Two kinds of rating
Cogito rates bias in two different ways, and it helps to know which one you’re looking at.
Outlet ratings
The colored label on every story in the feed and on the homepage. It describes the publication as a whole, so it is fast and consistent, but it can’t see what any single story says.
Article analysis
What the Bias Detector does. It reads the words of one article, so it can catch a slanted piece in a center-rated paper. It is slower, and like any model it can be wrong.
Outlet ratings
Cogito keeps a list of 1,609 news websites, each rated on a five-point scale compiled from independent media-bias ratings. When a story appears, Cogito matches the website it came from against the list and shows that rating. Sites that aren’t on the list are marked Unknown rather than guessed.
| Label | What it means | Sites |
|---|---|---|
| Left | Consistently progressive in the stories it picks and how it frames them | 264 |
| Lean Left | Somewhat progressive | 520 |
| Center | No consistent lean, or a balance of views | 351 |
| Lean Right | Somewhat conservative | 219 |
| Right | Consistently conservative in the stories it picks and how it frames them | 255 |
| Unknown | Not on the list yet. This is not the same as Center. | – |
A rating describes an outlet, not a story. A center-rated wire service can run a one-sided piece, and an opinion column in any paper argues a side by design. To check a single article, use the Bias Detector.
The Bias Detector
When you paste a link, Cogito works through four steps.
-
Read the page
Cogito’s server downloads the article and pulls out the main text with trafilatura, an open-source extraction library, leaving behind menus, ads and comments. Some sites block automated readers or sit behind a paywall, and those can’t be analyzed.
-
Check that it’s political
The first model scores how much the text reads as political versus plain factual reporting. The article only counts as political when its political score beats its factual score by more than 20 percentage points. Otherwise Cogito calls it non-political instead of forcing a lean onto, say, a sports recap.
-
Read the lean
The second model is a political-leaning classifier published by researchers Matous Volf and Jakub Simko in 2025. It was trained on 12 datasets of news articles, social media posts and political statements, and it labels the text left, center or right. Cogito shows center as Neutral.
-
Break it into passages
With Deep Analysis Mode on, the text is split into passages of three sentences. Each passage goes through steps 2 and 3 on its own: non-political passages count as center, and political ones get their own left, center or right label.
The results show how many passages landed in each group, plus up to three passages that match the overall result, longest first. Each one shows the model’s confidence: High (96% or more), Medium (80% to 96%) or Low (under 80%).
Turning Deep Analysis Mode off skips step 4 and returns only the overall label, which is faster. The models run on Hugging Face, and if they haven’t been used for a while the first analysis can take up to a minute while they start up.
What the models respond to is language: word choice, framing and which points get emphasis. They don’t check whether claims are true.
The political quiz
The quiz asks you to rate ten statements from Strongly disagree to Strongly agree. Five are about economics: free markets, universal healthcare, taxes on the wealthy, social safety nets and business regulation. Five are about society: law enforcement, personal choice, traditional values, social change and civil liberties.
Each answer counts from −2 to +2. Some statements are worded so that agreeing moves you right or toward authority, and others so that agreeing moves you left or toward personal liberty. Each axis adds up its five answers and divides by ten, giving a score from −1 to +1.
Those two scores place you in one of four quadrants: Authoritarian Left, Authoritarian Right, Libertarian Left or Libertarian Right. If a score lands exactly on zero, the result says so instead of picking a side: Left or Right when the social score is zero, and Centrist when the economic score is zero.
The news feed only uses the left or right half of your result. If you land on the left, Only opposing views shows you right-leaning and center outlets. If you land on the right, it shows you left-leaning and center outlets. A Centrist result has no single other side, so the feed keeps mixing all sides. Your answers are scored on Cogito’s server, and the result is saved only in a cookie in your browser.
The news feed
Searches run through the Brave Search News API, which returns up to 50 recent stories per search. Cogito labels each one with its outlet’s rating, then orders them in one of two ways.
- Filter off
- Stories rotate between left-leaning, right-leaning and center outlets, so no side takes over the top of the page. Unrated outlets come last.
- Filter on
- The search is re-ranked with a Brave Goggle, a set of custom search rules that boosts outlets rated on the other side from yours. Cogito then keeps only outlets rated on that side or Center and removes the rest, and it takes one story per outlet in turn so no single site fills the page.
- Nothing found
- If a topic has no coverage from the other side, Cogito says so at the top of the results and shows a labeled mix of all sides instead.
Results for the same search are reused for five minutes, which keeps the site within the news service’s limits.
The homepage
Once a day, starting at midnight New York time, Cogito searches five broad topics (politics, elections, government, Congress and the president) and labels every result. The top stories alternate between left- and right-leaning outlets, with center and unrated outlets filling gaps, and no single group may take more than 60% of the top twelve. The Across the spectrum row picks one story each from Center, Left, Lean Left, Lean Right and Right when those are available.
Limitations
- Lean isn’t accuracy
- A left- or right-leaning article can be factually correct, and a neutral one can be wrong.
- Models make mistakes
- The models learned mostly from U.S. political writing. Satire, quotes from the other side, politics outside the U.S. and very short texts are harder for them. That’s why the Bias Detector shows the passages behind its result.
- Outlet ratings are coarse
- One label covers everything a publication prints, and outlets change over time.
- The quiz is short
- Ten statements can only sort you into a broad quadrant. It is a starting point, not a validated survey.
- Search sets the limits
- Cogito can only show what the news search returns, and some topics get little coverage from one side.