On the other hand, offenders could plausibly argue that they factored in the prospect of
such “earned credits” when entering their pleas. Even when release programs are
“discretionary,” they may influence offenders, who rely on their potential.191
Although the issue is not free from doubt, the First Step Act apparently does not
guarantee that education and treatment programs will be offered. In fact, criminal justice leaders
and leading reformers have criticized DOJ for not making such programs available, thereby
blunting the underlying purpose of the First Step Act to encourage inmates to pursue
programming that will diminish the likelihood of their recidivism.192 At this time, the
programming which drives the dynamic factors of PATTERN remains discretionary. Thus, the
opportunity to “earn” credits through work and education programs cannot be considered part of
the original sentence and is not therefore protected by the Ex Post Facto Clause. In short, given
that the sentence length formally will not change and that the prospect of earning credits remains

189

Burnette v. Fahey, 687 F.3d 171 (4th Cir. 2012). See also Morales v. Francis, 2005 U.S. Dist. LEXIS 26253
(S.D. Tex.) (retroactively limiting right to serve out sentence in halfway house does not violate the Ex Post Facto
Clause).
190
Taking away discretionary credits, however, clearly violates the Ex Post Facto Clause. See Lynce v. Mathis, 519
U.S. 433 (1997) (retroactive cancellation of discretionary credits violates the Ex Post Facto Clause).
191
See Paul D. Reingold & Kimberly Thomas, Wrong Turn on the Equal Protection Clause, 106 CAL. L. REV. 593
(2018) (arguing that the presence of discretion should not be dispositive for Ex Post Facto analysis, but rather the
potential for earlier release).
192
Ames Grawert, What Is the First Step Act — And What’s Happening With It?, BRENNAN CENTER FOR J. (June 23,
2020), https://www.brennancenter.org/our-work/research-reports/what-first-step-act-and-whats-happening-it
(discussing concerns raised by various parties that DOJ and BOP provide insufficient programming opportunities
and that those that are available are often under-resourced and understaffed).

44

speculative, Congress can alter the earned credit program without running afoul of the Ex Post
Facto Clause.
In any event, Ex Post Facto principles would not likely prevent the Bureau of Prisons
from altering its PATTERN methodology. In Miller v. Florida,193 the Court invalidated
retroactive application of a change in sentencing regulations that presumptively increased the
offender’s sentence.194 In contrast to Miller v. Florida, however, an update of PATTERN might
result in finding that recidivism is more likely for some inmates, but less for others. The very
purpose of AI is to improve efficacy with greater information. Changes to the methodology in
PATTERN after an offender commits a relevant offense would not violate the ex post facto
prohibition, even if the prediction of future dangerousness increases for some inmates. Indeed,
DOJ has already altered PATTERN scores for some based on a change in methodology.195 The
Supreme Court’s decision in Garner v. Jones196 suggests that, only if such change in
methodology has created “a significant risk of prolonging . . . incarceration” within the “whole
context of the parole system,”197 would the Ex Post Facto Clause be violated. In Garner, the
Court upheld a Georgia law extending the interval between periods of parole consideration.
Although Congress cannot abolish the First Step Act’s enhanced good time credits policy
retroactively, it can alter how the government calculates the likelihood of recidivism, which it
employs in the earned credit program. Thus, Ex Post Facto principles may prevent the
government from retroactively removing the enhanced good time credits and, while Due Process

193

Miller v. Florida, 482 U.S. 423 (1987).
Id. at 429-36.
195
First Step Act Fiscal Year 2020 Interim 90-Day Report, DEP’T OF JUSTICE, 3-4 (June 2, 2020),
https://www.ojp.gov/pdffiles1/nij/254799.pdf.
196
Garner v. Jones, 529 U.S. 244 (2000).
197
Id. at 251.
194

45

principles in part constrain the government from rescinding the dynamic aspects of the First Step
Act, the Ex Post Facto prohibition likely does not come into play.
* * * *
If courts hold, as we have argued, that the First Step Act triggers Due Process protections,
Congress or the DOJ might react by inserting more discretion into the system. Instead of
predicating early release on a low likelihood of future recidivism, Congress could determine that
early release be permitted only when BOP officials deem it warranted, much as in the parole
system that existed for years. No liberty interest would arise because the government would not
have tied its own hands.198 Nothing precludes Congress from altering the statutory scheme and
changing PATTERN so as to avoid giving rise to the Due Process protections we have discussed.
Indeed, many states reacted to the Supreme Court’s recognition of a liberty interest in
parole by creating more discretionary parole decisions. The response of Georgia is illustrative.
Under the regulatory system, the Board of Pardons and Parole had promulgated a Georgia Parole
Decision Guidelines System, setting forth a step-by-step system to evaluate whether an inmate
was entitled to parole.199 The Board assigned an inmate a crime severity level, and a parole
success likelihood score. Based on a combination of the two, the Board then established a target
release date.200 The parole statute directed that the “guidelines system shall be used in
determining parole actions on all inmates.”201 Despite the seemingly objective standards, the

198

Texas, for instance, created a roughly comparable risk assessment tool based on static and dynamic factors to
govern parole, but explicitly vested discretion in parole authorities to disregard the guidelines when they deemed it
appropriate – “Parole panel members retain the discretion to vote outside the guidelines when circumstances of an
individual case merit their doing so.” Revised Parole Guidelines, TEXAS BOARD OF PARDONS AND PAROLES,
https://www.tdcj.texas.gov/bpp/parole_guidelines/parole_guidelines.html (last visited July 1, 2021).
199
Sultenfuss v. Snow, 35 F.3d 1494, 1495 (11th Cir. 1994) (en banc).
200
Id. at 1497.
201
Ga. Code Ann. 42-9-40.

