Governing Through Narrative

A text-as-data analysis of how Bukele's government constructed and coordinated its public narrative — and what independent media said instead. 162k tweets, 140k news articles, 2015–2025.

See the findings Explore all visualizations
162k
Government tweets
5
Institutional accounts
140k+
News articles
20+
News sources
2015–2025
Coverage
52%
Post-capture convergence

A country at the breaking point — and the man who said he'd fix it

Before Nayib Bukele, El Salvador was one of the most dangerous countries on earth. In 2015, the country recorded 103 homicides per 100,000 people, the highest rate in the Western Hemisphere. Gang extortion was so widespread that small businesses paid protection fees as a matter of survival. Ordinary Salvadorans avoided whole neighborhoods. The political establishment — the right-wing ARENA party and the left-wing FMLN, trading power since the 1992 peace accords — had no credible answer.

Bukele won the 2019 presidential election with 53 percent of the vote in the first round, the first candidate in Salvadoran democratic history to win without a runoff. A former publicist and mayor of San Salvador, he ran as a genuine outsider against both parties. His appeal was not ideological. It was anti-institutional: he promised to dismantle a corrupt system that had failed working-class communities for three decades.

The security improvements under his administration are real. After the Estado de Excepción was declared in March 2022, homicide rates fell from 53 per 100,000 in 2018 to approximately 2 per 100,000 in 2024. Many Salvadorans who could not previously walk in their own neighborhoods describe a transformation in daily life. Bukele's approval ratings have not meaningfully declined despite sustained international criticism — because the criticism often comes from people who were not the ones afraid.

"When human rights organizations criticize prison conditions, Bukele frames them as wealthy elites who prioritize criminals' rights over honest citizens' safety. The argument lands because it contains a grain of truth."

From the paper: Section 2.3

What distinguishes Bukele from a conventional security-focused leader is his mastery of spectacle. A former publicist, he self-labeled "the world's coolest dictator" on social media, deployed slick propaganda videos to brand CECOT — a 40,000-capacity megaprison — as a symbol of state power, and announced Bitcoin legal tender at a theatrical beachside event with pyrotechnics. Governance, for Bukele, is inseparable from performance. That is precisely the communicative mode that spin dictatorship theory predicts.

This project uses computational text analysis to ask whether that performance leaves a measurable footprint in government language — and whether the gap between what the government says and what independent media reports can be quantified at scale.

El Salvador by the numbers

103
homicides per 100k in 2015 — highest in the Western Hemisphere

53.1
homicides per 100k in 2018, year before Bukele took office

~2
homicides per 100k in 2024 (government-reported; independent analysts note potential undercounting)

80k+
people arrested under Estado de Excepción since March 2022 — largest per-capita mass incarceration in the hemisphere

85%
of the vote Bukele won in the February 2024 re-election

Three computational signatures of spin dictatorship

Drawing on 162k government tweets and 140k news articles, the analysis finds three patterns consistent with what Guriev and Treisman (2022) call spin dictatorship: coordinated messaging, agenda-setting without accountability language, and narrative construction at scale.

01

Institutional capture left a measurable linguistic footprint

After Bukele's loyalists took over the Asamblea and FGR in May 2021, those accounts began speaking like him.

On May 1, 2021, Bukele's new legislative supermajority dismissed El Salvador's Constitutional Court and Attorney General within hours of convening. We measure the linguistic impact of that capture by computing cosine similarity between each institutional Twitter account and @nayibbukele before and after the takeover. Every account converged — but the captured institutions converged most.

@AsambleaSV +52%
Before: 0.293After: 0.445
@PresidenciaSV +52%
Before: 0.404After: 0.613
@Gobierno_SV +29%
Before: 0.540After: 0.698
@FGR_SV +26%
Before: 0.253After: 0.319
Cosine similarity to @nayibbukele. Light blue = before May 2021; dark blue = after. Bars scaled to @Gobierno_SV's post-capture value (0.698 = 100%).

The two accounts whose leadership was most directly replaced show the largest jumps (both +52%). The FGR's smaller increase (+26%) reflects a natural floor: prosecutorial language is structurally constrained by legal terminology. That it converged at all is notable. Spin dictatorship theory predicts exactly this: captured bodies adopt the captor's vocabulary.

This finding quantifies something previously documented only qualitatively. Communicative convergence may be a more sensitive early-warning signal of authoritarian consolidation than formal institutional change — messaging can shift before legal structures fully align.

AccountBeforeAfterChange
@AsambleaSV0.2930.445+51.9%
@FGR_SV0.2530.319+26.1%
@PresidenciaSV0.4040.613+51.7%
@Gobierno_SV0.5400.698+29.3%

Cosine similarity to @nayibbukele. Computed over unigram frequency vectors after stop-word removal.

02

The Estado de Excepción: government engages security, avoids accountability

Government tweets discuss the crackdown — but almost never use the words "arbitrary," "innocent," or "torture."

During 2022–2024, we compare how often key terms appear in government tweets versus domestic media articles (per 1,000 units). The pattern is asymmetric: the government actively uses security vocabulary (pandillas, terroristas) but systematically avoids the language of rights and accountability.

