“We do not want a personal and abusive campaign. Let’s talk about what we can do for the country.”

“The Tories are the party of wealth and privilege. Labour is the party of hard work, fairly paid.”

The above statements are from two leaders of the same party, the British Labour Party – Tony Blair and Ed Miliband – both running for the office of Prime Minister some 20 years apart. The rhetoric of political leaders has, in recent years, become more an exercise in division than in persuasion. Is this because voters’ preferences have evolved toward greater polarisation, or are political leaders themselves actively fuelling political polarisation? And if so, why?

Standard models of electoral competition, rooted in the tradition of Anthony Downs, suggest that candidates competing for the same pool of voters should converge to the same moderate positions. Divergence, in this view, reflects fixed ideological commitments – politicians accepting some electoral cost to stay true to their beliefs (Alesina 1988). A different view of elections says that competing politicians can diverge and polarise the electorate, because they target different groups of voters and seek to energise their core supporters rather than fighting for the same swing voters (Glaeser et al. 2005, Gennaioli and Tabellini 2025). In this second case, electoral campaigns give politicians strong incentives to sharpen their distinctive messages and to polarise the electorate.

Testing these competing theories is difficult. Speeches, manifestos, and party platforms are infrequent and are less followed than social media by people interested in politics. In a new paper (Boeri et al. 2026), we exploit a rich source of high-frequency data that cuts through this problem: the Twitter/X posts of 367 political leaders across 21 Western democracies, spanning 2013 to 2022. With roughly 3.4 million tweets, grouped at quarterly intervals around staggered national election dates, we can track how the rhetoric of politicians evolves as elections approach and recede. A distinctive advantage of our approach is that it focuses on political leaders, as opposed to parties like most of the literature on populism to date (Guriev and Papaioannou 2022). Rhetoric is a personal tool of an individual leader rather than the official platform of an organisation.

Measuring rhetorical polarisation

We focus primarily on the populist versus non-populist divide, which is now the defining axis of political competition in Western democracies. Populist leaders in the sample range from Donald Trump and Jair Bolsonaro to Marine Le Pen, Giorgia Meloni, and Bernie Sanders. Non-populists include mainstream centre-left and centre-right leaders across all 21 countries. There is not a one-to-one correspondence between populist parties and populist leaders; we allow for non-populist leaders in populist parties and populist leaders in non-populist parties. Our classification draws on Funke et al. (2023), supplemented by GPT-assisted coding with human validation.

Our central challenge is measurement. Any two groups of people will use somewhat different words by chance alone; the question is whether observed differences are larger than random variation would predict. We address this using the methodology of Gentzkow et al. (2019). For each politician in each quarter, we compute a “partisanship” score: the probability that an impartial observer, seeing a single two-word phrase (a bigram) from that politician’s tweets, could correctly identify whether he/she is populist or non-populist. A score of 0.5 means the language gives no information; higher values mean the language is increasingly distinctive.

The electoral cycle in political rhetoric

Our key finding is that political rhetoric becomes significantly more polarised in the run-up to national elections and reverts toward its average level once elections are over.

Populist versus non-populist partisanship starts to rise two quarters before the election and peaks in the election quarter itself. After the election, it declines back to baseline. This pattern holds after controlling for observable characteristics of politicians, for politician and calendar-quarter fixed effects, and for country-specific electoral-window effects. Thus, it captures changes in individual rhetoric, removing the effect of unobserved and time invariant features in the rhetorical style of each individual politician, as well as of unobserved aggregate shock, including of polarising events like the 2015 refugee crisis or the COVID-19 pandemic that may happen to coincide with elections.

To appreciate the magnitude, consider what a tweet can tell you. Two years before an election, on average reading a single tweet raises the probability of correctly identifying a politician as populist or non-populist from 50% to about 57%. During the election quarter, that same single tweet raises predictability to 62% – nearly doubling the informational content. After reading two tweets in the election quarter, predictability climbs to 67%. These are substantial differences for a measure that aggregates across thousands of bigrams and hundreds of politicians in many languages and countries.

The magnitude of the partisanship values in the electoral quarter is comparable to that observed during the 2015 Syrian refugee crisis in Europe, one of the most politically salient shocks of the decade. Elections, in other words, move rhetorical polarisation by roughly as much as a major external crisis does.

Why polarisation, not convergence?

The analysis is also revealing of some of the mechanisms behind this electoral surge in rhetorical polarisation.

First, the polarising effect is stronger in plurality-rule and presidential systems, where competition typically involves two dominant parties or candidates, than in proportional representation systems with many parties. This rules out the simple explanation that divergence reflects multi-candidate dynamics. Two-party competition should, if anything, push towards convergence under standard Downsian logic. Instead, the data suggest that Twitter/X allows politicians to reach different audiences, giving each party an incentive to sharpen its message for its own supporters rather than chase the median voter.

Second, the increased polarisation near elections operates through two distinct channels. Populists and non-populists become more likely to talk about different topics (between-topic divergence) and to frame the same topics differently (within-topic divergence). Non-populist topics – predominantly policy-oriented ones such as climate change, foreign policy, and inflation – become even more characteristic of non-populist politicians as elections approach. Populist topics – immigration, anti-establishment rhetoric, self-promotion – remain distinctively populist throughout.

