How The Balanced News App Detects Bias Across 50+ Sources

TL;DR: The Balanced News reads the same story from 50+ Indian news outlets, groups matching articles together, then uses AI to score each version's political lean and sentiment. Instead of telling you what to think, it shows you the full picture and lets you decide.

The Problem With Reading Just One Source

Pick any political story trending in India right now. Read it on NDTV, then read the same story on Republic. You will walk away thinking two completely different things happened.

This is not new. What is new is that most people do not realize it is happening. They open one app, read one version, and assume they got the full story.

The Balanced News was built to fix this specific problem. Not by creating yet another "neutral" newsroom (that concept is a myth), but by aggregating coverage from across the political spectrum and making the differences visible.

Here is how the system actually works, from raw article to bias score.

Step 1: Crawling 50+ Indian News Sources

The pipeline starts with ingestion. TBN continuously crawls articles from over 50 Indian news publishers. These include national English-language outlets (The Hindu, Hindustan Times, India Today, NDTV, Republic, Times of India, The Wire, Swarajya, OpIndia, Scroll, The Print, Mint, Economic Times), Hindi outlets, and regional publishers.

The goal is breadth. A single story like "PM Modi Addresses Lok Sabha on West Asia Conflict" might generate 34 separate articles across these sources within hours. Each article enters the system with metadata: source name, publication time, full text, and the source's historical political leaning.

This is already a departure from platforms like AllSides or Ad Fontes Media, which primarily rate Western English-language outlets. TBN is built for the Indian media landscape, where the spectrum runs from OpIndia on the right to The Wire on the left, with dozens of outlets occupying different positions in between.

Step 2: Grouping Articles Into Stories

Raw articles are useless in isolation. The system needs to know which articles are about the same event.

TBN uses semantic similarity to cluster articles into "story groups." This is not simple keyword matching. Two articles could share zero common phrases but describe the same event. The grouping engine uses vector embeddings to understand meaning, not just words, and clusters articles that are about the same underlying news event.

A story group for the LaGuardia airport collision this week, for instance, pulled together 65 separate articles from different outlets into a single coherent story. The petrol price hike story grouped 44 articles. For political stories, the clustering is especially important because the same event gets repackaged with wildly different framing depending on the outlet.

Step 3: Scoring Political Lean With AI

This is the core of what makes TBN different.

Once articles are grouped, the system analyses each group's collective coverage and assigns a political lean score across three dimensions: Left, Centre, and Right.

These are not static labels assigned to outlets (though source-level historical lean is one input). The scoring happens at the story level, using large language models trained to evaluate framing, language choices, source attribution, and emphasis patterns.

Here is a real example from this week. PM Modi's address to the Lok Sabha on the West Asia conflict was covered by 34 sources. The AI scored the overall coverage as:

• Left: 14% | Centre: 64% | Right: 22%

The reasoning: coverage primarily presented government perspectives, with opposition viewpoints from Congress leaders calling for broader parliamentary discussion. The mix of ruling party narratives and calls for inclusive debate produced a centre-heavy distribution with notable right lean from pro-government framing.

Compare that with a non-political story like the LaGuardia airport crash (65 sources): Left: 0% | Centre: 100% | Right: 0%. Factual, operational coverage with no political framing at all.

This is what makes live, per-story scoring more useful than a static bias chart. An outlet rated "centre" overall might produce distinctly right-leaning coverage on a specific topic. TBN catches that.

How This Differs From AllSides and Ad Fontes

Platforms like AllSides use blind surveys, editorial panels, and independent reviews to place outlets on a five-point scale (Left to Right). Ad Fontes Media uses multi-analyst content analysis across bias and reliability axes. Both are respected, but they share a limitation: they rate outlets, not individual stories.

A 2025 systematic review in Social Science Computer Review found that most automated bias detection tools still struggle with consistent definitions of what "bias" means. TBN sidesteps this partly by not claiming any single article is "biased" or "unbiased." Instead, it shows you where the coverage cluster sits on the spectrum and lets you read across the range.

Step 4: Sentiment Analysis

Beyond political lean, every story group gets a sentiment score from 0 to 100.

