The US federal government spends $90 billion a year on technology, and 75% of it goes to maintaining legacy systems before a dollar reaches the modernisation AI is meant to bring.
Mehr, Harvard Ash Center, 2025
Single-sourceFully one-third of the Social Security Administration's workforce of nearly 22,000 was expected to retire by 2022, while citizens ask for more digital government.
Mehr, Harvard Ash Center, 2025
Where one AI platform ran social insurance, 71% of requests were filed online, against 28% before it, and individual tasks took about 93 seconds to process.
Zheng et al. (AAAI/IAAI), 2025
21 experts in AI, policy and disinformation rated technical deepfake detection the least effective defence, at about 3.4 out of 7.
Loth, Kappes & Pahl (ACM WWW), 2026
Participation, faster casework, self-service and translation: what government AI delivers when it works.
São Paulo's Participe+ platform lifted participation in the city's citizen budget from 2,097 people in 2019 to 12,354 in 2020, a 500% increase.
Sousa et al., 2025
Single-sourceAfter a year in three Chinese cities, the SmartHS platform served 2.92 million social-insurance cases with only 37.5% of the original front-desk staff, and cut residents' travel by 20.6%.
Zheng et al. (AAAI/IAAI), 2025
Peer-reviewedIBM Watson's MySurrey app in British Columbia answers 65 percent of the questions that already have answers on city websites, after studying over 3,000 documents on 16 services.
IBM case, cited in Mehr, Harvard Ash Center, 2025
Company-reportedOnly 27% of US citizens were satisfied with government digital services, but 92% said improved digital services would positively change their view of government.
Accenture survey, cited in Mehr, Harvard Ash Center, 2025
Company-reportedUnbabel combines crowdsourcing and machine learning to translate into 14 languages, cited as a model for translating government information.
Mehr, Harvard Ash Center, 2025
Company-reportedOf 250 real AI use cases across European public administrations, 76 (30%) were used to streamline internal management and administrative procedures.
van Noordt & Misuraca, Government Information Quarterly, 2025
Peer-reviewedA sovereign public-services assistant built on an open 70B model correctly declined 98% of out-of-domain questions rather than inventing answers.
Branco et al., 2026
Single-sourceParticipatory budgeting, recognised by the UN as a democratic-governance best practice, had spread to more than 7,000 cities by the 2020s.
Sousa et al., 2025
Single-sourceNo verified figure exists for how much of a measured turnout lift lasts into the next cycle. The available panels stop at twelve months, and longer projections were excluded rather than extrapolated.
We could not establish what a digital-service gain costs per resident. Agencies do not report AI spend at that grain, and vendor breakdowns cannot be joined to public budgets.
Synthetic campaign content, compute with no owner, uneven participation and machines checking facts.
Among the 139 most influential political accounts on Instagram in the 2024 US election, memes were 27.6% of content and AI-generated images about 8.4%, against 10% and 5.2% across all users.
Chang et al., 2024
Single-sourceA review of 775 non-US data centres covered $360 billion of investment; 36% of that value had no identifiable operator at all.
Richardson et al. (RAND), 2026
Single-sourceParticipatory budgeting spread unevenly: by the early 2000s, about 140 Brazilian municipalities had it; two decades on, nearly half of the 250 largest cities did, handing citizens 5–15% of budgets.
Sousa et al., 2025
Single-sourceOf AI incidents logged in the public AI Incident Database from May to November 2025, 61% were misuse, 27% unreliable or hallucinated outputs and 5% supply-chain compromises.
Huwyler, 2025
Single-sourceAsked to classify more than 20,000 previously fact-checked claims, ChatGPT agreed with human raters about 70% of the time: plausibility, not proof.
Summerfield et al., 2024
Single-sourceNo verified figure exists for political advertising bought through resellers outside the disclosing platforms. Every published number is a modelled estimate.
The share of synthetic political content that was labelled as such cannot be established: most platforms carry no machine-readable label field.
Risky vendors, synthetic images, a few superspreaders and surveillance inside democracies.
xAI's Grok won a Pentagon contract worth up to $200 million in the same week the chatbot posted antisemitic content, calling itself “MechaHitler”, that the company had to remove.
Newsweek, 2025
Single-sourceIn a large X dataset on the 2024 US presidential election, about 12% of shared images were AI-generated.
Chen, Ye, Tsai, Ferrara & Luceri, 2025
Single-sourceAround 10% of users were responsible for sharing 80% of the AI-generated images: a small group of superspreaders drives synthetic politics.
Chen, Ye, Tsai, Ferrara & Luceri, 2025
Single-sourceReal-time facial-recognition vans are used by 13 local UK police departments, one of several AI surveillance systems now operating inside democracies.
Sania et al., 2026
Single-source53% of the projects reviewed, and 48% of AI projects, had no identified operator: compute that shapes policy, owned by no one the public can name.
Richardson et al. (RAND), 2026
Single-sourceReading notices of funding opportunity directly, researchers found 407 grant opportunities that fund AI or encourage grantees to use it, more than aggregate reporting suggests.
Bateyko & Levy (ACM FAccT), 2025
Peer-reviewedIn the absence of federal regulation of AI vendors, procurement remains one of the few levers governments have to push for safety, non-discrimination, privacy and accountability.
Johnson, Silva & Heidari, TechPolicy.Press, 2025
Single-sourceA twelvefold savings range, plans that skip the money and incidents nobody reports.
Automating federal employees' tasks could save between 96.7 million and 1.2 billion hours a year, worth $3.3 billion to $41.1 billion: a twelvefold spread that shows how little is settled.
Deloitte, cited in Mehr, Harvard Ash Center, 2025
Company-reportedReviewing 22 federal agency plans for complying with White House AI guidance, researchers found only two mentions of grants or financial assistance, though agencies fund AI through grants.
Bateyko & Levy (ACM FAccT), 2025
Peer-reviewedSupply-chain compromises were 5% of logged incidents, and bias and drift barely appear at all, likely because they are internal failures that are rarely disclosed.
Huwyler, 2025
Single-sourceExperts rated AI-generated text a 6.1 out of 7 threat, a systemic risk that pollutes everyday discourse; deepfake video scored 6.4 but is more readily debunked.
Loth, Kappes & Pahl (ACM WWW), 2026
Single-sourceOf more than 220,000 AI studies in Web of Science, about 6,000 concern governance or regulation, and 558 relate to the EU AI Act's categories.
Kim et al. (IEEE), 2025
Peer-reviewedBoth sides of this page are sourced to the same standard: every figure names its origin and its confidence, and where no verified number exists the page says so. If the missing records were published tomorrow, which of these numbers would you expect to move?
The Deloitte, Accenture and IBM figures come via Mehr's Harvard Ash Center paper, as it cites them.