Third, both groups of politicians adopt a more populist style as elections near. Specifically, they use fewer phrases associated with formal, institutional, elite-coded language – phrases like “member state,” “prime minister,” or “head of state” – and shift toward more direct, emotionally resonant communication. Non-populist politicians in particular appear to partially imitate populist rhetorical style, even as they diverge in substantive content. This is consistent with research by de Vries and Hobolt (2020) on how mainstream parties respond to populist challengers, and with broader arguments about the adoption of populist communication strategies across the political spectrum.

Twitter as a targeting tool

Why do we find polarisation where some other studies – notably Di Tella et al. (2025), examining candidate websites and party manifestos in US and French elections – find convergence? The answer likely lies in the nature of social media. Tweets can be targeted, monitored, and adjusted in real time. Until 2023, Twitter analytics provided breakdowns of followers by political interest, location, and other characteristics. Politicians could assess whether their messages were reaching their intended audiences. Allcott et al. (2025) document that, in the 2020 US presidential election, most Facebook and Instagram political advertising was directed at each party’s own supporters rather than at persuadable voters. Our results suggest Twitter/X serves a similar function for political communication.

Manifestos and candidate websites, by contrast, are official documents that face scrutiny from a broad audience, including political opponents and mainstream media. The incentives for convergence – or at least moderation – are stronger there. Social media, by lowering the cost of selective communication, may therefore be reshaping the incentive structure of electoral competition in ways that systematically amplify polarization.

This is not merely an academic concern. A growing literature documents that politicians’ messages and propaganda fuel polarisation among citizens (Zhang et al. 2025). If elections give politicians stronger incentives to polarise their rhetoric precisely when public attention is highest, the feedback loop between elite polarisation and mass polarisation may be tighter than previously recognised.

Implications

Our findings carry some broad implications. Platform design matters for democratic competition – and the mechanisms operate also on the supply side of politics, not only on the demand side. The targeting technologies embedded in social media do not merely expose voters to more extreme content; they also alter the incentives of the politicians who produce that content. By enabling selective communication with core supporters at low cost, social media makes base mobilization more attractive relative to median-voter convergence.

These supply-side incentives interact with demand-side effects to compound polarisation. Gennaioli et al. (2026) document that merely making a social conflict salient – without providing any new information – sharply widens disagreement on both factual and normative questions. Politicians need not spread misinformation to polarise voters; they can simply choose which conflict to put at the top of the public agenda.

It is not a mere coincidence that the rise in political polarisation over the past two decades coincided with the diffusion of social media (Melnikov 2025).  The new communication technologies have changed the incentives of politicians, making base mobilisation more attractive relative to centrist persuasion, while at the same time increasing the effectiveness and diffusion of emotional and extreme content. Debates about transparency in political advertising and about the algorithmic amplification of extreme content need to grapple with this supply-side dynamic, alongside the more studied demand-side effects.

References

Alesina, A (1988), “Credibility and policy convergence in a two-party system with rational voters”, American Economic Review 78(4): 796–805.

Allcott, H, M Gentzkow, R Levy et al. (2025), “The Effects of Political Advertising on Facebook and Instagram before the 2020 US Election”, NBER Working Paper 33818.

Boeri, T, N Nikiforova, and G Tabellini (2026), “Do Elections Moderate or Polarize Political Rhetoric?”, CEPR Discussion Paper 21317.

de Vries, C E and S Hobolt (2020), Political Entrepreneurs: The Rise of Challenger Parties in Europe, Princeton University Press.

Di Tella, R, R Kotti, C Le Pennec, and V Pons (2025), “Keep your enemies closer: strategic platform adjustments during US and French elections”, American Economic Review 115(8): 2488–2528.

Funke, M, M Schularick, and C Trebesch (2023), “Populist leaders and the economy”, American Economic Review 113(12): 3249–3288.

Gennaioli, N and G Tabellini (2025), “Identity Politics”, Econometrica 93(6): 1937–67.

Gennaioli, N, F Schwerter, and G. Tabellini (2026), ““Us versus Them”: Salient Conflict and Belief Polarization”, CEPR Discussion Paper 21324.

Gentzkow, M, J M Shapiro, and M Taddy (2019), “Measuring group differences in high-dimensional choices: method and application to congressional speech”, Econometrica 87(4): 1307–1340.

Glaeser, E L, G Ponzetto, and J Shapiro (2005), “Strategic Extremism: Why Republicans and Democrats Divide on Religious Values”, Quarterly Journal of Economics 120(4): 1283–1330.

Guriev, S and E Papaioannou (2022), “The Political Economy of Populism”, Journal of Economic Literature 60(3): 753–832.

Melnikov, N (2025), “Mobile internet and political polarization”, SSRN Working Paper 3937760.

Zhang, L, G Gratton, P Grosjean, and H Yousaf (2025), “Adding Fuel to the (Gun) Fire: How Politicians Polarize the Public Debate”, Working Paper.