Government tweets
Media articles
pandill-
92.0
119.4
seguridad
64.2
195.2
excepci-
26.8
118.6
derechos
14.5
158.9
arbitrar-
0.9
41.2
tortur-
0.2
19.1

Per 1,000 tweets or article-units. Bars scaled to media "seguridad" (195.2 = 100%). Red labels = accountability vocabulary.

The government is not silent on the Estado de Excepción. It actively uses security vocabulary. But the avoidance is selective: it targets precisely the sub-vocabulary that invites accountability. Government tweets use "tortur-" at 1% the rate of independent media. They use "arbitrar-" (arbitrary) at 2% the rate.

This pattern — actively engaging a topic while systematically suppressing accountability language about it — is what Guriev and Treisman (2022) describe as positive agenda-setting. The frame is constructed, not absent.

TermGov / 1kMedia / 1kRatio
pandill-92.0119.40.77×
seguridad64.2195.20.33×
excepci-26.8118.60.23×
derechos14.5158.90.09×
capturado20.760.70.34×
arbitrar-0.941.20.02×
inocente0.923.90.04×
tortur-0.219.10.01×

Gov tweets n = 34,084; media articles (domestic) n = 14,101. 2022–2024.

03

Government messaging shifted sharply at each political turning point

"Bitcoin" spiked in 2021 then disappeared. "Pandillas" tripled after the Estado de Excepción. "Derechos" quietly declined.

Term frequency per 1,000 government tweets by year reveals a government that adjusts its vocabulary in response to political events — not gradually, but in sharp discrete jumps. Bitcoin dominates 2021 messaging then nearly vanishes. Gang vocabulary triples after March 2022. Accountability terms like "derechos" (rights) and "democracia" hold flat or decline while security vocabulary rises.

Term 2019 2020 2021 2022 ↑EDE 2023 2024 2025
pandill- 9.614.012.5 44.2 76.351.986.5
excepci- 0.73.71.8 24.7 29.827.516.2
seguridad 41.938.144.9 57.0 66.975.848.5
bitcoin ⚡ 0.00.0 24.7 4.7 0.50.52.4
derechos 16.614.511.4 16.8 14.110.02.8
corrupci- 11.93.9 16.9 14.5 24.710.614.6
democracia 2.01.42.9 1.4 2.03.31.2
N tweets 13,707 17,278 13,767 16,250 10,262 7,572 2,473

Rate per 1,000 tweets across all five government accounts. Blue column = Estado de Excepción declared. Red labels = accountability terms (derechos, democracia).

Explore the full analysis

All visualizations are built from the raw corpus data. Government and media vocabulary is compared using log-odds ratios with Dirichlet smoothing (Monroe et al., 2008). Event framing windows are ±30 days around each event.

Event Framing: Government vs. Media Word Clouds

For each key event, which words appear far more in government tweets vs. independent media? Compare the CECOT opening, the Bitcoin launch, the institutional capture of May 2021, and more.

Full screen ↗

What Makes Each Account Distinctive

Log-odds ratio analysis: which words does each government account use far more than the others? @FGR_SV's prosecution templates vs. @nayibbukele's global-positioning vocabulary vs. @AsambleaSV's procedural normalization.

Full screen ↗

How Bukele Tweets About International Critics

Strategy and target heatmap. When does he delegitimize vs. deflect vs. mock? Which critics (HROs, foreign governments, press) get which treatment?

Full screen ↗

Government Vocabulary Over Time

Track how key political terms rise and fall across accounts and years. Bitcoin's 2021 spike, pandillas tripling after the Estado de Excepción, derechos quietly declining.

Full screen ↗
Note: All visualizations use the original Spanish text. To translate labels and hover text to English, right-click any visualization and choose Translate to English (Chrome / Edge), or use your browser's built-in translation feature on the full-screen view.

What Makes Each Outlet Distinctive

El Faro vs. La Página vs. El Mundo vs. Diario Colatino. Which words does each outlet use far more than the others?

Word × Year Heatmap (Press)

How does news coverage of key terms shift over time across domestic and international outlets?

Top Bigrams per Twitter Account

Recurring two-word phrases reveal structural patterns: FGR prosecution templates, Bukele's international positioning, AsambleaSV procedural boilerplate.

Top Bigrams per Media Outlet

"Régimen de excepción," "derechos humanos," "centros penales" — the phrases that define critical coverage. Compare across outlets.

Distinctive Bigrams by Source

Log-odds analysis of the most uniquely identifying two-word phrases per account and outlet. What does each source say that no one else does?

Bukele Critics: Volume Over Time

When does Bukele ramp up attacks on international critics? Volume spikes track key political stress points: the May 2021 coup, EDE criticism, IMF pushback.

Rhetorical Strategies by Target

Delegitimize, deflect, mock, or threaten? Which strategies does Bukele deploy against HROs vs. foreign governments vs. international press?

Event Framing: Butterfly Chart

Head-to-head comparison of government vs. media vocabulary for each event window, displayed as a symmetric diverging chart.

