Industries / Government, education and community / Government / Corrections
Release and reentry
Reentry is the move from prison back into the community: the papers, housing, health coverage, work and services that people need in the first weeks and months after release.
6
Milestones achieved, from 1845 to 2019.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1845
A home after prison
In June 1845 the Female Department of the Prison Association of New York opened the Home for Discharged Female Convicts in a rented house in Manhattan, giving up to 30 women at a time somewhere to live after release and help finding work. New York's Landmarks Preservation Commission says historians consider it the first halfway house, a supervised home for people leaving prison, in the world. Renamed the Isaac T. Hopper Home in 1858, it moved to 110 Second Avenue in 1874.
Credit
Isaac Tatem Hopper and his daughter Abigail Hopper Gibbons, Quaker reformers
2008
Federal money for reentry
The US Second Chance Act, signed on 9 April 2008, funded grants to state and local governments and nonprofits for reentry planning and services. An evaluation for the National Institute of Justice randomly assigned 966 people to seven programs funded under the act or to a control group. After 30 months those in the programs had received more services and had higher employment and earnings, but they were no less likely to be rearrested, reconvicted or reincarcerated, and they had somewhat more arrests and convictions, perhaps because closer supervision detected more.
LessonMore services did not by themselves mean fewer returns to prison, and closer watching can raise the count of violations it is meant to lower.
Credit
US Congress; evaluated by Ronald D'Amico and Hui Kim (March 2018)
2015
Rates that do not compare
Recidivism, a return to crime after release, is counted in different ways: as rearrest, reconviction or reimprisonment, over periods from six months to five years. Seena Fazel and Achim Wolf looked for official rates in the 20 countries with the most people in prison and found that only 2 of the 20 reported them; across the 18 countries with data, they found that "few countries were comparable". Follow-up time alone changes the answer: of 401,288 people released from state prisons in 30 US states in 2005, 68% were arrested within 3 years and 83% within 9 years, according to the Bureau of Justice Statistics.
LessonOne country's recidivism rate cannot be set beside another's unless the definition, the group counted and the follow-up time match.
Credit
Seena Fazel and Achim Wolf (PLOS ONE); the US Bureau of Justice Statistics
2016
Japan legislates against reoffending
Japan's Act for the Prevention of Recidivism (Act No. 104 of 2016) requires the national government to adopt a plan to prevent reoffending, covering education, vocational training, jobs, housing and access to health and welfare services, and asks prefectures and towns to make their own. It tells the state to favor "cooperating employers", firms that hire people with criminal records, in government contracts, and to give people leaving prison special consideration for public housing. In 2012 the government had set a target of cutting the share of people back in prison within two years of release to 16% or less by 2021; the Ministry of Justice's 2024 summary gives the latest rate as 13.0%.
Credit
Japan's National Diet and Ministry of Justice
2016
Hiding records backfires
"Ban the box" laws stop employers from asking about criminal records on job applications. Using the timing of state and local laws, Jennifer Doleac and Benjamin Hansen found that the laws lowered the chance of being employed by 3.4 percentage points for young, low-skilled Black men and by 2.3 points for young, low-skilled Hispanic men. In a field experiment with about 15,000 made-up online applications in New Jersey and New York City, Amanda Agan and Sonja Starr found that white applicants received 7% more callbacks than similar Black applicants from employers that asked about records, and that the gap grew to 43% at companies that removed the question.
LessonWhen employers cannot see a record, some guess it from race, so the rule can hurt the people it was meant to help.
Credit
Jennifer Doleac and Benjamin Hansen (NBER working paper, 2016); Amanda Agan and Sonja Starr (Quarterly Journal of Economics, 2017)
2019
Records sealed automatically
Pennsylvania's Clean Slate law, signed by Governor Tom Wolf in June 2018, made it the first US state to seal some criminal records by computer, without the person having to apply. From late June 2019, charges that did not end in conviction, summary offenses and most nonviolent misdemeanor convictions could be sealed, convictions after ten years without a new offense and once fines were paid; the public can no longer see them, though police still can. WHYY reported that 40 million charges were eligible. The need showed in Michigan, where J.J. Prescott and Sonja Starr found that only 6.5% of people eligible to clear a conviction did so within five years, and that those who did saw their wages rise by over 22% within a year.
LessonA right that people must apply for reaches few of them; doing it automatically reaches the rest.
