VeritypeThe Veritype guide
Last updated: September 2, 2026
This page covers how to use Veritype, how it arrives at the numbers it shows you, and — the part most tools leave out — where those numbers are weakest. If you only read one section, read what Veritype cannot do.
What Veritype does
You choose a page, or select some text on it. Veritype sends that text to an AI model, which reads it for word choice, which sources are quoted or left out, and how the story is framed. The side panel then shows you where that framing sits on a left-to-right spectrum, the specific signals behind that assessment, the strongest reasonable counterarguments to what you just read, and a hedged read on whether the writing shows signals common in AI-generated text.
It analyses text you point it at. It is not a fact-checker. It does not verify whether a claim is true, and a piece can be scrupulously accurate and heavily framed at the same time — those are different questions, and Veritype only answers the second one.
Reading the spectrum
The analysis produces a score from −1 (strongly left) to +1 (strongly right). The panel never shows you that number. It shows a mark on a rule, and a coarse label:
| Score | Label |
|---|---|
| −1.0 to −0.5 | Strongly left-leaning |
| −0.5 to −0.15 | Left-leaning |
| −0.15 to +0.15 | Center |
| +0.15 to +0.5 | Right-leaning |
| +0.5 to +1.0 | Strongly right-leaning |
The label is deliberately coarse, and the mark is not. Five buckets is roughly the most a single reading of a single article honestly supports, so that is what the words say. But a piece at −0.2 and a piece at −0.45 are both “Left-leaning”, and they are not the same piece — which is why the mark sits where it sits on the rule rather than snapping to the middle of its band. If you want the finer reading, look at the position. If you want the defensible one, read the label.
“Center” is a description, not a grade. It does not mean accurate, balanced, fair, or good. It means the framing did not pull noticeably in either direction. A centrist piece can be wrong, and a strongly-leaning piece can be entirely factual. Veritype is describing how something is written, never scoring how much you should trust it.
The signals, and why yours are ordered the way they are
Underneath the summary are up to ten specific signals — the evidence for the assessment. Each is tagged with one of five dimensions:
- Sourcing — who is quoted, who is not, how claims are attributed
- Framing — emphasis, ordering, what is presented as the default view
- Data — figures and statistics, and how they are selected or compared
- Language — word choice, loaded terms, tone
- Methodology — how the underlying claim or study was arrived at
Not every dimension applies to every piece. Signals are listed most significant first, because most people read the first few and stop.
If you set a persona in the panel’s settings, that ordering changes — a journalist sees sourcing first, a researcher sees data and methodology first, a student sees language first. The analysis itself is identical either way. Personalisation reorders what you are shown; it does not change the score, the signals, or what the model was asked. We think that distinction matters: a tool that quietly told different people different things about the same article would be doing something rather different from what this one claims to do.
Counterarguments
The model is asked for the strongest reasonable counterarguments to the piece — not a rebuttal of its politics, and not a both-sides gesture. If an article argues something well, the counterarguments should be the ones a thoughtful opponent would actually raise.
The same instruction that governs the rest of the analysis applies here: describe the framing, do not take a side. Veritype is not trying to talk you out of what you read.
The AI-content read, and its limits
Veritype gives a hedged read on whether text shows signals common in AI-generated writing — repetitive phrasing, generic transitions, unusual uniformity, a lack of specific detail.
It reports one of three words, from a score between 0 and 1:
| Score | Shown as |
|---|---|
| below 0.3 | Low |
| 0.3 to 0.6 | Some |
| 0.6 and above | Notable |
Those cutoffs were chosen by judgement, not derived from calibration data. We do not have any yet. If we ever do, the boundaries may move and we will say so here. We publish them because a methodology page that quietly withholds its own numbers is not one.
You will never see a percentage, and that is deliberate. Current AI-detection is not reliable enough to support “73% AI-written”, and printing a number like that would be false precision stacked on top of an already uncertain assessment. Treat this as a weak signal, never a verdict — and never as grounds for accusing someone of anything.
Similar Sources
Similar Sources looks for real coverage of the same story from an outlet at a chosen point on the spectrum, so you can read across a story rather than through one window. Every result is a genuine article the tool verified; nothing is invented.
Two different messages are worth telling apart:
- “No matching coverage found at this position.” The search worked and found nothing. Usually that means what it says — thin parts of the spectrum genuinely have less coverage of some stories, and local, independent or primary-source reporting often has no counterpart anywhere. That absence is a finding about the story, not a failure of the tool.
- “The search failed. Try again in a moment.” That is the tool breaking. Retry it.
Which outlets can appear, how they are placed on the spectrum, and why a few carry a note about their factual record are all covered in the outlet methodology, which is where those decisions live.
Your reading
Not in the released extension yet. This section describes what the next release adds. The build currently on the Chrome Web Store came before it, so if you installed Veritype from the store you will not find a reading summary in your panel — that is expected, and nothing is broken. It is documented here ahead of time so the behaviour is written down before it ships rather than after.
When it lands, the panel will keep a running summary under each report of what you have been reading — how many pieces, roughly where they sit, and which outlet comes up most.
The summary draws two marks on the same rule. The solid mark is you — the average of your recent reading. The hollow one is the article on screen, so you can see at a glance whether what you are reading now sits where you usually sit.
This is kept on your own computer by default and is never sent anywhere. If you turn on keeping your reading on your account, the same rows also sync to Veritype so they survive a reinstall and follow you between computers, and a fuller view opens up on the web with a trend over time and your full outlet breakdown. What is stored is the day, the lean, and the outlet — never the page address, the headline, or the text. It is off unless you switch it on, and deleting it deletes it everywhere. The details are in the privacy policy.
The panel
- Whole page or selection. Analyse the article you are on, or select a passage and analyse just that. Selection mode is useful for one quoted paragraph in a longer piece.
- Recent. The last few reports, so you can go back to one without re-running it. This is cleared when you close your browser — “Copy all” beside it gets them out first.
- Your daily allowance. The badge shows how many analyses you have left. When you reach the limit the panel tells you when it resets and what your options are, and the report already on screen stays put — hitting a limit does not take away what you were reading.
- Pinned actions and persona. In settings. Both change what you see and the order you see it in, never the analysis itself.
What Veritype cannot do
The honest section. These are the known weaknesses, not a disclaimer.
- It is not a fact-checker. It never verifies whether a claim is true. Accuracy and framing are different axes.
- It is weakest on what is not there. Bias by omission — the story not covered, the source never called — is the hardest kind to see from a single text, and it is the kind Veritype most often misses.
- It reads one piece at a time. A single article is a small sample of any outlet, any writer, or any argument. One reading is not a verdict on a publication.
- The AI-content read is a weak signal. See above. It is uncalibrated, and it should never be treated as proof of anything.
- Long pieces get truncated. Very long articles are cut before analysis, so the assessment reflects the part that fit.
- Some pages do not work. Paywalled articles, PDFs, and pages that are mostly interactive app rather than prose often have no article text to extract.
- It is a language model, and it varies. The same piece analysed twice can produce slightly different wording and a slightly different score. Where the mark sits on the rule is a reading, not a measurement.
- It is trained on, and best at, English-language political writing. Coverage from other traditions and languages is weaker ground.
Where your result came from
Results are cached and shared across everyone who analyses the same text, for up to 30 days. That is what keeps Veritype affordable, and it means a report can be accurate and several weeks old at once. Each report says whether it was freshly generated or came from that cache.
Every analysis uses the same model and the same instructions — including when you bring your own API key, where the request goes directly from your browser to the provider and never touches Veritype’s servers.