How divided can a society be and still be a democracy?

Somewhere between an argument and a civil war there is a line, and democracies are built on the assumption that everyone can see it.

The line is nowhere in the constitution. No document instructs a defeated prime minister to telephone the winner, or tells a government with the votes to pack a court to decline, or requires a party beaten by forty thousand votes to behave as though the count were honest. These are habits — and, on the evidence of the past decade, the part of democracy that fails first.

Which reframes the usual question. It is not whether an election will be held; elections are held nearly everywhere, including in places nobody would call free. It is how much division a political community can carry before the people inside it quietly stop wanting the arrangement to work.

The rules that are nowhere written down#

The conceptual housekeeping here arrived last year. In Defining and Measuring Democratic Norms, published in volume 28 of the Annual Review of Political Science, Gretchen Helmke and Josiah Rath argue that political science has been using the word "norm" far too loosely to explain anything with it. The complaint is not pedantic. If a norm is any regularity in behaviour, every change in behaviour becomes a violation, and the concept explains nothing. They call for more precision in defining norms and more care in labelling them democratic, and they organise the literature around who transgresses and who enforces — elites or citizens — because a norm with no enforcer is a wish.

The best-known inventory remains that of Steven Levitsky and Daniel Ziblatt in How Democracies Die (2018): mutual toleration, the acceptance that one's rivals are legitimate contenders rather than enemies of the nation, and institutional forbearance, the restraint that stops a government from using the full extent of the powers it legally possesses. Both are, strictly, unnecessary. A government that packs a court, impounds funds or governs by decree may break no law at all. Forbearance is the decision not to.

To that pair add a third, less often named: courts are everyone's courts. Not a chamber one side owns and the other endures.

Two things are called polarisation, and only one of them is a problem#

The word does too much work. Ideological polarisation is distance between positions: two blocs that want genuinely different tax systems, borders, energy policies. That is not a pathology. That is what elections are for. A society whose parties agreed about everything would have no use for voting.

Affective polarisation is something else. In the review that made the term standard — Shanto Iyengar, Yphtach Lelkes, Matthew Levendusky, Neil Malhotra and Sean Westwood, writing in the Annual Review of Political Science in 2019 — it is measured not as distance between policies but as dislike and distrust between people: partisanship as a social identity, with the ordinary machinery of in-group loyalty and out-group suspicion attached. One can be ideologically far apart and perfectly civil, or ideologically rather close and loathe the other side.

Hold the distinction. Almost every confused argument about polarisation collapses the two.

The staircase, and where the step is missing#

Jennifer McCoy, Tahmina Rahman and Murat Somer named the destructive variant in the American Behavioral Scientist in 2018: pernicious polarisation, in which a society's cross-cutting cleavages — region, class, faith, occupation — stop cutting across and stack up into a single line, leaving two camps that each regard the other as the country's principal danger.

The descent has a grammar, worth setting out plainly, because each step sounds reasonable to the person taking it.

You are wrong. This is democracy working.

You do not belong here. The first break: the argument has stopped being about the policy and become about the standing of the person holding it.

Your election is illegitimate. Mutual toleration is gone, and with it the mechanism by which power changes hands without force.

Removing your rights is legitimate. The vote is now a formality. The polity has decided that entitlements are to be weighed by group advantage.

The empirical question is how tightly the second step is fastened to the fourth. On the newest evidence: less tightly than the alarm suggests, and differently in different countries.

The link is contextual, not mechanical#

The most ambitious test to date appeared in the European Journal of Political Research in May. Eelco Harteveld, Lars Erik Berntzen, Andrej Kokkonen, Haylee Kelsall, Jonas Linde and Stefan Dahlberg ran a survey experiment across nine democracies — Brazil, France, Germany, Italy, Poland, Spain, Sweden, the United Kingdom and the United States — with roughly 18,000 participants, asking what actually follows when you turn the temperature down.

The intervention worked, up to a point. It reduced hostility towards opposing partisans in six of the nine countries, by between three and ten points on the standard feeling thermometer, and respondents became less inclined to avoid or discriminate against political opponents.

