Known limits
what the checks cannot tell.

01How to read this page

The safety levels are OK, Caution and Danger. A check that could not run is listed as "not checked" and never counts as OK. The first eight limits concern the message, e-mail and screenshot checks; the rest concern links, media and claims.

021. A payment request with an unlabelled landline is Caution, not Danger

Some messages claim to be from a company, ask for a payment, and give a fixed-line, 08xx or 800 number without a label such as "Customer line" or "Infolinka". These stay at Caution, even when the message names the company.

• The reason: genuine bills often give their service number this way.
• A request to confirm, verify, unblock or authorise something, or to have a code or card details ready, still raises such a message to Danger.

032. A brand is never shown as "through its own channels" when one of its numbers cannot be verified

A genuine message that lists a service number the product has not verified keeps its brand claim at Caution. This also applies to a genuine message with a correct link to the company's own site.

043. Links to public forms, file shares and page builders are always Caution

This covers hosts where anyone can publish a page, a form or a payment link, for example Google Forms, Microsoft Forms, Dropbox shares, GitHub Pages, PayPal.me and Revolut.me. Legitimate forms and shares (a school survey, a colleague's photos) are Caution too.

054. A few genuine messages can be Danger when the AI reads them strictly

• A school or office form on a Google or Microsoft platform is Danger if it asks for money or sign-in details and the AI reads it as sent by Google or Microsoft.
• A friend's payment request is Danger if the AI reads it as pressure (for example "today, I'm leaving tomorrow").

065. Brands that are not on the built-in brand list are judged less

For a company that is not on the list, its links and bank accounts are not compared with its own. For example, an insurer that is not listed and asks for a payment to a foreign account gets Caution, not Danger. An "account abroad" note is added when the message is in Slovak or Czech. You can extend the list in server/src/lens/link/brands.json.

076. Verified service numbers cover six mobile operators and must be re-checked

• The numbers were read on each operator's own contact page on 12 November 2026. Re-read them from time to time, because operators change their lines. They are in server/src/lens/message/service-numbers.json, each with its source page.
• O2 Slovakia's numbers are not verified, because its website does not answer automated reads. A genuine O2 Slovakia bill shows "a mobile number as a company line, which could not be verified".

087. Phone numbers are read in Slovak, Czech and UK formats

National Slovak, Czech and UK formats are read, and any number in international format (+…). Numbers written in other countries' national formats are not read, so they do not count in the result.

098. An app download from a company's own listed domain is not flagged

An install link (for example an .apk file) on a domain the brand list names as the company's own is not counted as an install from elsewhere. An install from any other site in the company's name is Danger. Links to the official app stores (Google Play, App Store) are not Danger by themselves.

109. Links

• A brand-new fraud page may be on no list yet: a link with no warnings can still be a scam.
• Some country domains (for example .sk, .de, .eu, .io) publish no registration date, so their age is not known.
• Pages are read as text; what a page shows only after scripts run may not be seen.
• The page read always uses the main assistant model, also for Pro installs.

1110. Images, video and audio

• AI-made and face-swap scores are probabilities. A real photo can score high and an AI-made one low.
• Videos are checked as files up to 60 seconds; links to video platforms are not downloaded. Frames are sampled, so a short altered moment can be missed.
• The web lookup shows where an image appears, but gives no dates, so Truthbeam never says when an image first appeared.
• Metadata can be edited, so it never counts for more than Caution.
• Most files carry no Content Credentials; without trusted signers in the console, no file reaches "Proven origin".
• There is no check for AI-written text: detection on short or Slovak and Czech text is not reliable.

1211. Claims

• Fact-checkers cover only some claims, and coverage differs by language; Slovak and Czech ratings are mapped only for Demagog.
• A fact check counts only when it is about the same claim. A review worded very differently, or in another language, may be dropped; the check then says "No reliable source found" rather than giving a wrong "Rated false".
• "Mostly true" ratings count as disputed, so "Supported by sources" is never overstated.
• Claim finding needs the language model; with the assistant switched off, no claim is checked.