Spuuun. The news, with receipts
How to read this
Pick a story. This sets the outlets that covered it side by side. Compare by groups them by the one fact on record about each outlet that you choose: the language it published in, where it is based, or its owner. A group is only the outlets that share that fact.
Every count comes from the outlets' own headlines and articles, and opens onto the outlets behind it, each with a link to its article. The statements outlets carried are listed in the Analysis pane. If an outlet is missing from a statement's list there, the statement was not matched in that outlet's articles, and our matching can miss one.
Headlines appear in English where we machine-translated them, with the original beside each one. Everything in this panel is a computed inference and passes no judgement on any outlet.
Select a story to see how outlets headlined and covered it.
Controls
These two are mentioned together in tracked coverage.
A line marks co-occurrence. It does not show that one caused the other. Lines reflect only the sources we track.
How each outlet covered it
Select a story, a theme or a name to compare how outlets covered it.
This selection's tone
Select a theme, story or sourceTone over time
Overall tone
Select something to see its tone.Tone toward the top subject
Select a theme or storyTone over time
Overall tone
Select something to see its tone.Connected entities
- Select an entity to see its connections.
Mentioned in
- Select an entity to see its stories on the canvas.
About this view
- This selection's tone. The average of every per-article tone reading for the articles in this selection, the same coverage the spine shows, across all our methods. This measures the tone of coverage and nothing more.
- Tone over time. Each point is the average tone of the coverage in that time window. Above the line = warmer tone; below = cooler. Tone ≠ truth. A computed inference over our tracked sources only.
- In-depth AI critique. An AI model re-reads a theme's coverage several times and reports a consistency-checked tone rationale, run only on the most salient themes. AI-assisted, so treat it as a signal. You decide what it means.
- Tone toward the top subject. A wider zoom-out: how all our tracked sources frame this subject, beyond just this selection's coverage. A bigger magnitude means a more confident reading, and no reading is certain. The small button beside a subject brings its stories up on the canvas.
- Connected entities. Other names on this subject's stories on the canvas, with how many stories they share. Click one to bring the shared stories up on the canvas. Mentioned in lists this subject's stories on the canvas.
- How each outlet covered it. For each outlet: how many articles it ran and when we first saw one, whom its quotes name, and the people and organisations its own articles name. A line under "Where an outlet differs" means it quoted different people, or most often named someone else, than the outlets on the same stories. These are counts from the articles we track. They say nothing about which outlet is right.
Three ways we read tone
We read tone three independent ways and use each for what it does best. All are approximate, computed inferences about the tone of coverage. They say nothing about what is true, and we never combine them into a single "bias" score.
-
Document tone…
Document tone. A fast transformer reads the whole article and scores its overall tone, then attributes that tone to each subject it names. Quick and broad, but article-level: tone ≠ truth, and it is not aimed at the subject.
Computed from the coverage we track. You decide what it means.
Document tone. A fast transformer reads the whole article and scores its overall tone, then attributes that tone to each subject it names. Quick and broad, but article-level: tone ≠ truth, and it is not aimed at the subject.
-
Toward-subject (zero-shot)…
Toward-subject (zero-shot). A natural-language-inference model judges whether the text reads positive, negative or neutral specifically toward this subject. Scales to every subject, but is an approximate, computed inference.
Computed from the coverage we track. You decide what it means.
Toward-subject (zero-shot). A natural-language-inference model judges whether the text reads positive, negative or neutral specifically toward this subject. Scales to every subject, but is an approximate, computed inference.
-
In-depth stance (LLM)…
In-depth stance (LLM). A large language model reads for nuanced aspect/stance toward the subject and gives a short rationale. Richer and slower, run only on the most salient stories; AI-generated, so treat the rationale as a signal.
Computed from the coverage we track. You decide what it means.
In-depth stance (LLM). A large language model reads for nuanced aspect/stance toward the subject and gives a short rationale. Richer and slower, run only on the most salient stories; AI-generated, so treat the rationale as a signal.
The outlets whose coverage feeds the currently selected bubble. Click a section, theme, or story on the left and this narrows to that level. Each outlet is listed once: where it is based, its owner, its language and when we first saw it, then its articles, each linked to the original. A large selection lists the articles of its 20 highest-ranked stories.