But the democratic pay-off was patchy. In four countries — the United Kingdom, Poland, Germany and Sweden — participants showed higher support for the right of opposing parties to take part in elections. In Italy and the United States the effect was minimal or ran the other way, even though the treatment itself had worked. The authors' conclusion is the one worth carrying: the consequences of affective polarisation may be more context-dependent than the literature has assumed. Reduced hostility does not automatically convert into democratic commitment. The exchange rate is set locally.

Norms behave like sticky attitudes#

A second result, published in the Journal of Experimental Political Science at the end of July, cuts in a related direction. Libby Jenke moved beyond the single-shot design that dominates this field, running a three-wave panel on a US sample recruited through Prolific — around 1,563 respondents between the first two waves — correcting misperceptions about the other side's demographics, policy positions and adherence to democratic norms repeatedly rather than once.

The honest summary is narrower than the headline. Jenke found no significant first-stage effects: the repeated treatments did not shift affective polarisation at all, and because the first stage failed, downstream effects on norm support could not be identified. She attributes part of the null to measurement, having used the gap between in-party and out-party warmth where earlier successful studies used out-party ratings alone.

That is not proof that norms are immovable. It is evidence that the brief corrective interventions performing so well in one-off experiments do not obviously accumulate, and that whatever holds these attitudes in place is more durable than a paragraph of accurate information.

Chile: the effect is real, and it does not need parties#

Against the nulls, set a finding that goes the other way and is cleaner for it. Loreto Cox, Pedro Cubillos and Carmen Le Foulon, writing in Political Science Research and Methods, ran an experiment two weeks before Chile's constitutional plebiscite of 4 September 2022. The treatment was almost crude: respondents were asked to list a few things they disliked about voters on the other side. Controls were asked about telemarketing calls.

Among the 85 per cent of the sample with a clear voting preference, the prompt raised affective polarisation — and support for democracy fell by roughly 0.12 to 0.14 of a standard deviation. The significance lies in the setting. Chile is a country of famously weak party attachment. You do not need a lifelong partisan identity for us-and-them feeling to erode the commitment to democracy itself. An invitation to dwell on the other side will do.

America takes its own temperature#

In its September 2025 wave, the Bright Line Watch project asked academic experts and the public to rate American democratic performance from 0 to 100. Experts gave 54; the public 49. The striking figure is the split beneath the public number: Republicans rated the system at 58.5, Democrats at 43.6 — a gap of 15 points, the widest the project has recorded, against a previous maximum of 12 in July 2018 and an average across earlier surveys of 8.

Two features of the wave deserve attention. The first is asymmetry. Asked about mid-decade redistricting, 92 per cent of experts judged the Texas plan a threat to democracy, while opinion on California's counter-move split three ways — 38 per cent a threat, 32 per cent a benefit, 30 per cent neither. Among the public, 64 per cent of Republicans approved of their party redistricting first; 8 per cent of Democrats did.

The second is that the erosion is not one-sided. Democratic support for several norm violations has risen since 2021 in the same series: approval of redistricting in defiance of the courts from 12 to 27 per cent, of restricting Fox News from 37 to 50 per cent. None commands majority support. But the direction is the direction.

Two elections that ran the tape backwards#

The pessimism is easy to over-fit. International IDEA's Global State of Democracy 2025 counts 74 national elections in 2024 and reports that representation scores fell to their lowest level in more than twenty years, with seven times more countries declining than advancing; press freedom deteriorated in 43 countries, the broadest decline since the dataset began in 1975. And yet the same report identifies two reversals, both produced by voting.

Brazil in 2022 was as bitterly divided as any democracy in the sample, and the division did not evaporate: on 8 January 2023 crowds stormed Congress, the Supreme Court and the presidential palace. The institutions held. In September 2025 a five-judge panel of the Supreme Court convicted Jair Bolsonaro of leading a criminal conspiracy to overturn the 2022 result and sentenced him to 27 years, with one justice dissenting. IDEA records Brazilian advances on ten democracy factors between 2022 and 2024 — a reminder that courts can be the load-bearing institution, as Brazilian constitutional argument has long recognised.

