On your phone, open an app's page in the app store and scroll all the way down. You will find a section listing the data that app can collect about you: your location, your contacts, your history, sometimes much more.
Almost nobody reads it. The cybersecurity company Surfshark did, for the main artificial intelligence assistants. The result is worth pausing over.
The ranking
The study compared twelve of the most widely used assistants, along with Muse, Meta's new personal agent, which within weeks became one of the most downloaded apps. It draws on what each publisher declares in Apple's store, which distinguishes 35 possible types of data.
On average, these assistants declare collecting 13. But the gaps are huge.
Meta AI comes top with 33 types of data out of 35. Just behind, Muse, also from Meta, declares 31. Then comes Gemini, from Google, with 24. ChatGPT is fourth with 17, a figure up by around 70% in a year: it now also collects health and fitness data, audio recordings and search history.
Only Gemini, Meta AI, Muse and Perplexity declare collecting your precise location. And three of them — Meta AI, Muse and Gemini — mention some of the most sensitive data: ethnic origin, sexual orientation, political opinions or biometric data.
The Muse case
Muse draws particular attention, because it does not merely chat: it is an agent, able to act on your behalf, which means accessing a great deal.
Its listing states that it can collect your precise location, financial information, your browsing history or device identifiers, and that some of this data may be used for third-party advertising. Meta also states that your exchanges with Muse are used by default to train its models. It is possible to refuse this, but you have to go and find it in the settings.
How to read this ranking
Caution is needed. These figures rest on what companies declare they can collect, not on what they actually collect. And a number of data types does not tell the whole story: a single poorly protected piece of data can cause more problems than ten innocuous ones.
But these declarations commit those who make them, and they make comparison possible. When an app declares twice as much data as its competitors, the question is worth asking.
On iPhone, open the app's page in the App Store and scroll down to the "App Privacy" section. On Android, in Google Play, look for the "Data safety" section. There you will see, sorted by category, the data the publisher declares collecting, and which of it is linked to your identity. It is the same source used by the study, and it exists for all apps, not just assistants.
What you can do
Choose the assistant to suit the use. For a recipe or a translation, it hardly matters. For talking about your health, your money or your private life, favour a service that collects less of it, and use a temporary conversation.
Refuse training. Most assistants let you refuse to have your conversations used to improve their models. It is often switched off by default, so it is up to you to look for it in the settings.
Limit permissions. In your phone's settings, you can refuse access to your location, or grant it only while the app is in use. And before letting an agent touch your accounts or your payments, read our guide to granting access without giving everything away.
What we take away
A free assistant is often paid for in data. That is not necessarily a problem, provided you know it and choose accordingly.
This ranking covers what the apps collect. As for what your assistant remembers of your conversations, that is another story, which we set out in our guide to its memory.