"Democratic backsliding" does not mean a coup. It means a legally elected chief executive weakening checks through a sequence of ostensibly lawful steps — replacing court personnel, adjusting electoral rules, reshaping media ownership. Each step stays within legal form and has a defensible rationale on its own, so there is no "coup moment" to react to; scholars call this executive aggrandizement. The crucial difference from a coup is that there is no threshold: opponents permanently face the question of whether the line has been crossed yet.
And it cannot be observed directly, only measured — which means the measuring instrument shapes the conclusion.
| Index | Method | Cost |
|---|---|---|
| V-Dem | Thousands of experts score hundreds of sub-indicators | Catches "still a democracy but degrading"; relies on subjective judgement |
| Freedom House | Scores against a political-rights and civil-liberties checklist | Comparable across years; compresses very different countries into one band |
| Polity-type indices | Look only at constraints on executive power | Immune to coder mood; insensitive to societal change |
The readings therefore diverge: on the V-Dem 2025 report's classification, roughly 72% of the world's population lives in countries it codes as autocracies, and only 29 qualify as liberal democracies — while on a coarse indicator such as "are there competitive elections," the absolute number of democracies remains near its historic high.
The decline camp: erosion happens in the gaps between coarse indicators. Elections can stay regular and the count honest while courts are packed and districts redrawn; by the time the coarse indicator moves, elections can no longer reverse it.
The sceptics (Levitsky and Way's The Myth of Democratic Recession; Treisman's tests against long historical data): first, a baseline illusion — the early 1990s were a historical outlier, and falling back from a peak is not decline; second, mood drift in expert coding — scores can slide when the atmosphere sours even though institutions have not changed.
The trade-off: one side accepts more false positives to raise detection; the other accepts more false negatives to stay robust.
"Backsliding means elections are cancelled" — the contemporary form keeps elections, and may cite the electoral mandate as grounds for weakening checks. The other is treating different indices as one ruler: they define democracy differently, so only the trend within a single index is readable.
Technocracy is not "experts in office." It is the insulation of a class of decisions from the electoral cycle, handing them to bodies that voters neither directly mandate nor easily remove — what scholars call non-majoritarian institutions.
It solves time inconsistency: promising today not to loosen next year is not credible, because by next year loosening will look like the better move. An engineering analogy: burning a setting into read-only firmware — the fact that it could be changed at any time is itself what stops downstream users from trusting it. The cost is symmetric: credibility comes from irrevocability, and irrevocability breaks the chain of accountability.
Fiscal councils and audit bodies do the opposite: they assess but decide nothing, trading effectiveness for the least contested legitimacy. The wider the delegation, the thinner the legitimacy.
The delegation case: wherever costs fall now and benefits fall later, electoral incentives lean toward deferral. Handing the decision to a body that need not seek re-election is really a voter's pre-commitment — the equivalent of putting the alarm clock across the room.
The democratic-control case: every real policy decision has distributive consequences, and those are not technical questions. Experts can agree within a method, but choosing the objective function and deciding who absorbs the adjustment costs is a value judgement; dressing it up as technical removes the most debatable part from debate.
The trade-off: one sacrifices immediate accountability for consistency over time; the other reverses it.
"Technocracy is neutral" — methods can be neutral, objectives cannot; where you set the relative weight on two conflicting goals is itself a distributive position. "Independent means unconstrained" — independence usually means instrument independence (how), not goal independence (what).
Representative government leaves two distinct gaps: distorted preference transmission (one vote bundles hundreds of issues into a single ballot) and missing deliberation (voters lack the time and information to dig in). Two routes each address one of them. A referendum is aggregative: everyone votes on a single question, which fixes the bundling problem but exposes the result to differences in campaign spending. A citizens' assembly by lot is deliberative: dozens to hundreds of randomly selected people get weeks of time and expert hearings, which fixes information and depth but not authorisation — they carry no electoral mandate.
Only three design variables matter: who sets the question, who takes part, and whether the conclusion binds.
The OECD database now holds 733 such cases, the vast majority of them advisory. The more technical and reversible the question, the better it travels; the closer to basic rights, the worse.
The expansion case: the distortion is structural — a legislator's first incentive is re-election, not thinking an issue through. People chosen by lot face no re-election pressure, and their conclusions diverge from their own starting positions far more than ordinary polling does, which suggests deliberation genuinely changes judgement.
The case for representation (the Madisonian logic): representation is a deliberate filter. It protects unpopular minorities, who fare worst under direct majority votes; it allows cross-issue bargains — conceding A to win B, which is how real majorities are assembled and which single-issue voting cannot do; and it supplies continuing, attributable responsibility, since once an assembly disbands nobody answers for the outcome.
The trade-off: one trades certainty of accountability for quality of judgement; the other reverses it.
"A referendum result is the public will" — the outcome depends heavily on who sets the question, how it is worded, and when it is put, so who holds the power of proposal often matters more than who holds the vote. "Sortition is just grabbing people at random" — it is stratified random selection: after a mass random invitation, participants are sampled by age, region and education to approximate the population, correcting precisely for the bias that volunteering introduces.
Traditional lawmaking assumes a rule stays valid long after passage, so everything must be thought through in advance. When change outpaces the legislative cycle, that assumption fails. The shared idea behind governance innovation is to write uncertainty into the design: replace one-shot general decisions with experiments bounded in scope and time. Besides pilots and controlled evaluation, three tools recur:
| Posture toward new technology | Approach | Cost |
|---|---|---|
| Prior authorisation | No licence, no operating | High barrier for entrants; regulator must foresee the risk list |
| Sandbox | Admit a few, set rules while watching | Selective enforcement; sustained close contact between regulator and regulated |
| Ex-post liability | Allow first, compensate or punish after | Harm is often irreversible; small actors cannot pay |
The sandbox's cost deserves stating plainly: entrants receive compliance exemptions that non-entrants do not, which sits awkwardly with formal equality before the law, and dense interaction between a regulator and a handful of firms is exactly the condition under which regulatory capture arises. Digital government is no different — the once-only principle dismantles the privacy barrier that inter-agency friction had unintentionally provided, and Estonia's compensation is to substitute auditability for inaccessibility.
The iteration case: without trials there is no evidence, and legislation without evidence is guesswork. Because amending law is politically expensive, rules with no built-in correction mechanism tend to survive long after they stop working.
The stability case: a large part of law's value comes precisely from not changing — predictability is what lets people invest long-term and sign long contracts. An institution that iterates constantly may lower total welfare through unpredictability even if its average decision is better; and rules governing property, procedure and basic rights are the worst candidates for split testing, since a trial means contemporaries living under different rules.
"Digitisation equals better governance" — moving a redundant process online only makes a bad process run faster; improvement comes from redesigning the process and from the legal authority to share data across agencies. "Innovation equals deregulation" — sandboxes and sunset clauses are also tightening tools: they make rules easier to withdraw.