How We Detect Bias at JQJO

We do not permanently label any outlet as “left” or “right” as a whole. Bias is evaluated article by article, story by story, using a mix of AI and human editorial review.

1. What We Mean by “Bias”

When we talk about bias, we are not talking about:

• whether a government or political party likes the story
• whether an outlet has a certain reputation
• whether a headline makes someone “feel” attacked

We are specifically looking at how a piece of journalism frames reality compared to other credible coverage of the same event.

Key questions we ask:

• Which facts are emphasized, downplayed, or ignored?
• Is one side’s argument presented in detail while the other is summarized or dismissed?
• Are emotionally loaded words used for one side and neutral terms for the other?
• Are claims from one side scrutinized while the other side’s claims are taken at face value?

We then classify articles within a story into three broad buckets:

• Left-leaning framing
• Center/neutral framing
• Right-leaning framing

This is per-article, not a permanent tag on an outlet.

2. Article-First, Not Outlet-First

Most bias trackers slap a label on the entire outlet (“this site is left/right/center”) and call it a day.

We don’t do that.

Our approach:

1. Every story cluster (a single “event” with many articles) is treated as fresh.
2. Each article in that cluster is analyzed individually.
3. The same outlet might have:
   • a center-leaning straight news report on one topic, and
   • a left or right-leaning piece on another.

We may use outlet history as context, but the bias label we show is tied to the specific article(s) informing that segment of coverage, not a blanket judgment on that brand for all time.

3. End-to-End Workflow for Bias Detection

3.1 Ingesting and grouping coverage

1. Our systems ingest up to 100,000+ articles per day from a wide range of sources.
2. AI models detect when multiple articles describe the same underlying event or issue and cluster them into one story.
3. Each cluster may contain coverage from national, local, international, and niche outlets.

Only after this clustering do we start bias analysis.

3.2 AI pre-analysis of each article

For each article in a cluster, AI runs a first-pass analysis that looks at:

• Language and tone
  • adjectives and verbs used for people, parties, policies
  • emotional or moral language (“outrageous”, “evil”, “heroic”, “unpatriotic” etc.)
  • repeated phrases strongly associated with partisan narratives
• Issue framing
  • which problems are highlighted (e.g. inequality vs regulation, crime vs policing, etc.)
  • which solutions are presented (state intervention vs markets vs personal responsibility, etc.)
  • whether coverage leans toward progressive/liberal frames, centrist frames, or conservative/traditional frames
• Source quoting patterns
  • which voices get direct quotes versus paraphrase
  • whether opposing views are given meaningful space or token lines
  • existence of one-sided sourcing (e.g. only government, only activists, only party officials)

The AI then produces a draft classification for that article:

• likely left-leaning
• likely center/neutral
• likely right-leaning

This is a draft. It is not published directly.

3.3 Human editorial review and override

Our editorial team reviews:

• a subset of key articles for each story (especially those that AI flags as strongly biased), and
• the overall pattern of coverage across left/center/right estimates.

Editors:

1. Read the article(s) in question in full, not just excerpts.
2. Compare them against coverage of the same story from other outlets.
3. Ask:
   • Is the language loaded or neutral?
   • Are major facts omitted or misrepresented?
   • Are opposing views treated fairly?
   • Does the outlet’s framing line up with known progressive, centrist, or conservative narratives on this issue?

They then either:

• Confirm AI’s suggested bias label, or
• Override it and assign a new label based on human judgment.

Where a piece is genuinely balanced or purely factual with clearly separated opinion, it will be classified as Center.

4. How We Turn Article Labels into the Story Bias Bar

On JQJO, you often see something like:

Left 33% – Center 50% – Right 17% Sources: 6

This reflects:

• The proportion of articles in that story cluster that our process categorized as:
  • left-leaning,
  • center/neutral,
  • right-leaning.

Example:

• Cluster has 6 articles we considered relevant enough to weigh:
• • 2 are left-leaning
  • 3 are center
  • 1 is right-leaning

We show:

• Left 33% (2/6)
• Center 50% (3/6)
• Right 17% (1/6)

The exact percentage is derived from article-level labels, not some global outlet score.

5. What We Look for When Calling Something “Left” or “Right”

We are not guessing based on vibes. We look for repeatable patterns that are widely recognized in political science and media analysis.