• Below 40: Negative tone (crisis, criticism, conflict)
• 40 to 60: Neutral
• Above 60: Positive tone (celebration, achievement, optimism)

The Modi Lok Sabha story scored 42 (slightly negative, reflecting the tension around the West Asia conflict and opposition criticism). The gold price crash story scored lower. A cricket event like CSK's ROAR 2026 fan gathering scored significantly higher.

Sentiment scoring adds a second dimension. Two stories might both be centre-leaning politically, but one could be framed in doom-and-gloom language while the other reads as measured and analytical. TBN surfaces this difference.

Step 5: The Balanced Summary

For every story group, the system generates what it calls a "balanced summary." This is not a copy-paste from any single article. It is an AI-generated synthesis that attempts to include perspectives from across the coverage spectrum.

The LaGuardia crash summary, for instance, incorporated details from 65 different articles into a single coherent paragraph covering the collision timeline, fatalities, injuries, emergency response, and investigation status. No single outlet covered all of these angles equally well.

For political stories, the balanced summary is more deliberately constructed. It includes the ruling party's position, opposition responses, and factual context that might be emphasized by some outlets and downplayed by others.

What "Left," "Centre," and "Right" Mean in Indian Context

This is an important distinction TBN makes that Western bias tools miss entirely.

In Indian politics, Right-leaning coverage tends to be pro-BJP, pro-government, and aligned with Hindutva ideology. Left-leaning coverage is typically opposition-aligned, critical of the ruling party, and often associated with outlets like NDTV, The Wire, or Scroll.

This is roughly the inverse of the Western framework where "right-wing media" means Fox News and "left-wing media" means MSNBC. An Indian reader using AllSides' definitions would get confused. TBN's scoring is calibrated for the Indian political spectrum.

As we showed in our previous analysis, How Indian Media Is Framing the West Asia Conflict, the same international event gets filtered through domestic political lenses in ways that are unique to India's media ecosystem.

The Tech Under the Hood

TBN's bias detection pipeline uses a combination of technologies:

Article clustering relies on vector embeddings and semantic similarity to group articles about the same event, regardless of headline phrasing.

Bias scoring uses large language models (LLMs) that evaluate each article cluster's framing patterns, language intensity, source attribution, and emphasis. The LLM does not just count keywords. It reads the articles, understands context, and evaluates how the story is being told differently across outlets.

Sentiment analysis uses a similar LLM-based approach, scoring emotional tone on a continuous scale rather than simple positive/negative buckets.

This is consistent with the direction the field is moving. Research from the University of Pennsylvania's Annenberg School confirms that "the methods that could do this sort of classification at scale didn't work sufficiently well until the latest generation of large language models arrived." Earlier approaches using keyword frequency or basic NLP were too crude to catch framing bias, which is the most common and most dangerous type.

What TBN Does Not Do

Transparency matters, so here is what the system deliberately avoids:

• It does not rate outlets as permanently biased. A source's lean can shift depending on the topic.
• It does not fact-check. Bias detection and fact-checking are separate problems. TBN focuses on the former.
• It does not tell you which version is "correct." It presents the spectrum and trusts you to evaluate.
• It does not hide sources you might disagree with. The whole point is showing you perspectives you would not seek out on your own.

This is a philosophical choice. As we wrote in How News Apps Decide Which Stories You Never See, most news platforms use engagement-driven algorithms that create filter bubbles. TBN does the opposite: it deliberately surfaces the coverage you are most likely to skip.

Why This Matters

India has over 500 million smartphone users consuming news daily. Most of them read from one or two sources, usually recommended by an algorithm optimised for engagement, not understanding.

The result is a population that is simultaneously hyper-informed and deeply fragmented. Everyone has access to information. Almost nobody has access to the full picture.

TBN's bias detection system does not solve this problem completely. No tool can. But it makes the invisible visible. When you can see that a story is being told three different ways by three different outlets, you stop assuming your version is the only version.

That shift, from passive consumption to active evaluation, is what media literacy actually looks like in practice.

---

Sources: - AllSides Media Bias Chart - Ad Fontes Media Methodology - Automated Detection of Media Bias: A Systematic Review (2025) - AI-Powered Bias Detector Transforms News Analysis, UPenn Annenberg - Media Bias Detector, CHI 2025

Explore TBN's bias scoring in action at thebalanced.news. Every trending story shows the full political lean breakdown, sentiment score, and coverage from across the spectrum.