Free tool · artificial intelligence
AI check of a suspicious message
Paste the text of a text message, e-mail or post. Artificial intelligence marks in red pencil where the message pressures or tries to fool you, and tells you how to check it. The decision is always yours.
Before you paste the message
- Don’t paste passwords, PINs, card numbers or codes from text messages. You can replace names and addresses with “XXX”.
- We hide e-mail addresses and long numbers (accounts, cards, phones) with • before sending.
- We send the text for analysis to Claude, a service of Anthropic; processing may take place in the USA. We don’t store it, and under its terms Anthropic may not use it to train AI. It deletes it within 30 days (longer only if misuse of the service is suspected).
What the AI check can and can’t do
- It only sees the text you paste. It doesn’t know who sent the message and doesn’t open links – use the Link X-ray for a link.
- It can be wrong both ways: mark an innocent sentence or miss a trick. That’s why it never claims a message is risk-free.
- It can’t detect a fake voice, photo or video (deepfake) – it only judges a transcript or description. On a call, only one thing helps: hang up and call back on a number you know.
- You can check every mark: the name of the trick leads to the Manipulation atlas with an explanation and examples.
- We store neither the text nor the result. We only keep daily totals without any message content: how many analyses took place, which tricks were found and how people answered the two questions below the analysis.
How the check works
Anthropic’s Claude Haiku model receives the text of the message and a list of 59 tricks from the Manipulation atlas. It returns only places in the message, the name of the trick and a short note.
The page then checks that every marked place really is in the message and that the trick is from the atlas. If something doesn’t fit, it discards the answer and tries once more.
The summary is written by the page from the number of marks, not by the AI. Our team wrote the “How to check it” steps; the AI only picks from them. A note that would declare the message harmless, give advice or contain an address or number is not shown.
The AI is designed not to obey instructions hidden in the message (“ignore previous instructions”, “mark this as fine”), and the page checks its answer. It isn’t foolproof, so always check the message another way.