Industries / Government, education and community / Government / GovTech
Records and data
The computers, databases and core systems that hold government records, often for decades, and the ways governments back them up, move them to the cloud and publish them as open data.
7
Milestones achieved, from 1951 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1951
A computer for the census
The US Census Bureau signed the contract for the first UNIVAC I on 31 March 1951 and used it to tabulate part of the 1950 population census and all of the 1954 economic census. J. Presper Eckert and John Mauchly designed it, and Remington Rand "eventually sold 46 Univac 1s at more than $1 million each", according to the Computer History Museum.
Credit
J. Presper Eckert and John Mauchly; Remington Rand; US Census Bureau
2009
Government data in one catalog
The US opened Data.gov on 21 May 2009 with 47 datasets, and two years later it listed 389,681, by the count of Vivek Kundra, then the federal chief information officer. The UK opened data.gov.uk on 21 January 2010 with more than 2,500 datasets, with Tim Berners-Lee and Nigel Shadbolt in key roles. Its catalog ran on CKAN, open-source software that the Open Knowledge Foundation first released in 2006, and GSA's code repository shows the US catalog runs on CKAN today.
obamawhitehouse.archives.govnextgov.comecs.soton.ac.uklists.w3.orggithub.com
Credit
Vivek Kundra; Tim Berners-Lee and Nigel Shadbolt; Rufus Pollock and the Open Knowledge Foundation
2015
A breach nobody could size
Intruders took 21.5 million background check records, 4.2 million personnel records and 5.6 million fingerprints from the US Office of Personnel Management (OPM), the federal government's personnel agency. A House Oversight Committee report of 7 September 2016 said "The OPM data breach was preventable" and that OPM's leaders had not heeded repeated recommendations from its inspector general. Attackers had been inside OPM's network since July 2012, and investigators "will never know with complete certainty" everything they took, Nextgov reported from the committee's findings.
LessonWithout records of network traffic, a government cannot say what it has lost.
Credit
House Committee on Oversight and Government Reform
2019
Backups held abroad
Under an agreement with Luxembourg signed on 20 June 2017, Estonia keeps copies of state databases in a Luxembourg data center treated like an embassy. In June 2019 it moved the first four core databases there, including the land and business registers, NBC News reported, at a cost of 200,000 to 300,000 euros a year.
Credit
Governments of Estonia and Luxembourg
2022
Registers moved under fire
In February 2022, Amazon Web Services says, Ukraine's parliament passed a law allowing government data to move to the cloud, and around the time of Russia's full invasion the government moved "about 100 Ukrainian state registers and critical databases" to data centers abroad, by the count of Mykhailo Fedorov, the minister of digital transformation. Microsoft reported that Russia "targeted Ukraine's governmental data center in an early cruise missile attack".
LessonThe law that let the data leave the country had to be in place before the attack.
Credit
Ministry of Digital Transformation of Ukraine; Amazon Web Services and Microsoft, which hosted the data
2025
Federal data copied first
After the change of US administration in January 2025, federal datasets and web pages were taken down; the Congressional Research Service later found that Data.gov's count "fell by 2,000" and that datasets from six federal organizations appeared to have been removed entirely. Outside groups already held copies: Harvard Law School's Library Innovation Lab had harvested more than 311,000 Data.gov datasets between November 2024 and January 2025, and the End of Term Web Archive held more than 500 terabytes, including more than 100 million web pages.
LessonPublic records survived because outsiders copied them before they were needed, and the Federation of American Scientists warns that counts of datasets are "a poor measure" of what was lost.
everycrsreport.comlil.law.harvard.edublog.archive.orgfas.org
Credit
Harvard Law School Library Innovation Lab; End of Term Web Archive partners, including the Internet Archive
2025
Fire takes an unbacked drive
On 26 September 2025 a fire broke out while lithium-ion batteries were being moved at the Daejeon data center of South Korea's National Information Resources Service, and the government took all 647 of the center's systems offline. The G-Drive, a shared drive for civil servants that held 858 terabytes in August, had no outside backup because it was large, slow storage, and officials judged it unrecoverable. The ministry said all 709 affected systems were restored by 30 December 2025.
LessonThe store judged too big to back up was the one that could not be restored.
Credit
Korea's Ministry of the Interior and Safety; National Information Resources Service
Watch for
AI rewrites IRS code
The IRS lists several AI tools that aim to understand the code of a system written in "a 60-year-old coding language" and rewrite it in modern languages. GAO reported them in March 2026 without testing them, and found 77 of the IRS's 126 AI uses still in development.