46

Board added a note to its guidelines: “The Board specifically reserves the right to exercise its
discretion under Georgia law to deny parole even though Guidelines Criteria are met by an
inmate. It is not the intention of the Board to create a “liberty interest” of the type described in
Greenholtz v. Nebraska Penal Inmates (citation omitted).”202
Similarly, under a complicated point system, the District of Columbia had provided that
certain offenders “shall” be granted parole.203 In response to Greenholtz, the City Council
changed “shall” to “may.”204 Here, as well, BOP could make it clear that, despite the PATTERN
algorithm, it retained the ultimate discretion whether to permit early release.
Such response, however, is inconsistent with the rehabilitative ideal underlying the First
Step Act. Congress made clear that it wanted inmates to enroll in programs that would reduce
the risk of recidivism and facilitate offenders’ reentry into society. Congress presented inmates
with a choice – take steps during the period of incarceration to reduce the likelihood of
recidivism or serve out the sentence in prison as originally meted. Changing to a discretionary
system, as in the Georgia and DC contexts, would blunt the force of the congressional goal and
reintroduce the ills of excessive discretion, which had led Congress to scrap the parole system in
the first instance. Inmates should be able to trust that successful completion of recidivism
reduction programming is the quid pro quo for early release.205 Otherwise, future inmates would

202

Sultenfuss, 35 F.3d at 1503.
See Ellis v. District of Columbia, 84 F.3d 1413, 1415-16 (D.C. Cir. 1996).
204
Technical Amendments Act of 1994, D.C. Acyt, 10-302, Section 52(c)-(e), codified at D.C. Mun. Regis. Tit. 28,
section 204.19-.21.
205
Congress was aware of the connection between treating inmates with respect and rehabilitation. See, e.g.,
Transforming Prisons, Restoring Lives: Final Recommendations (Charles Colson Task Force on Federal
Corrections), NATIONAL INSTITUTE OF CORRECTIONS, 36-37 (2016); https://nicic.gov/transforming-prisonsrestoring-lives-final-recommendations-charles-colson-task-force-federal#:~:text=on%20Federal%20Corrections,Transforming%20Prisons%2C%20Restoring%20Lives%3A%20Final%20Recommendations%20of%20the%20Cha
rles%20Colson,Task%20Force%20on%20Federal%20Corrections&text=%22After%20decades%20of%20unbridled
%20growth,States%20faces%20a%20defining%20moment. The Task Force’s findings and ultimate
203

47

have no reason to rely on promises of prison authorities and less reason to pursue the recidivism
reduction programming, with a greater likelihood that they will recidivate in the future.
Indeed, many of the theories underlying rehabilitation focus on the importance of treating
inmates with dignity.206 Respectful interaction increases inmates’ motivation to work towards
rehabilitation. Part of the First Step Act furthers that goal by encouraging offenders to take
agency of their own futures. Thus, Congress should welcome the limited constitutionalization of
early release because that very determination may well further the rehabilitative goals in the Act.
CONCLUSION
In sum, the First Step Act reaches into history to bring back rehabilitation as a goal of our
federal criminal justice system. The centerpiece of that effort – inclusion of dynamic factors to
encourage inmates to pursue steps while incarcerated that should minimize the chances of
recidivism – has yet to be fully implemented, but holds significant promise for the future.
At the same time, Congress’ efforts will straightjacket its own flexibility in the future to
alter prison policies for those already in the system. Prison authorities must disclose to offenders
who are excluded from eligibility under both the statute and PATTERN the inputs upon which
the exclusion was based, and inmates excluded from eligibility under PATTERN must be
afforded a limited opportunity to present information that they should be entitled to earn credits

recommendation for a renewed emphasis on the components of rehabilitation theory formed the starting point for the
First Step Act, as evidenced by the relevant House Committee Report. H.R. REP. 115-699, at 22-24 (2018).
206
See, e.g., Edgardo Rotman, Do Criminal Offenders Have a Constitutional Right to Rehabilitation?, 77 J. CRIM. L
& CRIMINOLOGY 1023, 1034 (1986); Jeremy Coylewright, New Strategies for Prisoner Rehabilitation in the
American Criminal Justice System: Prisoner Facilitated Mediation, 7 J. HEALTH CARE L. & Pol’y 395, 406 (2004);
Anita Hurlburt, Building Constructive Prison Reform on Norway’s Five Plllars Cemented Through Aloha, 19 AsianPacific L. & Policy J. 194, 232-33 (2018) (focusing on Norway’s prison policies in advocating for reform of
Hawaii’s criminal justice policies); Major Evan R. Seamone, Reclaiming the Rehabilitative Ethic in Military
Justices: The Suspended Punitive Discharge as a Method to Treat Military Offenders with PTSD and to Reduce
Recidivism, 208 MIL. L. REV. 1, 62 (2011); Amanda Ploth, Note, Why Dignity Matters: Dignity and the Right (or
not) to Rehabilitation from International and National Perspectives, 44 N.Y.U. J. INT’L L. & POL. 887 (2012).

48

towards early release despite what the algorithm indicates. Moreover, the earned credit system
itself creates a liberty interest, which prison authorities must respect once the recidivism
reduction programs have been completed. And, while Congress can jettison both the static and
dynamic factors prospectively, it must pursuant to the Ex Post Facto prohibition honor the
increase in good time credits for those who committed offences prior to the operative date.
Perhaps unwittingly, Congress’s measures in the First Step Act have constitutionalized in part
the former discretionary decision to release inmates before the end of their incarceration. Such
constitutionalization will result in greater administrative costs. That price seems more than
reasonable to create a quasi-contract between the government and offender to help the offender
take the steps needed to reenter society and avoid the trap of recidivism.

49