5.1 Left-leaning article patterns (for that story)

Things we might see:

• Emphasis on:
  • inequality, marginalized groups, systemic injustice
  • government responsibility to intervene
  • corporate or wealthy interests as primary drivers of harm
• Language that:
  • highlights discrimination, climate risk, worker rights, social safety nets
  • frames policy as morally necessary for vulnerable groups
• Framing where:
  • conservative arguments are minimized or presented mostly to be refuted
  • progressive policy positions are described more sympathetically or as the default

5.2 Right-leaning article patterns (for that story)

Things we might see:

• Emphasis on:
  • law and order, border control, national sovereignty
  • personal responsibility, free markets, smaller government
  • threats to tradition, family, or religious values
• Language that:
  • highlights crime, disorder, overregulation, cultural decline
  • emphasizes costs of welfare or social programs over benefits
• Framing where:
  • progressive positions are minimized or portrayed as naive/dangerous
  • conservative policy positions are treated as common sense or default

5.3 Center / neutral article patterns

Articles we mark as Center typically:

• separate facts from opinion clearly
• give meaningful space to multiple sides of the issue
• use neutral, descriptive language rather than emotionally loaded terms
• avoid uncorroborated claims or speculative leaps
• acknowledge uncertainty, limitations, or ongoing investigation

A “center” label does not mean “perfect” journalism, but it signals that we do not see strong partisan framing in how the story is told.

6. Special Cases and Edge Situations

6.1 Opinion pieces vs straight news

• If it is clearly labelled as Opinion / Editorial by the outlet, we treat it under a different lens:
  • Bias is expected, so the question is where it leans, not whether it leans.
• Opinion pieces can still feed into our understanding of how a story is being framed, but they are not treated as neutral reports.

When possible in the UI, we distinguish between:

• News coverage
• Opinion coverage

6.2 Breaking news

When events are breaking:

• There may be fewer articles and less time for manual review.
• We may temporarily rely more heavily on AI’s draft labels plus quick human checks.
• As more coverage appears and editors have more time, bias labels and distributions can be updated.

6.3 Satire, parody, and clickbait

• Satire is generally excluded from bias metrics for serious stories.
• Parody or memes are not treated as news sources.
• Pure clickbait with no meaningful reporting may be filtered out or given minimal weight.

If an outlet mixes satire and news poorly (user can’t easily tell), we reduce its relevance for bias calculations.

6.4 Mis- and disinformation

If coverage clearly:

• spreads false claims refuted by strong evidence, or
• relies on fabricated sources or debunked conspiracy narratives,

we treat that outlet’s article as low-credibility for that story. It may be:

• excluded from bias charts,
• flagged internally, or
• used only at the editorial team’s discretion (for example, as an example of misinformation patterns, not a neutral “right” or “left” perspective).

7. Human Responsibility and Limitations

Our process is serious, but not perfect.

7.1 Things we openly acknowledge

• Bias detection is partly an art: some calls are close.
• Even our editors bring their own experiences and perspectives.
• AI can inherit patterns from the data it was trained on.

This is exactly why we:

• rely on multiple editors, not a single person’s judgment,
• combine quantitative signals (language, sources, framing) with qualitative review,
• treat labels as reviseable rather than frozen forever.

7.2 Corrections and disputes

If you believe we misclassified coverage on a specific story:

• you can contact us with:
  • the JQJO URL,
  • the specific outlet/article,
  • and a reasoned explanation of why you think the label is wrong.

We will review substantial, good-faith feedback and adjust bias labels where appropriate.

8. Why We Show “Coverage of Story: From Left / Center / Right”

Most readers live inside one media bubble: they see the same style of coverage over and over, and they rarely realize how differently the same story is framed elsewhere.

Our “Coverage of Story” section:

• groups headlines and links into:
  • From Left
  • From Center
  • From Right

Each group includes:

• direct links to original articles, and
• often a short framing description.

This is meant to expose the information environment, not tell you what to think. You see:

• how left-leaning outlets describe the story,
• how center outlets phrase it, and
• how right-leaning outlets talk about the same event.

Step outside your bubble in one screen.

9. How This Fits With Our Mission

We’re a small team, but we believe that in 2025 and beyond, the volume of misinformation, spin, and partisan narrative is only going to increase.

So we built JQJO to:

• collect a huge volume of coverage,
• cluster related articles into single events,
• analyze how different sides frame the same reality, and
• present that to you in clear, transparent ways.

Our bias detection process is designed to be:

• article-level, not outlet-level
• transparent enough to explain
• strict enough to be useful, but humble enough to admit its limits.

If you rely on JQJO, you should always know not just what happened, but how different sides are telling you the story and who that storytelling may be serving.