Measuring the erosion of democratic discourse: what 872 million news events reveal

How the conflict framing elasticity (β) of news media tracks democratic backsliding across 40 countries between 2017 and 2026.

Hannah Arendt wrote that democracy fails when a society loses its shared ground for facts. The mechanism is not censorship in the obvious sense but the slow corrosion of common reality: events still happen, but their framing fractures. If she is right, then media coverage should carry a measurable signal of democratic health, and that signal should move first.

For the past two years I have been working on extracting that signal from the GDELT database, which records every news event indexed by the major English-language news aggregators since 1979. The result is a single number per country per year that I call β, the conflict framing elasticity: how reactively the tone of news coverage responds to the conflictual intensity of events. A country with a high β publishes more negative stories when violent or contentious events occur; a country with a low β maintains a steadier tone regardless of what happens.

What the data shows

Three findings together form what I think is a useful early-warning signal:

First, β is rising globally. The median β across forty countries grew by about +0.0032 per year between 2017 and 2026 (p = .015). Modest in size, but consistent: the world's news media is becoming more reactive to conflictual events, not less.

Second, the rise is concentrated in hybrid regimes and flawed democracies. These are the countries Samuel Huntington described as the "contested middle" of the third wave of democratization. Full democracies are stable; authoritarian regimes are stable at a much more negative floor; the middle is where the motion happens. Eight countries cross the threshold for statistically significant β acceleration: Honduras, Uganda, Malaysia, Tanzania, Bolivia, Turkey, Mexico, and Portugal.

Third, the discourse field has structure. When we plot every country's tone and Goldstein (conflict-cooperation) position alongside its direction of motion, two attractor basins emerge: a "healthy democracy" attractor in the upper-right (positive tone, cooperative coverage) and an "authoritarian floor" attractor in the lower-left (negative tone, conflictual coverage). The middle of the field is unstable: countries are pulled toward one basin or the other rather than settling in between. This is the geometric expression of what Thomas Carothers and Benjamin Press described in their 2022 Carnegie report as the bistability of contemporary democratic discourse.

Why this method works (and what it cannot do)

The method has a single methodological subtlety worth knowing about. GDELT switched from version 1 to version 2 in February 2015, and the switch produced a sudden 6-point shift in measured tone that has nothing to do with politics — it was a measurement change. Working around this required a z-score correction (comparing each country to the global average within the same year) and the use of Goldstein scores, which are derived from event codes and were not affected by the version change. Once these corrections are in place, the long-run signal is clean.

What β does not do is read intent. It tells you that a country's coverage has become more reactive to conflict; it cannot tell you whether that is because the country is genuinely more conflictual, because the media is more polarized, or because reporting practices have changed. Used alongside other indicators (the Economist Intelligence Unit's Democracy Index, V-Dem, polity scores), β adds a real-time media-side signal that the other indices cannot offer.

Why it matters

β can be computed in near-real-time from public GDELT data. The Economist Intelligence Unit's Democracy Index is annual and based on expert surveys. V-Dem is annual and based on coding. β is daily, mechanical, and free. It will not replace those measures, but it provides a leading-indicator complement that they cannot match.

The eight countries flagged as canaries deserve closer attention. Honduras and Uganda show the largest accelerations; Turkey and Mexico are also there. Of those, only Turkey and Tanzania appear on most existing backsliding watchlists. The others would not be obvious from existing indicators alone.

The full preprint, all the code, and the underlying GDELT queries are available at github.com/alejandrofeged/conflict-framing. Feedback and replication attempts welcome.

---

Alejandro Feged-Rivadeneira is a researcher at Tecnológico de Monterrey working on computational social science, political economy of media, and human ecology.