Key political events in this study

The framing analysis centers on eight moments where government and media narratives diverged most sharply. Red markers indicate democratic backsliding events.

June 2019

Bukele inaugurated

First president since the peace accords elected without ARENA or FMLN. Runs on anti-corruption, social media-first strategy. Wins 53% in the first round.

February 2020

Military enters the legislature

Bukele surrounds the Asamblea with troops to pressure passage of a security loan. First major constitutional crisis; international condemnation.

2020–2021

COVID-19 response

Daily Twitter dashboards become a central communications tool. Emergency decrees concentrate executive power. Bukele's approval holds above 80%.

May 2021

Institutional takeover

New legislative supermajority dismisses the Constitutional Court and Attorney General on Day 1. Both accounts shift messaging within weeks. IACHR condemns the move as a violation of constitutional norms.

September 2021

Bitcoin becomes legal tender

El Salvador is the first country to adopt Bitcoin nationally. Government frames it as a development breakthrough. The IMF and international investors express concern. "Bitcoin" spikes to 24.7/1k tweets — then nearly vanishes by 2023.

March 2022 – present

Estado de Excepción

Constitutional rights suspended after a gang violence spike. 80,000+ arrests — the largest per-capita mass incarceration in the Western Hemisphere. Government: historic security success. Independent media and Amnesty: arbitrary detentions, deaths in custody, systematic torture.

2023

CECOT opens

The Terrorism Confinement Center (40,000 capacity) is launched with a government propaganda video set to driving music. Bukele tours international leaders through it. Human rights organizations document beatings and overcrowding.

February 2024

Bukele re-elected — 85%

Wins by a historic margin. The Constitutional Court that permitted his candidacy for a second term was composed of his own 2021 appointees. International observers note unfair advantages despite calling the vote technically free.

A multi-source Salvadoran political text corpus

Assembled from six source streams: government Twitter, GDELT BigQuery, HuggingFace datasets, direct outlet crawls, URL repair via the Wayback Machine, and internationally-filtered press. Cleaned, deduplicated, and filtered for political relevance to El Salvador.

Government Twitter accounts

AccountRoleTweets
@AsambleaSVLegislature54,466
@FGR_SVPublic prosecutor36,519
@Gobierno_SVExecutive comms33,370
@PresidenciaSVExecutive comms30,438
@nayibbukelePresident personal6,995

Collected via Twitter/X API v2 full-archive search. Budget constrained to $30 total; full timeline pagination per account. Corpus spans 2015–2026.

News article corpus (139,637 articles)

Source streamArticles
HuggingFace (El Mundo, EDH)85,912
International press (Reuters, DW, Guardian…)30,849
Original scrape (sitemaps + GDELT)8,815
Repaired URLs (redirects + Wayback)8,157
New outlets (Factum, Foco STV…)5,904

Collection methodology

  • Twitter/X API v2 full-archive search with pagination
  • GDELT BigQuery export — 1.86M event rows filtered to relevant Salvadoran URLs
  • HuggingFace datasets (justinian336/salvadoran-news-*)
  • Direct sitemap and archive crawls for outlet-level collection
  • URL repair: 11,572 broken links recovered via HTTP 301 redirect-following and Wayback Machine CDX API
  • International content filter: requires 3+ mentions of "El Salvador" or 2+ "Bukele" with key terms

Limitations

  • Twitter API budget capped at $30 — @nayibbukele coverage is truncated relative to full history
  • La Prensa Gráfica is partially paywalled
  • Gatoencerrado.news blocked by Cloudflare
  • Corpus captures text production, not consumption — no audience data
  • Cosine similarity measures lexical overlap, not semantic content
  • Before/after comparison is not a clean natural experiment

How the analysis works

Text analysis methods

  • Log-odds ratio with 0.5 Dirichlet smoothing — identifies vocabulary that is distinctively more common in one source than all others (Monroe, Colaresi, and Quinn, 2008)
  • Cosine similarity over unigram TF vectors — measures overall vocabulary overlap between two accounts or documents
  • Event framing windows — ±30 days around key events; TF-IDF weighted word clouds; separate vectors for government tweets and media articles
  • Bigram analysis — adjacent non-stop token pairs reveal phrase-level patterns (FGR prosecution templates, Bukele's international positioning)
  • Temporal heatmaps — word × year normalized frequency, showing how vocabulary priorities shift over time

Built with Python, Plotly, WordCloud, and pandas. All visualizations are self-contained interactive HTML. Spanish tokenization strips URLs, @mentions, hashtags, and applies a 3-character minimum token length with a comprehensive stop-word list.

Theoretical framework

This project tests predictions from Guriev and Treisman's (2022) theory of "spin dictatorship": the idea that contemporary autocrats govern primarily through narrative control rather than violence. Spin dictators are predicted to show coordinated messaging across institutions, positive agenda-setting that avoids accountability language, and social media attacks on international critics rather than physical repression.

All three predictions find support in the Salvadoran corpus.

Access the paper

Download paper (.pdf)

Course project for Text as Data, Spring 2026.