Credit
Community Legal Services of Philadelphia, which helped write the law, and the Pennsylvania General Assembly; J.J. Prescott and Sonja Starr of the University of Michigan Law School
Watch for
Medicaid before release
In April 2023 the US Centers for Medicare & Medicaid Services (CMS) invited states to use Medicaid, the public health insurance for people with low incomes, to pay for some services in the 90 days before a person's expected release, a partial exception to the rule that Medicaid does not pay for care during incarceration. In KFF's summary, the aims are better care transitions, steadier coverage, better health and less recidivism. KFF counts 19 state approvals under the Biden administration and five more in September 2026; Michigan withdrew its approved plan in December 2025.
Achieved whenan independent evaluation compares people released with and without pre-release coverage and publishes the effects on their care after release and on returns to prison.
Claimed by
CMS, April 2023, renewed with new approvals in September 2026
A chatbot by prison phone
Beyond Bars AI offers Nova, a voice chatbot that people in prison reach through the Securus phone system, with 10 free minutes a month; family members can sponsor 60 minutes a month for $25. Its founder, CJ Jones, says the company aims to make AI available to people in prison nationwide, and it has partnered with a nonprofit to promote Nova in Missouri's prisons. A writer held in Missouri who tried it found its advice generic but helpful, and his roommate used it to size up the job market where he will live after release.
Achieved whenan outside test of its answers to reentry questions, such as how to get an ID or apply for benefits, publishes how often they are correct.
Claimed by
Beyond Bars AI, as reported by the Prison Journalism Project, June 2026
Open
01
Test reentry chatbots
A public test of AI chatbots on 300 scripted questions that people ask when leaving prison, such as how to replace a birth certificate, apply for food benefits, seal a record or find housing, written for five US states and two other countries. Legal-aid staff grade each answer against official sources and mark errors that would send someone the wrong way.
The briefCredit · unclaimed02
Count who leaves with ID
A state-by-state table of how many people released from state prisons each year left with a photo ID, a birth certificate and a Social Security card, drawn from published reports and from public records requests for aggregate counts only.
The briefCredit · unclaimed
Open 01 · the brief
Test reentry chatbots
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI chatbots on 300 scripted questions that people ask when leaving prison, such as how to replace a birth certificate, apply for food benefits, seal a record or find housing, written for five US states and two other countries. Legal-aid staff grade each answer against official sources and mark errors that would send someone the wrong way.
Why it is still open
ViaPath reports how often its ViaChat career guide is used, not how often it is right, and a writer who tried the Nova voice chatbot in a Missouri prison in June 2026 found generic advice and ran out of minutes mid-answer. LindaBot, a chatbot that University of South Australia researchers designed with women who had been in prison, to help with tasks such as replacing ID, was still a prototype in April 2024 with no test results reported. Searches of arXiv and Crossref found no published accuracy test of reentry chatbots.
Built with what exists
General models through their public APIs, such as Claude, Gemini and GPT; open-weight models run locally with llama.cpp; public reentry chatbots where an outside tester can reach them; state agency web pages as the answer key; a speech-to-text model to transcribe voice tools, such as ElevenLabs Scribe v3 (see the Note "Fewer word errors").
First step
Write 30 questions on getting a state ID after release in one state, using its motor vehicle and corrections pages, and run them through three chatbots.
Done looks like
300 questions, at least five tools, every answer graded blind by two graders, and the share correct and the share with a harmful error published for each tool, with the questions and the grading rules.
Open 02 · the brief
Count who leaves with ID
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A state-by-state table of how many people released from state prisons each year left with a photo ID, a birth certificate and a Social Security card, drawn from published reports and from public records requests for aggregate counts only.
Why it is still open
The Reason Foundation's February 2026 survey of all 50 states describes their laws and programs for ID at release, but for several states it reports no outcome data; for Alabama it says "there is no available data indicating the number of people who have been provided documents upon release". A few states publish counts. In Hawaii, whose law requires help with ID, 52% of people released between November 2019 and September 2020 left without an ID card, and about half of the 692 people who left its prisons from November 2024 to October 2025 still left without one, according to a report to the state legislature described by Honolulu Civil Beat; Michigan's corrections director said in December 2025 that 99% of people leave its prisons with a state ID card or driver's license. No table brings such counts together.
Built with what exists
State corrections agencies' annual and legislative reports; public records requests sent and tracked through a service such as MuckRock; a spreadsheet; a model that reads long reports to find ID figures, with every figure checked by a person.
First step
Send aggregate-count requests to five states whose laws require help with ID before release.
Done looks like
A count, a written refusal or a "no records" reply from all 50 states, published by state and year with each source document, and the request letters shared for others to reuse.