MyDataLabs
MyDataLabs Quantifying the Unquantifiable
Lab Note LN-02 · Politics

Is the World Polarising? 25 Years, 173 Countries

Yes: hostility between political camps has risen in every region since 2000, mostly after 2010. Of ten candidate explanations, parties' hate speech is the one that holds up; people going online is not.

12 min read
World average, 2000 → 2025
+0.69
-0.41 to +0.28 on a −4…+4 scale
Clearly rose / clearly fell
62 / 9
of 173 countries, beyond measurement noise
Countries with hostile camps
40% → 54%
deeply hostile (≥ +1): 13% → 32%
Of the rise after 2010
85%
all 6 world regions rose

In short Yes, polarisation is rising across the world. In 173 countries tracked by V-Dem since 2000, the average moved from -0.41 to +0.28, crossing from friendly to hostile territory. 62 countries became clearly more polarised and 9 clearly less; every region rose, and 85% of the rise came after 2010. As for why, of ten candidate explanations one passes every test we ran: parties’ use of hate speech against religious, ethnic and other groups. Its rise alone accounts for about a third of the global increase. Disinformation and weakening democratic institutions move with polarisation but do not hold up as independently. And the most popular explanation, people simply being online, shows no link at all. Polarisation also feeds these factors back: it is a loop, not a one-way chain.

Part 1: Is polarisation rising?

V-Dem asks country experts every year whether society is “polarized into antagonistic, political camps” — whether supporters of opposing sides avoid friendly contact at family gatherings, at work and in their free time (v2cacamps). Roughly, below zero people across the divide mostly get along; above zero they are more likely to be hostile than friendly. We use every country V-Dem covers in all 26 years from 2000 to 2025: 173 countries, democracies and autocracies alike.

The world average was flat until about 2005, edged up to -0.30 by 2010, then rose steadily to +0.28. The median country tells the same story (-0.37 to +0.33). Every country counts once: we do not weight by population, which would let a handful of very large countries decide the answer for the world.

World polarisation, 2000–2025 173 countries. Shaded: the middle half of countries (25th to 75th percentile). Above zero, camps are more hostile than friendly.

An average can rise because a few countries move a lot. That is not what happened. The whole distribution shifted: the share of countries where camps are more hostile than friendly went from 40% to 54%, and the share scoring +1 or higher, where hostility is the norm, more than doubled, from 13% to 32%.

Share of countries at each level of polarisation Percent of the 173 countries in each band, by year. Click a band in the legend to hide it; the axis rescales to the lines still showing.

Everywhere, but not evenly

All six regions ended higher than they started. Latin America and Eastern Europe rose most; Western Europe and North America rose almost a full point but remain the least polarised region; Sub-Saharan Africa, which started highest, barely moved.

Average polarisation by region V-Dem’s politico-geographic regions; number of countries in brackets. Click a region in the legend to hide it; the axis rescales to the regions still showing.

How sure can we be about each country?

Each score is an estimate from a handful of experts (5 per country in 2025, typically), and V-Dem publishes an uncertainty range with every one. Counting only changes where the 2025 range clears the 2000 range, 62 countries became more polarised and 9 less. The other 102 moved too little to tell. That still makes risers outnumber fallers by about seven to one.

