In short
We use AI for a few helper features. Account suspensions and hiding content are decided by people, not by AI.
Data-saving mode
We currently use Gemini through a free access tier. Google applies different data terms to free access than to paid access. That is why we currently send as little as possible to the AI: new posts are not checked automatically, no AI checks uploaded images, and we do not translate chat messages.
Which features use AI
- FriendsBot: a chat partner whose replies are written by an AI model. The model runs on a computer of the operator, not at Google. It is pinned at the top of your chat list. To answer, your latest messages in the FriendsBot chat and its replies to them are sent via our server to this computer. The replies can be wrong.
- Web pages in the FriendsBot chat: if you send FriendsBot a web address, our server fetches exactly that one public page, without cookies and without logging in. The page’s rules for automated access (robots.txt) are respected. A text excerpt of at most about 6,000 characters is passed to the AI model. The page sees the address of our server, not yours. Internal addresses, links containing login data and non-public Friendstime pages are not fetched. At most 10 pages per hour.
- Translation: when you tap, Gemini translates a post or comment. In data-saving mode we do not translate chat messages.
- Reply suggestions: currently switched off.
- Hashtag and caption suggestions: when you tap while writing a post. You decide whether to use them.
- Fact check: in data-saving mode our server does not send new posts to Gemini automatically. The team can check individual public posts. A label saying a post may be false only appears once a team member has confirmed it.
- AI image check: switched off in data-saving mode. Looking for provenance data and removing the location keep running on our server without AI.
What data goes to the AI
Google Gemini only receives what the feature needs: the text or image you hand over, and for the fact check the text of the public post. More in the privacy policy.
Your messages to FriendsBot and its replies are not sent to Google. The computer that runs the model collects the requests from our server itself and sends the reply back. The collecting program does not store the messages on that computer; they are stored with us like other chat messages. A push notification about a new reply only contains a fixed sentence, not the text.
We do not store the fetched page text. As before, only the chat is stored, meaning your message with the link and FriendsBot’s reply. For the limit, we only note the time and result of each fetch, without the address, and delete this after 7 days.
Labelling AI-generated images and videos
When you upload an image to a post, we read the provenance data in the file before removing the other metadata such as the location. We look for the IPTC “Digital Source Type” entry, Content Credentials (C2PA) and markers of known image generators. Only the result is stored.
For videos in posts we look for the same data in the first and last megabytes of the file.
If something is found, the image shows “AI-generated” or “Edited with AI”. On the website the image also carries this information in machine-readable form in the data-ai-generated attribute. You can also mark an image as AI-generated yourself when posting.
If there is no such data, that does not mean an image is real. Many programs do not write provenance data, and some apps remove it.
Your duty with AI content
If you created an image, video or audio with AI, or changed it so that it could look real, mark it as AI-generated when posting. Not allowed: AI fakes that show real people saying or doing things they never said or did.
Human oversight
- No AI result suspends an account or hides content.
- A negative fact-check result only becomes visible once a team member has confirmed it.
- Appeals are decided by people.
Turning AI features off
In the app, the first AI feature asks whether you want to use it. If you decline, translation and reply, hashtag and caption suggestions stay off until you turn them on again in the settings.
General information on the AI Act is on the EU AI Act page.