Poland is the subtler case, because the election was not clean and the outcome was still democratic. The OSCE's Office for Democratic Institutions and Human Rights found the parliamentary election of 15 October 2023 competitive, with a wide choice of alternatives, while concluding that the ruling party enjoyed a clear advantage through undue influence over state resources and the public media. Power changed hands anyway. IDEA records improvement on six factors between 2023 and 2024.

Scale is not the same as fairness#

Against which set India, where elections run at a scale no other polity attempts and Freedom House nonetheless rates the country Partly Free, at 62 out of 100 in its 2026 edition — 31 of 40 on political rights, 31 of 60 on civil liberties, down a point on the year. Its assessment notes that while India is a multiparty democracy, the government led by Narendra Modi and the Bharatiya Janata Party has presided over discriminatory policies and rising persecution of the Muslim population.

Holding an election is not the same as maintaining an equal political field. V-Dem's 2026 report counts 44 countries autocratising against 18 democratising, with 7 per cent of the world's people living in liberal democracies — the lowest share in over half a century.

Defences, and one genuine dilemma#

Norms can also be improvised. In France, after Emmanuel Macron dissolved the National Assembly on 9 June 2024 and the National Rally and its allies led the first round with 33.2 per cent, 134 candidates of the New Popular Front and 82 of Ensemble withdrew before the run-off to avoid splitting the vote. The National Rally bloc finished with 142 seats in a hung parliament. An unwritten rule functioning as institutional defence — at a cost worth stating: voters in those seats had a choice removed by party managers.

South Africa tested a different norm. The ANC took 40.18 per cent and 159 seats in May 2024, losing the majority it had held since 1994, and formed a government of national unity spanning some seventy per cent of parliament. The winner-takes-all convention broke gracefully. But analysts are right to warn that a coalition of nearly everyone thins out accountability: when the government contains most of the parliament, it is not obvious who is opposing.

Mexico poses the hardest case, and it deserves both arguments. The 2024 constitutional reform put federal judges and Supreme Court justices to popular election; the first such contest was held on 1 June 2025. The case for it is not frivolous — an unelected, self-reproducing judiciary widely believed to serve those who could afford it is a democratic problem, and submitting it to the electorate is a recognisably majoritarian answer. The case against is structural rather than partisan. Turnout was about 13 per cent, the lowest in the country's democratic history, with more than a fifth of ballots blank or spoiled; and a judge who must assemble a constituency has, thereafter, a constituency. Nancy Bermeo's category from her 2016 essay in the Journal of Democracy fits uncomfortably well: executive aggrandisement, the hollowing of checks through changes that are legal, popular and individually defensible, and vexing precisely because they are legitimated by the very institutions democracy promoters prize. Her further observation is worth keeping: this contemporary mixture tends to be more reversible than the coups it replaced.

The opposite error#

There is a temptation, having read all this, to treat polarisation as the illness and consensus as the cure. That is its own way of ending democratic politics.

Real conflicts of interest exist. Someone pays for the transition, someone is refused at the border, someone's industry closes. A demand that these be settled by reasonableness rather than contest is, functionally, a demand that the losing side stop pressing. Insistence on consensus is among the oldest instruments for neutralising opposition, which is why the agonistic tradition has spent thirty years arguing that the aim is not to abolish conflict but to give it a form in which the opponent remains an adversary rather than an enemy.

The target, then, is not less polarisation. It is polarisation within democratic limits: fight hard, and concede.

The question that actually matters#

Democracy's institutional apparatus is in reasonable repair almost everywhere. Ballots are printed. Courts sit. What the evidence assembled here describes is not the failure of machinery but the slow withdrawal of the disposition that makes machinery worth having — a disposition that turns out to be stubborn, context-bound, and not easily restored by telling people accurate things about their opponents.

So the hard question is no longer whether there will be an election. It is whether we will still want democracy on the day we lose one.