Change in polarisation per country, 2000–2025 Number of countries by size of change. Dark: change larger than V-Dem’s measurement uncertainty. Light: within it.
Every country with a clear change, 2000 vs 2025 Each bubble is one country whose change exceeds V-Dem’s measurement uncertainty (62 rose, 9 fell). Above the diagonal line, polarisation is higher in 2025 than in 2000; bubble size is the size of the change. Hover for details.
Show the full country tables
Every clear rise (62)
Country20002025Change
Nicaragua-0.47+3.14+3.62
Honduras-2.58+0.84+3.42
Brazil-0.17+3.22+3.40
Türkiye+0.04+3.36+3.32
Poland+0.33+3.35+3.02
Ukraine-1.04+1.80+2.83
Hungary-0.29+2.47+2.76
United States of America-0.46+2.30+2.76
Germany-2.44+0.19+2.63
Romania-0.05+2.55+2.61
Bulgaria-1.24+1.29+2.53
Yemen+0.35+2.81+2.46
Mexico-0.29+2.13+2.42
Slovakia-0.95+1.42+2.36
Maldives-1.70+0.66+2.36
Mozambique-1.06+1.26+2.32
India-0.22+2.08+2.30
Hong Kong-1.23+1.00+2.23
Netherlands-2.31-0.09+2.21
Burkina Faso-0.67+1.53+2.20
Colombia-0.25+1.89+2.14
Pakistan-1.11+0.99+2.10
Libya+0.47+2.48+2.01
Czechia-2.12-0.13+1.99
Belarus+1.19+3.18+1.99
Argentina-0.03+1.90+1.93
Azerbaijan+0.26+2.18+1.92
Bolivia-0.28+1.62+1.90
Venezuela+0.38+2.23+1.85
Lithuania-2.67-0.87+1.80
Thailand-0.20+1.56+1.76
Slovenia+0.28+2.04+1.76
Costa Rica-2.02-0.27+1.75
Moldova-0.15+1.60+1.74
Portugal-2.37-0.64+1.73
Guatemala-1.00+0.68+1.68
Spain-1.75-0.09+1.66
Philippines-1.14+0.51+1.65
Guinea-Bissau-0.76+0.86+1.62
France-0.18+1.43+1.61
Sweden-2.28-0.67+1.61
Morocco-0.67+0.88+1.56
Bangladesh+0.86+2.37+1.50
Egypt-1.70-0.22+1.48
Bhutan-2.38-0.89+1.48
South Korea-0.04+1.44+1.48
Tanzania-1.54-0.10+1.44
Cameroon+0.67+2.08+1.42
Madagascar+1.08+2.38+1.29
Palestine/West Bank-1.23+0.04+1.27
Uruguay-3.06-1.81+1.25
Canada-1.62-0.40+1.22
Malawi-0.88+0.33+1.21
Papua New Guinea-2.08-0.88+1.20
New Zealand-2.61-1.42+1.19
Cambodia-0.17+0.99+1.16
Russia+0.13+1.24+1.11
Uzbekistan-2.53-1.43+1.10
Suriname-1.81-0.73+1.08
Senegal-1.23-0.14+1.08
Benin-1.31-0.27+1.04
Georgia+0.83+1.75+0.93
Every clear fall (9)
Country20002025Change
Iraq+2.66+0.20-2.46
Democratic Republic of the Congo+1.27-0.76-2.02
Togo+2.43+0.53-1.90
Angola+1.58-0.30-1.89
Zanzibar+0.22-1.35-1.56
Nepal+0.99-0.45-1.44
Somalia+2.83+1.45-1.39
Zimbabwe+2.43+1.24-1.19
Dominican Republic+0.02-0.90-0.92

The risers span every kind of country: established democracies (the United States, Germany), new ones (Poland, Romania, Brazil) and countries that became less democratic along the way (Nicaragua, Türkiye, Hungary). The fallers are mostly countries where hostility was extreme in 2000, several of them coming out of war (Iraq, Angola, the DR Congo, Nepal).

Part 2: What moves with it?

V-Dem rates many other things about each country every year. We picked ten that are commonly offered as explanations for polarisation and that V-Dem measures from 2000 onward, in three groups:

  • Technology: how much citizens consume online media; how much parties campaign on social media; how differently online outlets report the same news.
  • What political actors do: party disinformation online; government disinformation; foreign disinformation; parties’ hate speech against groups defined by race, religion, origin or similar.
  • Media and institutions: media bias against the opposition; government censorship of the media; the liberal democracy index.

For each, we ask the same narrow question: in the years when a country scores higher on this than it usually does, is it also more polarised than usual? Comparing a country with itself removes everything stable about it — history, culture, electoral system — and comparing within the same year removes shocks that hit everyone at once. The model covers 4,629 country-years in 179 countries. Coefficients are standardised so they can be compared: each is the change in polarisation, in standard deviations, for a one-standard-deviation rise in the candidate.

