Do AI assistants train on your messages? Most will not say

Of the 90 AI assistants catalogued here, 29 give no answer to the most basic question you can ask one: does it train on your conversations. Eighteen of those wrote a privacy policy and left the question out of it.

That is not a gap in the research. Every policy in the table below was fetched, read and dated, and the 11 products with nothing to read are counted separately.

What the policies actually say

Fourteen state plainly that they do not. That is the good half of this and it is worth saying before anything else.

The rows worth reading are the 6 that say they do, the 6 that do it unless you find the switch, and the 33 that answer a narrower question than the one asked. Those last are the ones to slow down on, because each is true as written.

Otty writes that it does not use personal data to train general-purpose AI models, which leaves its own models unaddressed. Flip explains that it sends your balances, transactions and messages to OpenAI and Anthropic under terms that prohibit those providers from training on it, and says nothing about Flip. Both sentences are accurate. Neither is an answer.

The ones that wrote a policy and skipped the question

This is the finding, and it took reading 79 documents to see it.

A company with no privacy policy has told you nothing, and you know where you stand. A company that publishes a policy has sat down with counsel, worked through what it collects, named its processors, set out your rights under GDPR and four American state statutes, and then not mentioned whether the conversations are training data. Eighteen listings did that.

Zo Computer is the clearest case. It runs a computer for you, holding your files. Its policy lists the categories of personal data it collects: an email address, trackers, usage data. The files are not in the list. Neither is training.

Some of this is a template. The state-by-state sections in several of these are the same boilerplate with the product name changed, and a template written for a shopping site does not have a paragraph about model training in it because a shopping site does not have this problem. That explains the omission. It does not answer it.

The one that decides by where you signed up

Arlo publishes the most detailed answer in the set, and it is not the one anybody would guess.

Its policy states that a new adult account created from a server-verified United States location starts participating by default, while an account created in the EEA, the UK or Switzerland has to opt in from settings. Same product, same price, and consent that turns on where your signup resolved.

What it participates in is unusual too. Arlo does not train on raw messages, and says so at length. It builds deidentified structural records instead, and its policy says it may license approved datasets of them to outside AI developers, who use them for their own model training as independent recipients rather than as Arlo's subprocessors. It is the only clause of its kind recorded here.

Read the whole clause before deciding what you think of it. It is also the most carefully written thing on this subject in the set, and a company that says this much has thought about it. The 18 have published nothing that shows they have.

Nobody is selling the opposite

Here is the part that makes this a market story rather than a compliance one.

Reading all 90 taglines on 9 September 2026, one product makes a data-privacy claim in the sentence it uses to describe itself: Bo, which says its AI provider is held to zero retention. One, out of 90. Every other product leads with what it can do.

Eighteen publish a security contact, which is the cheapest possible signal that somebody is responsible for this, and it costs a line in a text file.

Meanwhile 14 of them state they do not train on you, and almost none says so anywhere a buyer would see it. The position is empty because fourteen companies are standing on it without saying a word.

One of the six that state they do train is Kachi, which writes that it trains and improves its own models on conversations and outcomes. It is quoted on its listing like every other, and of the six it is the plainest, which is worth something on its own.

What is still not known

The training question is the one people ask. Two others matter more and are answered less.

Fourteen of the 79 readable policies address people who never signed up: the friend in the group thread, the colleague on the calendar invite, the person on the other end of the email your agent answered. Every one of these products processes those people's messages, and fourteen of them have written down what happens to them.

Retention is thinner still. Sixty-five listings state what happens to your data when you cancel. The rest leave it to a general sentence about keeping data as long as necessary, which is a legal formula rather than an answer.

What to watch

Three things, and each is checkable rather than a prediction.

The first is whether any of the 18 adds a sentence. It costs nothing, it is the highest-value line any of them could publish, and the board that ranks disclosure will move the day one does.

The second is whether anybody starts selling this. The first product in this category to put a training answer in its own headline takes a position the rest of this directory has left empty, and it will be visible from the homepage rather than from a policy page.

Last is whether consent by geography spreads. One product decides your default by where your signup resolved. If a second does it, a norm is forming, and it will have formed without anybody discussing it.

Training is one of two questions a policy can answer about what a product holds. The other is what it is given when you connect an account, and 38 of 90 listings answer that one: one takes the password, two never see it, and two never show it to their own model.