Should citizens trust a government decision that was AI-assisted, and who do they appeal to?

The shift

Producing a government decision — matching a case to the rules and generating a rationale that reads as considered — goes from scarce staff-judgment to near-free and instant. The rationale detaches from the deciding: fluent, defensible-sounding reasons are now abundant whether or not they describe why the outcome actually landed where it did.

The axioms

• A citizen trusts a decision because a human weighed their case and is answerable for it. Rests on an accountable human decision being the thing that produced the outcome — and on the reasoned letter being reliable evidence that one happened.
• Due process gives the citizen a right to appeal to a responsible party who can revisit the decision. Rests on there being someone with the standing and discretion to be persuaded and to reverse.
• The state owes a reason it can defend — transparency and explainability are a democratic duty, not a courtesy. Rests on the stated reason actually being the reason the decision was reached.
• A decision that uses public money or restricts rights must carry an accountable human's sign-off. Rests on accountability being a property only a liable person can hold.
• Democratic legitimacy for a decision comes from its being made under authority answerable to citizens. Rests on the chain from vote to official to decision being real, not nominal.

Invalid axioms

1. A reasoned, individualized letter is reliable evidence that a human weighed your case. Reading a file, matching it to rules, and drafting a specific, well-argued rationale is exactly what LLMs make abundant — so the artifact citizens used as their signal of human attention no longer carries that signal. The belief that this letter means someone considered me was true when composing it required a person to have read the file; it isn't anymore. The habit-trap: agencies, courts, and citizens still read the presence of a tailored rationale as proof of human deliberation, and design notice-and-explanation requirements around producing text rather than around evidencing that a human actually decided.

Unchanged axioms

1. A citizen trusts a decision because a human is answerable for it. Trust was never in the prose; it was in there being someone who can be complained to, sued, fired, or voted out. A model cannot be answerable, so an AI-assisted decision earns trust only to the extent a named human or institution owns the outcome — this gets more load-bearing as plausible rationales multiply, not less.
2. Due process requires a responsible party who can actually revisit and reverse the decision. The right to appeal is worth something only if it routes to someone with the discretion and standing to be persuaded. Generating a fresh rationale on appeal is now cheap; being genuinely open to reversing is not, and it's the part that was never automatable.
3. The stated reason must be the reason the decision was reached. Explainability as a duty survives, but its content just got harder: a defensible-sounding explanation is abundant, an explanation that faithfully reflects the actual basis of the decision is not. The duty was always to the real reason; AI makes the counterfeit cheap without touching the obligation.
4. A decision restricting rights or spending public money needs an accountable human's sign-off. Cheaper decision-drafting doesn't make liability cheaper. If anything the sign-off matters more, because the volume of AI-drafted decisions a single accountable human is nominally standing behind can now exceed what any human could actually have judged.
5. Legitimacy comes from authority answerable to citizens, exercised for real. Setting the rule and owning its application is a political act with standing behind it. AI can apply the rule and articulate the call; it cannot hold the authority or answer to the electorate for it.

New axioms

1. When citizens can't tell whether a human decided, government must solve for evidencing the human role, not just the output. The old proxy — a considered-looking rationale — is dead as a signal. Nothing yet reliably tells a citizen whether their denial was owned by a person who read the file or emitted by a system with a name rubber-stamped on it, and citizens can't calibrate trust without knowing which.
2. When generating a fresh rationale is free, government must solve for appeals that route to genuine reconsideration rather than a re-run of the same model behind a new signature. An appeal answered by regenerating a plausible justification for the original outcome is procedurally intact and substantively hollow — the citizen exercised a right that reversed nothing because nothing was actually reconsidered.
3. When a defensible-sounding explanation is abundant, government must solve for whether the stated reason is the operative reason. A generated rationale can be fluent, lawful-looking, and disconnected from what actually drove the decision — a post-hoc story, not a cause. Explainability requirements met by such text satisfy the letter of the duty while defeating its purpose, and a hallucinated citation or invented basis does harm before anyone checks.
4. When automated decisions issue at scale, government must solve for the legitimacy of decisions no accountable human meaningfully touched. Individual sign-off was designed for a volume a human could actually judge. A named official standing behind ten thousand AI-issued decisions in an afternoon is accountable in form; whether that's democratic legitimacy or its costume is the open question — and the answer is moving as agentic systems take more of the decision loop.

Where it breaks

Agencies satisfy explainability duties by attaching a generated, defensible-sounding rationale (NEW #3) while the citizen's right to appeal still assumes a responsible party who can revisit the actual basis of the decision (STILL HOLDS #2). The appeal then routes to a human handed a rationale that was never the real reason — so reconsideration operates on a fiction, and the more fluent the generated explanation, the harder it is for anyone in the chain to notice there's nothing underneath it to reconsider.

A second collision: the individualized rationale no longer evidences that a human decided (INVALID #1), yet legitimacy still depends on an accountable human actually owning each decision (STILL HOLDS #4). At AI decision volume, the sign-off becomes a name on output no person could have judged — accountability preserved in form, hollowed in substance — and neither the citizen nor the reviewer has a way to tell a genuinely owned decision from a costumed one. How fast this bites depends on how much of the decision loop agentic systems take without a human in it, which is moving quickly.

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