Methodology

What unslant does, and what it doesn’t

unslant shows how a text may shape your reading: loaded words, one-sided sources, numbers without context, fallacies, omissions. It does not say whether the author means well or ill, does not decide whom you should believe, and does not replace your own reading.

Bias is not the same as lying. A text can contain only true facts and still select and arrange them to push the reader toward a conclusion. Likewise, an opinion piece can argue for a position without being dishonest: there, unslant looks for distortions, omissions and fallacies, not for the existence of an opinion.

How an analysis works

1. Extraction. We separate the main text from menus, ads and footers. Tables become readable rows. PDFs are read in your own browser, in the right order even in two columns, without the headers and footers repeated on every page.
2. Identification. The AI identifies the type of text (news, feature, analysis, opinion, editorial, paper…), the country and the relevant political context. The type follows fixed criteria (for example: a signed piece arguing a position is opinion; a piece stating the outlet’s own position is an editorial) and uses the section shown by the page as a hint, when there is one. The reason appears in the result. If you know the type, you can choose it in the form, and your choice overrides the AI.
3. Specialists. The text is examined at the same time by several “specialists”, each looking at a few groups of biases only (language, sources, data, fallacies…). One of them also assesses political leaning, the share of fact and opinion, and possible omissions.
4. Images and charts. At the same time, the AI examines the content’s images: it describes each chart, table or photo, compares it with what the text says, and looks for misleading charts (an axis that does not start at zero, a cherry-picked time window, a distorted scale) and visual framing. It can only point to figures it actually received.
5. Verification. Every passage quoted as evidence is searched for in the original text. Anything not found is discarded (see below). Scores are calculated by the site itself, with a fixed formula.
6. Synthesis. A final summary brings the findings together and suggests how a problematic passage could be rewritten neutrally. The neutral version never adds sources, numbers or facts that are not in the original; the site checks this automatically and discards the suggestion if it invents anything.

There are 54 categories in 12 groups. Some only apply to certain content: the academic group is used only for papers, and the reasoning-bias group only for claims. See the full list in Types of bias.

Every finding must be backed by the text itself

By default, the analyzed text shows only moderate and strong highlights; mild ones can be revealed with a button. Passages flagged in more than one category appear as a single highlight with all explanations. Verbs such as “stated” or “claimed” are not flagged on their own, only when the sides are treated unequally.

The AI may only flag a bias if it quotes the exact passage that shows it. Before showing the result, the site searches for each quote in the original text, tolerating only differences in quotation marks, spacing and capitalization. If the quote does not exist, it is discarded and the category loses that finding. The number of discarded passages is shown at the end of each analysis.

This keeps the AI from “inventing” problems and lets you check every finding yourself: just click the highlighted passage.

Quotes belong to the speaker, not the author. When the text quotes a politician, expert or interviewee, that quote is not treated as the author’s opinion presented as fact, nor as the author’s language. The site checks this automatically and discards such findings. Quotes can still count when the problem is the author’s choice, for example giving voice to only one side.

How the 0–10 score is calculated

Each category gets a score on this scale:

• 0: no problem
• 1 to 3: mild or isolated problems
• 4 to 6: clear, recurring problems
• 7 and 8: strong bias that shapes the reading
• 9 and 10: severe, systematic distortion

Each score comes with a confidence level (low, medium or high). The AI distinguishes two situations: “no problem” (the text has material for that category and handles it well, which counts as 0) and “not applicable” (the text has no material for it, like data categories in a text without numbers). Categories that do not apply are left out of the calculation, and the result shows how many there were.

To keep scores consistent, the AI is given reference descriptions: a 2 is a balanced story with the odd unnecessary adjective; a 5 gives far more space to one side and uses some loaded words but still reports the facts; an 8 is campaign material dressed up as news.

Each group’s score combines its worst category (50%) with the average of its categories (50%), where low-confidence findings weigh less.

The overall score combines the weighted average of the groups (65%) with the average of the 3 highest-scoring groups (35%). The second part keeps a text that is heavily biased in a few aspects from looking balanced just because it is fine in the others. Current group weights:

• Language: 1
• Framing and structure: 1
• Sources: 1
• Omission and selection: 1
• Data and statistics: 0.9
• Fact vs. opinion: 1
• Emotional appeal and sensationalism: 0.8
• Logical fallacies: 0.8
• Interests and conflicts: 0.7
• Representation: 0.9
• Reasoning biases: 1.1
• Academic paper issues: 1.2

And the overall score bands:

• 0 – 2: Mild bias
• 2.1 – 4: Moderate bias
• 4.1 – 6: Marked bias
• 6.1 – 8: Strong bias
• 8.1 – 10: Very strong bias

The score measures how biased the text is overall, not just politically. A text can score high with no political leaning at all, for example through sensationalism or misleading use of numbers.