How strongly each candidate moves with polarisation Standardised coefficients with 95% intervals, errors clustered by country. Grey: each candidate on its own. Blue: all ten together, so each is measured holding the others fixed.

Taken one at a time, almost everything correlates with polarisation: disinformation, censorship, media bias, hate speech and democratic decline all come with more hostile camps. That is not very informative, because these things tend to happen together. Put all ten in at once and only one stays clearly significant: party hate speech (+0.37, p <0.001). Liberal democracy is borderline (-0.19, p = 0.055). The disinformation measures shrink by three-quarters or more.

Share of the global rise accounted for, by candidate Points of the 0.73-point rise in average polarisation, 2000–2025. Solid: significant in the joint model. Pale: not significant.

Part 3: What can be done?

Some of the ten candidates are conditions: how many people are online, how fragmented the news is, how democratic a country is. Others are choices made by people who can be named. Party hate speech, party disinformation, government disinformation, foreign disinformation and parties’ social media campaigning are all things a party, a government or a foreign state decides to do, and almost all of them reach voters through platforms that someone runs. These five can be controlled.

Together they account for 69% of the 0.73-point rise in average polarisation by a Shapley split (95% interval 52–91%), and 71% by the simpler bookkeeping. The two methods disagree on how to divide credit among these five, because they rose together, but not on their total. They also account for 74% of the year-to-year variation the ten candidates explain. Leaving out social media campaigning, which is only weakly linked to polarisation once the others are controlled for, hate speech and the three disinformation measures alone account for 50% of the rise (interval 35–65%) and 71% of the explained variation.

How much of the rise is in actors’ hands Share of the 0.73-point rise in average polarisation, 2000–2025, under each way of splitting it. Red and amber: controllable. Grey: the other five candidates and the part no candidate explains.

Who holds the levers

  • Parties and their leaders decide whether to use hate speech against groups and whether to spread false claims. Electoral codes of conduct, enforced by election commissions, and voters who punish such campaigns are the direct checks.
  • Governments decide whether to spread disinformation themselves, and set the rules for parties, broadcasters and platforms.
  • Digital platform owners decide what their systems amplify. Their ranking systems reward engagement, and content that attacks the other side engages most. Each word about the political out-group raised the odds of a post being shared by 67% in a study of 2.7 million Facebook and Twitter posts (Rathje, Van Bavel and van der Linden, 2021). Each moral-emotional word raised a tweet’s spread by about 20% (Brady and colleagues, 2017). False news was 70% more likely to be retweeted than true news (Vosoughi, Roy and Aral, 2018). When platforms amplify what spreads fastest, they amplify the hate and falsehood that this panel links to polarisation, so they share responsibility for it.

Two cautions. First, the arrow runs both ways: polarisation also produces more hate speech and disinformation, so cutting them would not remove the full 69%. The figure is the share of the rise that travels with these choices, more likely a ceiling on what changing them could achieve than a promise. Second, V-Dem does not measure platform algorithms; the virality figures come from the studies cited, mostly of English-language posts in the United States, not from this panel.

What these numbers cannot show

  • Experts rate both sides. Polarisation, hate speech and disinformation are all judged by country experts, typically 5 per country. An expert who sees a country as bitterly divided may also rate its parties as more hateful. Some of the link may be built into the measurement. Surveys of ordinary voters are the natural check.
  • The measures are coarse. About 61% of year-to-year polarisation scores do not change at all, and the uncertainty ranges average 1.12 points wide. The analysis has less information than the country-year count suggests.
  • Correlation, not cause. Comparing each country with itself removes a great deal, but not events that push polarisation and its candidates up together, such as a bitter election.
  • Not every explanation is measurable here. Inequality, immigration, ageing and the decline of traditional news are all plausible and are not in V-Dem’s yearly data, so they could not be tested.