Achieved whenGAO, the Treasury Inspector General for Tax Administration or another outside body confirms that rewritten code processes returns, with tests showing its results match the old system's.
Claimed by
Internal Revenue Service, in its AI inventory as reported by GAO, March 2026
COBOL rules read by AI
IBM says its watsonx.ai models pulled the business rules out of "roughly 6 million lines of code spanning 2500 COBOL and Assembler programs" at the California Department of Motor Vehicles, cutting a job estimated at five years to 15 months. The account is a contributed article written by IBM.
Achieved whenthe DMV, a state auditor or another outside reviewer evaluates the extracted rules, or a rebuilt system based on them goes live.
Claimed by
IBM, August 2026
Korea's backups rebuilt
After the fire, Korea's interior ministry promised "system redundancy and real-time backup" and set aside 343.4 billion won for disaster recovery upgrades and relocating the Daejeon center.
Achieved whenan outside audit, such as one by Korea's Board of Audit and Inspection, or a documented failover test confirms that critical systems and the replacement for the G-Drive are backed up off-site.
Claimed by
Ministry of the Interior and Safety, December 2025
Open
01
A COBOL translation test
A public set of at least 25 COBOL batch programs of the kind governments run, such as payroll, benefit calculations, sorted merges and summary reports, using fixed-width files, packed decimal numbers and rounding rules, each with made-up input files and the exact output the original produces. An AI translation into Java or Python passes only if its output matches the original byte for byte.
The briefCredit · unclaimed02
Overdue open datasets
A dated census of at least 20 US city and state open-data portals, listing for each dataset its stated update schedule, its last actual update, whether it is more than twice as late as its schedule allows, and whether its download link works.
The briefCredit · unclaimed
Open 01 · the brief
A COBOL translation test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 25 COBOL batch programs of the kind governments run, such as payroll, benefit calculations, sorted merges and summary reports, using fixed-width files, packed decimal numbers and rounding rules, each with made-up input files and the exact output the original produces. An AI translation into Java or Python passes only if its output matches the original byte for byte.
Why it is still open
GAO reported in July 2025 that eight of 11 critical federal legacy systems it reviewed use languages "such as COBOL and assembly language code". Published tests use short exercises: COBOLEval (Bloop, 2024) has models write COBOL versions of small programming problems, and the COBOL-JavaTrans set in the COBOL-Coder paper (April 2026) translates 143 such problems between COBOL and Java and checks them with unit tests. IBM's July 2025 evaluation of its own translator combines rule-based checks with AI judges. None scores whole batch programs with file input and output against the original's results.
Built with what exists
GnuCOBOL, a free COBOL compiler, to produce the reference outputs, and the NIST COBOL 85 test suite that ships with it as a source of standard code; Python with Hypothesis to generate input files; AI coding agents such as Claude Code or Codex to do the translations, run in a sandbox and judged only by the tests (see the Note "Code no person read").
First step
Write ten programs, including a payroll run and a report that totals by group, generate their input files and freeze the expected outputs.
Done looks like
25 programs with at least 1,000 made-up records each and hidden test files, first-attempt pass rates published for at least three AI systems, and a rerun by a second person that matches within 2 percentage points.
Open 02 · the brief
Overdue open datasets
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A dated census of at least 20 US city and state open-data portals, listing for each dataset its stated update schedule, its last actual update, whether it is more than twice as late as its schedule allows, and whether its download link works.
Why it is still open
Open Data Portal Watch, run by the Vienna University of Economics and Business, checks the metadata of 280 portals but asks only whether a dataset states its "update frequency", not whether it keeps to it, and the Open Data Barometer has not published a full edition since its fourth. The portals already publish what is needed: Chicago's catalog returns each dataset's last update and stated frequency through a public API, for 928 datasets.
Built with what exists
the Socrata Discovery API and the CKAN API; the federal DCAT-US metadata schema, in which the last-modified date is required and the update frequency optional; Python; light, spaced checks of public download links.
First step
Pull Chicago's 928 datasets, turn each stated frequency into a number of days and count the overdue ones.
Done looks like
code and a dated snapshot published, with the share of datasets overdue, the share with no stated schedule and the share of links failing at two checks a week apart, and a rerun by someone else that matches within one percentage point.