Governments across the world are deploying artificial intelligence at a pace few citizens realize. In the United States alone, federal agencies documented 3,611 separate AI use cases across 56 departments in 2025 — more than double the count from the year before. Estonia runs an AI-powered assistant across 21 public agencies. Singapore’s digital identity system handles 300,000 AI-assisted transactions a day. The question is no longer whether governments use AI, but how — and what that means for the people on the receiving end.
What governments are doing with AI
The broadest category is citizen-facing services. Virtual assistants and chatbots handle routine inquiries around the clock — checking benefit status, guiding permit applications, answering tax questions — without the wait times of a phone queue. Estonia’s Bürokratt is one of the most advanced examples: a voice and text assistant that lets citizens navigate the full range of government services through a single conversational interface, deployed across agencies from the tax authority to the police and border guard.
Document and form processing is the second major category. AI extracts information from submitted documents, flags incomplete applications, and routes cases to the right team — compressing what once took weeks to days for standard submissions. Welfare agencies use this to triage claims; immigration services use it to screen documents at volume.
Fraud and tax compliance is where governments have invested most heavily. The UK’s HMRC has run its Connect system for over a decade: it cross-references employment records, bank data, property filings, and other sources to identify taxpayers at elevated risk of evasion. Singapore’s SATIS system analyzed hundreds of thousands of URLs daily and disrupted more than 50,000 scam-linked websites by late 2024.
Law enforcement and public safety is the most sensitive category. Predictive analytics tools estimate where and when crime is likely to occur. Facial recognition is used at some border crossings and transport hubs. Risk-scoring systems assess recidivism probability for bail and sentencing decisions — applications that carry the most direct implications for individual rights.
What citizens gain
The clearest benefit is speed. Routine document processing that once took weeks can take days; automated eligibility checks can run in hours. Fraud detection, when accurate, directs resources to legitimate claimants faster. AI-powered chatbots provide answers at midnight without requiring office hours. The OECD tracks these benefits across its member countries and consistently finds productivity gains in routine administrative work.
The cost savings can also reach citizens indirectly. Analysts at Deloitte estimated that automating repetitive federal tasks in the US could save tens of billions of dollars annually — funds that could in principle be redirected to services.
What can go wrong
The risks are substantial and well-documented.
Algorithmic bias is the most systemic concern. AI systems trained on historical government data can reproduce — and entrench — existing inequalities. The clearest legal precedent is the Netherlands’ SyRI system, a welfare fraud detector struck down by a Dutch court in February 2020 on human rights grounds. The court found that SyRI violated citizens’ right to private life: it had been deployed exclusively in low-income and minority neighborhoods, collected excessive data, and operated without meaningful transparency. The UN called it a landmark ruling.
Opacity compounds the problem. Machine learning models are often difficult to explain even to the engineers who built them. When a benefit is denied or a tax return flagged, citizens are entitled to a reason — but “the algorithm decided” is not one.
Accountability gaps emerge when decisions are partially automated. It becomes unclear whether the agency, the official, or the AI vendor is responsible for a harmful outcome — and citizens may have limited recourse to challenge it.
Surveillance creep is a longer-term concern. Systems built to detect fraud or crime can, if unchecked, become infrastructure for broader monitoring of citizens who have done nothing wrong.
How government AI is different
Private companies can be avoided. Government services usually cannot. A citizen flagged as a fraud risk by an automated welfare system cannot simply switch to a competitor. This is precisely why the EU’s AI Act classifies government AI in welfare, migration, law enforcement, and justice as high-risk — requiring human oversight, transparency to affected persons, and documented risk management before deployment. Social scoring by public authorities and real-time biometric identification of citizens in public spaces are prohibited entirely.
The OECD’s AI Principles, originally adopted in 2019 and updated in 2024, make a point that often goes unnoticed: governments are “not only AI regulators and investors, but also developers and users” — and should hold themselves to the same accountability standards they set for the private sector.
Why it matters for Georgia
As an EU candidate country, Georgia will be expected to align with EU AI Act standards progressively — including the stricter rules on government-deployed AI. Public agencies introducing AI into benefits administration, tax, or law enforcement will need to meet the same transparency and oversight requirements now taking effect across Europe. Tbilisi has already shown active interest in this agenda: the city hosted a UN forum focused on AI in public governance, signaling that these questions are on the national policy radar. For a deeper look at what the EU framework means specifically for Georgian institutions, see our explainer on the EU AI Act and Georgian companies.
In the news
California became the first US state to deploy an AI assistant across every state government agency, in a landmark deal with Anthropic. See our report on California’s statewide Claude deployment.
FAQ
Are citizens told when AI is involved in a decision that affects them?
Rules vary by jurisdiction. The EU AI Act requires high-risk AI systems to disclose their use to affected individuals. Outside the EU, disclosure obligations are patchy and inconsistently enforced.
Can I opt out of government AI?
Rarely. Unlike private services, government agencies typically offer no alternative channel — citizens interact with whatever system the agency uses, regardless of personal preference.
What makes a government AI system trustworthy?
Transparency about when and how AI is used, mandatory human review for consequential decisions, independent auditing, and a clear accountability chain — a named official who is responsible when something goes wrong.
Is government AI more regulated than private-sector AI?
Yes, under the EU AI Act. Government applications in welfare, law enforcement, migration, and justice are automatically high-risk or prohibited — subject to stricter obligations than equivalent commercial AI. Other jurisdictions are still developing comparable rules.
Sources: Wikipedia — AI in Government · OECD AI in Government · EU AI Act · US Federal AI Use Case Inventory 2025 · Netherlands SyRI ruling · Estonia Bürokratt