What Predicts does with your messages
It answers the training question with a qualification worth reading twice.
Your message passes through 4 sets of hands before it is answered, and Predicts names 2.
Read from its published policy on 2026-09-10, and quoted rather than paraphrased.
Whose hands it passes through
Somebody carries the message, somebody runs the model that reads it, and where the agent uses a website on your behalf somebody operates that browser. A blank below is not an absence of a company, it is a company nobody has named.
- iMessage, WhatsApp, Telegram
- Loki Labs
- Not named
- Anthropic, Cloudflare
- Not named
- Poland and the EU, with GDPR and UK GDPR legal bases and the Polish supervisory authority named, alongside California CCPA and CPRA rights.
In its own words
“We do not sell personal information, and we do not use your messages to train third-party AI models.”
What it does not answer
Each of these is a question some policies in this directory answer. The count is how many of the 56 we have read close it.
- the people in your threads who never signed up
- where to report a hole
Silence is not a denial and it is not a promise. A document that does not address the people in your group thread has not agreed to anything about them.
How old this document is
Predicts last revised it on 2026-07-01, which was 71 days before we read it. 52 of the 56 policies here carry a revision date at all; the rest do not say when they were last changed, so there is no way to tell whether they describe the product as it ships today.
We started keeping a dated copy of this document on 2026-09-10. A policy that is quietly rewritten leaves nothing behind, so from here any change to it is recorded with the date it appeared.
Back to the Predicts listing, every policy read the same way, or why 22 of 65 give no answer at all.