What to take from this

Polarisation is rising across the world: in most countries, in every region, mostly since 2010, and in democracies more than in autocracies (an average rise of +1.08 against +0.34). The factor that tracks it most consistently is not technology but political choices — above all, parties using hate speech against groups — with economic downturns adding to it and disinformation travelling alongside. People going online has, by this measure, nothing to do with it. And polarisation feeds the very things that feed it. That is also the hopeful part: about 69% of the rise travels with choices that parties, governments and platform owners make, and can make differently.

Discussion & Methodology Notes

Reader Observations

Methodological feedback, questions, and empirical observations on this study. To maintain high-signal discourse and prevent promotional noise, all submissions are moderated before publication.

No observations published yet. Be the first to share an observation.

Contribute an Observation

Links, handles, and promotional copy are prohibited. Describe references in words.

0 / 1200
Minimum 15 characters · Reviewed by the editorial desk
Methodology · LN-02

How the panel was built and tested

V-Dem v16 country-year data, every country, 2000–2025.

Source and sample

Varieties of Democracy, dataset v16 (March 2026), country-year file. No country is filtered out by regime type: the question is about the world, so autocracies count. The trend charts use the 173 countries with a polarisation score in every year from 2000 to 2025, so that countries entering or leaving the data cannot move the average. The regressions use every country-year with all ten candidates observed (4,629 country-years, 179 countries). All averages are unweighted: each country counts once, whatever its population. Regions are V-Dem’s six politico-geographic regions (e_regionpol_6C).

V-Dem advises against relying on scores built from three or fewer expert coders. In the regression sample, 3.7% of country-years have at least one of the ten scores with coder counts used here in that position. They are kept in the main results for coverage. Without them (4,459 country-years left), the world average on the balanced panel moves from -0.40 to +0.30, and party hate speech keeps its joint coefficient (+0.37, p <0.001). The one change is liberal democracy, borderline in the main model (p = 0.055), which becomes significant (-0.22, p = 0.029). No other candidate changes its verdict.

Variables

LabelV-Dem variableQuestion
Polarisationv2cacampsIs society polarized into antagonistic, political camps?
Citizens' online media usev2smonexDo people consume domestic online media?
Parties' social media campaigningv2smcampDo parties and candidates use social media to campaign?
Online media fragmentationv2smmefraDo the major online outlets give different presentations of the news?
Party disinformation onlinev2smpardomHow often do parties use social media to spread false information domestically?
Government disinformationv2smgovdomHow often does the government spread false information domestically?
Foreign disinformationv2smfordomHow often do foreign governments spread false information in the country?
Party hate speechv2smpolhateHow often do major parties use hate speech against groups defined by race, religion, origin or similar?
Media biasv2mebiasIs coverage of the opposition biased?
Government media censorshipv2mecenefmDoes the government attempt to censor the print and broadcast media?
Liberal democracyv2x_libdemV-Dem's liberal democracy index (0 to 1).

V-Dem codes disinformation, hate speech, fragmentation, media bias and censorship so that high means less; we multiply them by −1 so that every candidate reads “higher means more”. “Beyond measurement noise” means V-Dem’s 68% intervals (v2cacamps_codelow, _codehigh) for 2000 and 2025 do not overlap. Economic growth is the change in log GDP per capita (e_gdppc), available to 2019.

Models

All models are OLS with country and year fixed effects and standard errors clustered by country. Coefficients on the Analysis tab are standardised (multiplied by the candidate’s standard deviation over that of polarisation, both pooled). The joint model explains R² = 0.894, against 0.855 for country and year effects alone. First differences regress the yearly change in polarisation on the yearly changes in all ten candidates with year effects. The lagged model regresses polarisation on its own lag and the lags of all ten; the reverse model regresses each candidate on its own lag and lagged polarisation. Lagged-dependent-variable models over a short panel carry a known bias (Nickell bias) and are offered as a check on timing, not as effect sizes. The accounting multiplies each joint-model coefficient by the change in the candidate’s average across the regression sample from 2000 to 2025.