Political leaning

Leaning is assessed in three parts. The first is which side of a dispute the text favors: when the text covers a concrete dispute (government × opposition, candidate A × candidate B), the AI says whether it favors one side, is balanced, or cannot be determined. To apply the same yardstick to both sides, they are always listed in the same order: first whoever holds power, then whoever is outside it. The classification must be backed by a passage from the text and states the intensity (slightly, moderately or strongly). The analysis summary is instructed to say the same thing as this marker.

The other two are ideological axes, filled in only when the topic really involves the economy or social values. An electoral fight, for example, is not an economic topic in itself; in that case the axis shows as “not applicable”.

• Economic: on the left, a larger role for the state in the economy, redistribution and social protection; on the right, free markets, lower taxes and spending, privatization.
• Social: on the left, progressive values and expanded minority rights; on the right, traditional values, religion, emphasis on order and security.

Three principles guide this assessment. First, it measures how the text treats the topic, not the topic itself: a balanced story about a left-wing policy is not a left-wing text. Second, the definitions are adapted to the text’s country, because political spectrums differ between countries. Third, “Unidentified” is used whenever there is not enough evidence. This is common and expected: many texts have no clear leaning, and the tool does not force a label.

How we pursue impartiality

• Same yardstick for everyone. The AI’s instructions are the same for any text, outlet or political side.
• Inversion test. Before flagging something, the AI is instructed to ask whether it would flag the same thing in a text favoring the opposite side.
• Opposite-sides test. We keep a set of real texts about the same event, from outlets on opposite sides and from a more neutral outlet. Every change to the site goes through an automatic test: both sides must get similar scores (at most 1.5 points apart) and opposite leanings, and the more neutral text must get a low score.
• Neutral colors. Left and right never appear in different colors. The site’s colors only indicate how severe a problem is.

To balance your reading

After the analysis, the AI runs a few web searches (at most 3) for other coverage of the same topic, from different outlets and perspectives, plus data and primary sources. For each text it finds, it explains what that text adds compared with the original.

• Real links only. The site keeps the list of results returned by the search and discards any link suggested by the AI that is not on that list.
• Different perspectives. The AI is instructed never to recommend outlets from a single side and to prioritize reliable sources. The original text’s outlet is excluded. When the text leans one way, the site requires at least one reliable outlet from the other side: if the search finds none, it runs one more search limited to outlets from the other side (in the US, for example, the Wall Street Journal or National Review for left-leaning texts; The Atlantic or NPR for right-leaning ones).
• Editorial line. Each suggestion shows the outlet’s editorial line (left, center-left, center, center-right, right, official, legal, fact-checking, academic, news agency). About 78 outlets and bodies come from a fixed list kept by the site, and each classification has a public source with a link: the Media Bias/Fact Check rating or, for Brazilian outlets it does not rate, the description of the editorial line on Portuguese Wikipedia, with the month of the rating. Public bodies (government, Congress, courts) appear as “official”. Outlets without a clear source are left off the list: for them the AI makes an estimate, marked as “estimate”.
• Not an endorsement. Recommending a text does not mean it is neutral or correct: it helps you see the topic from another angle. Read it with the same critical eye.

Claims

In the Claim tab you can analyze a short statement, such as something you heard in a conversation or saw on social media. Besides language and logic biases, unslant looks at reasoning biases (such as generalizing from a few cases) and runs a claim check: what kind of claim it is, whether it can be checked, what is known about the topic, why the idea circulates and where to check it.

The claim check relies on the AI’s general knowledge, which has a cutoff date and may be outdated or wrong. It does not consult sources during the check. That is why the AI is told to prefer “not enough evidence” when unsure. Right below, the “Where to check this claim” section lists real sources found on the web so you can check for yourself.

Limitations you should know about

• Images: only what is an image. The AI examines the images inside the article (up to 6) and, in PDFs, the pages with figures, charts or tables (up to 20). Embedded interactive charts (such as Datawrapper or Flourish), videos and text pasted in the Text tab have no images to analyze. Scanned PDFs (images only, no text) are not accepted.
• Long texts take longer. To respect the AI’s per-minute usage limit, long texts are analyzed in stages, and the screen says so when that happens.
• The AI can be wrong and biased too. Language models learn from human writing and carry its biases. The rules above reduce this risk but do not remove it.
• Results can vary slightly. Analyzing the same text twice may produce slightly different scores. Small differences (half a point, for example) should not be read as meaningful.
• Irony, satire and humor can be misread.
• Very specific local context (small-town politics, slang, very recent events) may escape the AI.
• Page reading can fail. Some sites block automated readers or hide text behind a paywall. If so, paste the text into the Text tab.
• The score is an approximation, useful for comparing and for flagging problems, but not a verdict on the text, the author or the outlet.

Technology

Analyses are performed by the Claude Sonnet 5.5 (Anthropic) language model. The AI’s answers follow a fixed format that is checked automatically; when an answer is malformed, the step is repeated. Page extraction uses Readability, the same technology behind Firefox’s reader view.

How to get the most out of unslant

• Read the evidence, not just the score. Click the highlighted passages and see whether you agree.
• Compare with other sources on the same topic, ideally from different perspectives.
• Use the analysis as a starting point for thinking, not as the final word.