What changes for government with AI?

The shift

Drafting, synthesizing, and translating the raw material of governing — case files, regulations, comments, correspondence, legislative text, benefit applications — goes from scarce staff-hours to near-free and instant. Government's core constraint was never having enough people to read, write, and respond at the volume citizens and problems generate; that constraint just loosened.

The axioms

• Casework and correspondence require a trained person to read the file and draft the response. Rests on synthesis-and-drafting being scarce, expensive labor.
• Rulemaking and legislative drafting are bottlenecked by the number of policy staff and lawyers who can produce and vet text. Rests on legal/technical drafting expertise being scarce.
• Public comment periods work because reading and weighing comments is slow enough to force selection and summarization. Rests on synthesis of large unstructured input being expensive.
• Bureaucratic delay is a proxy for care — waiting in line, processing time, and paperwork friction filter out low-stakes or fraudulent requests. Rests on process being costly to attempt at scale.
• Only government (or contracted experts) can translate technical/legal language into citizen-facing language, and vice versa. Rests on translation between registers being a scarce skill.
• A human civil servant must sign off on decisions that use public money or restrict citizens' rights. Rests on accountability being a property only a liable person can hold.
• Enforcement and service delivery require boots on the ground — inspectors, caseworkers, benefits officers physically verifying and acting. Rests on physical action being irreducible.
• Citizens trust government because a known, accountable institution — not an anonymous process — stands behind the decision. Rests on trust and legitimacy being tied to identifiable human/institutional accountability.
• Novel policy problems (a new technology, an emergent crisis) require judgment calls nobody has made before. Rests on judgment under genuine novelty being scarce.
• Democratic legitimacy requires elected humans, not administrators, to decide what the state should want (goals, tradeoffs, values). Rests on taste/goal-setting for the polity being a political, not technical, act.

Invalid axioms

1. Casework and correspondence require a trained person to read the file and draft the response. Reading a citizen's file, matching it against rules, and drafting a first-pass response is exactly what LLMs do well — synthesis plus drafting at near-zero cost. The habit-trap: agencies still staff call centers and casework queues sized for a world where every reply required a human to compose it from scratch, and budget backlogs as an inevitable cost of headcount rather than a solvable queuing problem.
2. Public comment periods work because reading and weighing comments is slow enough to force selection and summarization. Summarizing tens of thousands of comments, clustering duplicates, and surfacing genuinely novel objections is now fast and cheap. The habit-trap: agencies still treat comment-period staffing and timelines as if synthesis were the bottleneck, when the actual scarce step — deciding which objections change the rule — was never the reading, it was the judgment.
3. Bureaucratic delay is a proxy for care. Delay filtered out low-effort submissions because submitting was costly for the citizen too. AI collapses submission cost for citizens and processing cost for agencies simultaneously, so delay stops functioning as a quality filter — it becomes pure friction with no signal value. The habit-trap: keeping slow multi-step processes "for rigor" when the rigor was never in the waiting, it was in verification that AI can now do faster and better.
4. Only government or contracted experts can translate technical/legal language into citizen-facing language, and vice versa. Plain-language translation of statutes, benefit rules, tax code, and citizen intake in any language is now abundant. The habit-trap: agencies still commission expensive plain-language rewrites and translation contracts as one-off projects instead of treating translation as a default, always-on layer.
5. Rulemaking and legislative drafting are bottlenecked by the number of staff who can produce and vet text. First-draft regulatory language, redlines against existing statute, and cross-jurisdiction comparisons are now fast to generate. The habit-trap: legislative and regulatory offices still size drafting teams and timelines around hand-drafting, when the bottleneck has moved to review and political negotiation.

Unchanged axioms

1. A human civil servant must sign off on decisions that use public money or restrict citizens' rights. Accountability requires someone who can be fired, sued, or voted out. A model cannot be liable, so every AI-assisted benefits denial, enforcement action, or permit decision still needs a human who owns the outcome — this doesn't shrink, it becomes more load-bearing as AI-assisted drafts multiply.
2. Enforcement and service delivery require boots on the ground. Inspecting a building, arresting someone, delivering a vaccine, showing up at a flooded home — none of this is token generation. AI can prioritize which building to inspect first; it cannot inspect it.
3. Citizens trust government because a known, accountable institution stands behind the decision. Legitimacy is a relationship, not an output. An AI-drafted letter with a real agency's signature only carries authority because a human institution is answerable for what it says — strip that and it's just plausible text.
4. Novel policy problems require judgment calls nobody has made before. AI pattern-matches against precedent; a genuinely new situation (a new technology's regulation, a first-of-its-kind crisis response) has no precedent to match against. This is where confidently-wrong output is most dangerous, because there's no ground truth to check against yet.
5. Democratic legitimacy requires elected humans to decide what the state should want. Setting the goal — what tradeoffs a society accepts, whose interests win — is a political act, not a synthesis problem. AI can lay out options and consequences; it cannot have standing to choose on the polity's behalf.

New axioms

1. When drafting responses, rules, and filings is free for citizens too, government must solve for volume it can no longer use delay to throttle. AI lets citizens generate polished appeals, comments, FOIA requests, and grant applications at the same scale agencies can generate replies — the friction that kept both sides' volume in rough balance is gone from both directions at once.
2. When AI can draft a plausible-looking legal or regulatory analysis instantly, government must solve for verifying which drafts are actually grounded in current law versus confidently invented. Confidently wrong is the default failure mode; a hallucinated citation or misstated rule embedded in a benefits denial or court filing does real harm before anyone catches it.
3. When every citizen and every foreign actor can generate high-volume, personalized, plausible-sounding petitions, comments, and disinformation, government must solve for distinguishing genuine public sentiment from synthetic astroturfing at scale. The comment period, the petition, the constituent letter — all signals of what the public wants — get noisier exactly as government gains the tools to listen to them faster.
4. When AI can draft and personalize enforcement, surveillance, and eligibility decisions cheaply, government must solve for who is accountable when an AI-assisted decision at scale is systematically wrong for a whole category of people, not just one case. Individual sign-off doesn't catch a pattern error replicated across ten thousand cases in an afternoon.
5. When AI makes translating and personalizing government communication free, government must solve for which channel citizens can trust is actually from the state and not an impersonation. Cheap, fluent, personalized official-sounding text is also the exact tool of scam and impersonation — the same capability that lets government serve citizens better lets bad actors impersonate government better.

Where it breaks

Agencies now push AI-assisted decisions (benefits eligibility, enforcement flags, permit denials) at volumes only possible because drafting got cheap (INVALID #1) — while accountability still requires a human to actually own each outcome (STILL HOLDS #1) and nobody has solved for catching systematic errors across thousands of AI-assisted cases before real harm compounds (NEW #4). The sign-off becomes a rubber stamp on volume no single reviewer can meaningfully check, which is the opposite of what sign-off was for.

A second collision: agencies drop bureaucratic delay because it no longer signals rigor (INVALID #3) at the same moment citizens and bad actors alike can flood comment periods and petitions with cheap, plausible, personalized submissions (NEW #3). Removing friction from processing while submission friction also disappears on the public side means both queues fill at once, with no remaining mechanism — neither delay nor manual reading — to separate genuine signal from synthetic volume.

ai-axioms.vercel.app/axioms/government/government-ai-axioms