Full results

Candidate Alone (std) Joint (std) First diff. Lagged Reverse 2000–10 2011–25
Citizens' online media use +0.08 (0.186) +0.02 (0.700) -0.00 (0.911) +0.03 (0.063) -0.003 (0.743) +0.03 (0.680) +0.10 (0.208)
Parties' social media campaigning +0.13 (0.006) +0.05 (0.163) +0.05 (0.027) -0.00 (0.806) +0.004 (0.796) +0.16 (0.013) +0.09 (0.112)
Online media fragmentation +0.14 (0.108) +0.04 (0.492) +0.05 (0.179) +0.01 (0.460) -0.007 (0.487) +0.30 (<0.001) +0.09 (0.335)
Party disinformation online +0.38 (<0.001) +0.09 (0.154) +0.03 (0.341) +0.04 (0.040) +0.035 (0.009) +0.23 (0.090) +0.43 (<0.001)
Government disinformation +0.40 (<0.001) +0.07 (0.181) +0.08 (0.008) +0.00 (0.919) +0.037 (0.001) +0.20 (0.098) +0.40 (<0.001)
Foreign disinformation +0.29 (<0.001) +0.05 (0.295) +0.05 (0.129) +0.01 (0.692) +0.021 (0.029) +0.31 (0.029) +0.24 (<0.001)
Party hate speech +0.58 (<0.001) +0.37 (<0.001) +0.19 (<0.001) +0.05 (0.049) +0.028 (0.018) +0.56 (<0.001) +0.46 (<0.001)
Media bias +0.26 (0.002) +0.02 (0.803) +0.05 (0.064) +0.00 (0.951) +0.042 (0.005) +0.08 (0.331) +0.34 (<0.001)
Government media censorship +0.30 (<0.001) +0.06 (0.247) +0.04 (0.077) -0.00 (0.812) +0.054 (<0.001) +0.12 (0.212) +0.35 (<0.001)
Liberal democracy -0.54 (<0.001) -0.19 (0.055) -0.95 (0.002) +0.06 (0.691) -0.004 (<0.001) -0.31 (0.023) -0.50 (<0.001)

p-values in brackets. “Alone”, “Joint” and the two period columns are standardised; first differences and lags are in raw units. “Reverse” is the effect of last year’s polarisation on the candidate this year (for liberal democracy, a negative value means polarisation is followed by decline).

Checking the accounting with a Shapley split

The bookkeeping on the Analysis tab gives each candidate whatever the joint model assigns it. When candidates rise together, as hate speech, disinformation and social media campaigning did, that split depends on which variable wins the shared trend in the regression. A Shapley decomposition removes that dependence: the model is refitted on all 1024 subsets of the ten candidates (country and year effects kept), and each candidate is credited with its average addition to the predicted rise across every order in which it could enter. Both methods split the same total, 0.58 of the 0.73-point rise, but they share it out differently.

Candidate Bookkeeping Shapley Gap Shapley share of rise (95% interval) Share of explained variation
Citizens' online media use +0.022 +0.045 +0.023 6% (-8 to 18) <1%
Parties' social media campaigning +0.106 +0.141 +0.035 19% (6 to 39) 3%
Online media fragmentation +0.001 -0.005 -0.006 -1% (-4 to 2) 2%
Party disinformation online +0.090 +0.105 +0.015 14% (10 to 23) 14%
Government disinformation +0.043 +0.072 +0.029 10% (5 to 16) 14%
Foreign disinformation +0.043 +0.075 +0.032 10% (4 to 17) 6%
Party hate speech +0.238 +0.111 -0.127* 15% (8 to 23) 37%
Media bias +0.003 +0.015 +0.012 2% (0 to 7) 5%
Government media censorship +0.018 +0.027 +0.009 4% (1 to 10) 7%
Liberal democracy +0.014 -0.009 -0.023 -1% (-6 to 3) 12%

Points of the 0.73-point rise in the regression sample. Intervals from 200 bootstrap resamples of whole countries. * The bootstrap interval for the gap excludes zero. The last column splits the within-country R² of the joint model (0.27) the same way.

