Barometer
How often the press uses the mechanisms in our glossary, judged by an AI model on an automatic sample of articles rather than on the ones we choose to critique.
Experimental: these figures are a preview and may change as the method is refined.
Week of 26 October – 1 November 2026, about
1 in 8
articles flagged by an AI model for at least one mechanism (12.6%, ±1.0 pts)
4170 articles sampled and judged
Measured from 26 October 2026: Proof by invocation, Attribution quarantine, Horse-race framing, Product-verb euphemism, Self-declared neutrality, Loaded-vocabulary adoption, Unsourced motive attribution; the share flagged for at least one mechanism counts them too, so it isn't comparable with earlier weeks.
Mechanisms, ranked
Under 0.5% this week: Adjective inflation, Agentless passive 0.2%, Attribution quarantine, Bothsidesism, Buried lede, Commitment-free statement 0.1%, Fossil legalism, Horse-race framing 0.1%, Market personification 0.1%, Modal deletion 0.1%, Nostalgia offset 0.3%, Product-verb euphemism, Proof by invocation 0.1%, Self-declared neutrality 0.1%.
By country
Each bar stacks one country's mechanism rates on one scale, so lengths compare. A country here means the outlets in our panel, not its whole press.
- Scare quotes
- Verdict adjectives
- Headline asserts, body hedges
- Presupposition
- Loaded-vocabulary adoption
- Unsourced motive attribution
- Nostalgia offset
- Agentless passive
- Market personification
- Horse-race framing
- Commitment-free statement
- Modal deletion
- Self-declared neutrality
- Proof by invocation
- Adjective inflation
The outlets in each country:
- UK: BBC News, GB News, The Guardian, The Independent, The Mirror
- France: France 24, France Info, France-Soir, RFI, Valeurs Actuelles
- USA: ABC News, CBS News, Fox News, NBC News, New York Post, NPR, PBS NewsHour
How this is measured
- A fixed panel of 17 outlets in the UK, France and the USA. Every 2 hours, up to 7 articles per outlet are drawn at random from what its RSS feed published since the previous draw: a sample, not everything they publish.
- Each article is checked by an AI model for each glossary mechanism, and flagged when the model is confident it shows it. A person has checked only a small share of the flags so far: read the figures as indicative.
- The ± figure is the 95% margin of error of a random sample of that many articles. Articles from the same outlet resemble each other, so the real margin is wider: read it as a minimum.
- Each article is judged on its headline and full text (for the rare very long piece, its first and last parts). Paywalled articles are left out.
- French articles are judged in French, so France is not strictly comparable with the other countries.
- The sample started on 12 October 2026. Days are counted in Paris time. These rates come from the sample alone, independent of which articles we critique.
Our critiques are separate: 60 published this week, 493 in total.
Data as of 15 November 2026.
Every week
- 9 – 15 November 20261 in 94183 articles
- 2 – 8 November 20261 in 94249 articles
- 26 October – 1 November 20261 in 84170 articles
- 19 – 25 October 20261 in 104107 articles
- 12 – 18 October 20261 in 94196 articles