Detecting media bias at scale: the political-lean news API

The same event is rarely reported the same way twice. Comparing how state media, independent outlets, and Western wires frame a story used to be manual analyst work. Here's how to do it with a query.

Why bias labels matter for data teams

If you're tracking geopolitical risk, running an LLM, or monitoring narratives, raw articles aren't enough — you need to know who is saying something and from what angle. A casualty figure from a defense ministry, an opposition outlet, and a wire service are three different signals. Without a bias label, your pipeline treats them identically, and your analysts spend hours on source triage.

A 9-category lean, applied at the source level

NewsAgent Data classifies every article's political_lean across nine categories:

• state — state-controlled media
• official — government & institutional press offices
• centrist — wire services & non-partisan press
• liberal · conservative · nationalist · opposition · tabloid · neutral

Because it's applied consistently across thousands of Russian and English sources, you can slice any feed by editorial stance without maintaining your own source list.

Comparing narratives in three calls

Take any event — articles covering it share a cluster_id. Pull the same cluster filtered by lean and you get an instant side-by-side of how each camp framed it:

GET /v1/feed?cluster_id=183420&political_lean=state
GET /v1/feed?cluster_id=183420&political_lean=opposition
GET /v1/feed?cluster_id=183420&political_lean=centrist

Three requests, three framings of one event — no NLP, no manual tagging.

Use cases

• Geopolitical risk: isolate the official state narrative on a topic (political_lean=state&topic=sanctions).
• LLM grounding: feed contrastive pairs (same event, opposing lean) as labeled training or retrieval data.
• Media monitoring: measure how a story spreads across the spectrum over time, using cluster_size and fetched_at.

How the labels are assigned

Lean is mapped at the source level from a curated classification, then inherited by every article from that source — deterministic and reproducible, not a black-box model guess. That makes it auditable: you always know why an article carries the label it does.