The two methods disagree on party hate speech, and only there is the gap clearly larger than noise. Bookkeeping credits it with 0.24 points (33% of the rise); Shapley with 0.11 (15%, interval 8–23%). The credit it loses goes mostly to social media campaigning and the three disinformation measures, which rose over the same years. So the share of the global rise that can be put on hate speech lies somewhere between about 15% and 33%, depending on how the shared trend is divided. Its lead is clearer in the year-to-year movement: it accounts for 37% of the within-country variation the ten candidates explain, more than twice any other. Social media campaigning gets a large share of the rise only because it climbed so far; it explains 3% of the variation and is only weakly linked to polarisation once the others are controlled for.

Two samples are used on this page. The trend charts use the 173 countries observed in every year from 2000 to 2025, where the world average rose by 0.69. The regressions, the bookkeeping and the Shapley split use every country-year with all ten candidates (179 countries), where the average rose by 0.73. Run on the balanced countries only, the Shapley split gives the same picture.

What changed from the first draft

The first draft studied twenty large “democracies”. That sample turned out to include three countries V-Dem now classes as electoral autocracies, and its results changed depending on how democracy was used to pick countries. Since the question is whether the world is polarising, this version uses every country instead. The draft also overstated several points, corrected here: it used conventional rather than clustered standard errors; described hate speech as attacks on political opponents, when V-Dem’s item measures hate speech against social groups; said no variable captured citizens’ internet use, when v2smonex does; and attributed Australia’s and Japan’s stability to institutions that were never measured. Unsourced figures (“3× engagement”, “70%+ of partisans”) were removed.

Reproduction

Every number on this page is written by scripts/build_polarisation_story.py from the V-Dem file into a single data file the page reads, so no figure is typed by hand.

Sources and licence

  • Coppedge, Michael, John Gerring, Carl Henrik Knutsen, Staffan I. Lindberg, Jan Teorell, David Altman, Fabio Angiolillo, Michael Bernhard, Agnes Cornell, M. Steven Fish, Linnea Fox, Lisa Gastaldi, Haakon Gjerløw, Adam Glynn, Ana Good God, Allen Hicken, Katrin Kinzelbach, Joshua Krusell, Kyle L. Marquardt, Kelly McMann, Valeriya Mechkova, Juraj Medzihorsky, Anja Neundorf, Pamela Paxton, Daniel Pemstein, Josefine Pernes, Johannes von Römer, Brigitte Seim, Rachel Sigman, Svend-Erik Skaaning, Jeffrey Staton, Aksel Sundström, Marcus Tannenberg, Eitan Tzelgov, Yi-ting Wang, Tore Wig, Steven Wilson and Daniel Ziblatt. 2026. “V-Dem [Country-Year/Country-Date] Dataset v16”. Varieties of Democracy (V-Dem) Project. https://doi.org/10.23696/vdemds26.
  • Pemstein, Daniel, Kyle L. Marquardt, Eitan Tzelgov, Yi-ting Wang, Juraj Medzihorsky, Joshua Krusell, Farhad Miri and Johannes von Römer. 2026. “The V-Dem Measurement Model: Latent Variable Analysis for Cross-National and Cross-Temporal Expert-Coded Data”. V-Dem Working Paper No. 21, 11th edition. University of Gothenburg: Varieties of Democracy Institute.
  • Coppedge, Michael, et al. 2026. “V-Dem Codebook v16”. Varieties of Democracy (V-Dem) Project.
  • For the social media and disinformation items (v2sm*): Mechkova, Valeriya, Daniel Pemstein, Brigitte Seim and Steven Wilson. 2019. “Measuring Internet Politics: Introducing the Digital Society Project”. Digital Society Project Working Paper 2019(1).

V-Dem publishes its dataset under CC BY-SA 4.0, which requires attribution and that derivative products carry the same terms. The figures and the data file behind this note are derived from it and are shared on those terms. V-Dem and its institutions are not responsible for this analysis or its conclusions.