Is an AI assistant reading group chats a privacy risk?

What happened with Meta Muse and your friends' data?

Meta's new AI agent, Muse, has turned a quiet privacy question into a loud one: when you plug an assistant into your digital life, whose data are you really handing over? Muse launched in the US in September 2026. It can connect to email, calendars, contacts and messages, then act for you. Independent researchers extracted its internal instructions and found that it builds a profile page for every person in a user's life, from partners and parents to coworkers, and refreshes those pages every hour. Coverage this week summed it up: you're asked to give up your own data, and also your friends' data and that of almost anyone you've ever been in touch with.

Meta says Muse is opt-in, that users choose what to connect, and that connections can be limited or removed. Privacy advocates aren't satisfied, and their main point is hard to argue with. The friends, relatives and colleagues pulled into those profiles have no real way to opt out. A few incidents have sharpened the debate. A tech columnist said Muse synced messages he believed he had declined to share, which Meta disputes. A reviewer watched it give his home address to a stranger, using a permission he had technically granted without realizing what it covered. Another investigation found that asking Muse to 'forget' something doesn't necessarily delete the original message where you shared it.

You might never install Muse, but the same question applies in WhatsApp. Any assistant you add to a group reads your neighbors', teammates' and friends' messages. Any assistant that messages people for you is talking to someone who never chose it. Agent Sonic does both, so we'd be dodging if we didn't answer the question directly. Here's our view. An AI reading other people's messages isn't the problem in itself. A colleague who takes meeting notes reads everyone's words too. The problem is an assistant that reads messages nobody knew it could see, uses them for purposes nobody agreed to, and keeps profiles nobody can delete. The rest of this guide is a practical consent check to run before you add any assistant to your chats.

What does an AI assistant actually see in your WhatsApp chats?

An AI assistant in WhatsApp sees only what you put in front of it. That usually falls into three buckets: the groups it's a member of, the people it messages for you, and the files you forward to it. That's a smaller footprint than an agent connected to your email, contacts and desktop, but small isn't zero. Each bucket involves other people's information, and each raises a different consent question. It helps to look at them one at a time, because the right ground rules depend on which one you're using. A recap of the building group is very different from forwarding a client's signed contract, even though both feel like 'just asking the AI.'

Groups are the most visible case. When you add an assistant like Sonic to a WhatsApp group, it appears in the member list like any other participant, so nobody has to wonder whether it's there. It reads the messages in that group so it can do group jobs: catch-up recaps, polls, shared lists, expense splits and automatic reminders. Send 'Sonic, summarize what I missed in the building group' and it works from that group's conversation. It doesn't go through your other groups or your one-to-one chats along the way. Compare that with giving an agent your entire inbox, where every thread you've ever had becomes material to work with, whether or not it has anything to do with the task.

Outreach and documents are the less obvious cases. With outreach, the other person knows exactly what's happening. Sonic writes to them from its own number, introduces itself as your assistant, asks your question and reports the answer back to you, so they know who they're dealing with. Documents are trickier, because a forwarded PDF often contains far more than you meant to share. A lease includes a tenant's ID number. An invoice shows a client's home address. A shift spreadsheet lists everyone's phone numbers. The assistant reads the whole file, not just the line you cared about. There's nothing sinister about that; it's how reading works. It does mean the decision to forward is yours, and it's worth a moment's thought.

Is it ethical to use AI on a WhatsApp group without telling anyone?

No. If an assistant reads a group's messages, the group deserves to know, and a quick heads-up is the minimum. Quietly adding a bot that summarizes everyone's messages is like recording a dinner party with a phone hidden under the table. Maybe nothing bad happens, but you've made a choice for everyone there without asking them. People write differently when they know software is reading along, and that's exactly why they should be told. The good news is that telling them costs you almost nothing. In most groups, the reaction to 'I've added an assistant to handle reminders' is relief, not outrage. Nobody enjoys being the one who chases everyone about rent.

Context matters too. A visible assistant doing admin in a group of friends is about as controversial as the friend who keeps the shared shopping list. A business group is different. If the chat includes clients, customers or staff, you may have obligations under local data protection law. That's especially true in Europe, where the rules generally get stricter once you're using data for more than purely personal reasons. We're not lawyers, so check what applies where you work. One sensible default holds almost everywhere: say what the assistant does, say what it doesn't do, and make it easy for people to object. If someone in a 40-member residents' group says they're uncomfortable, take that seriously and talk it through.

Telling the group doesn't need a policy document. A message like this does the job: 'Hi all, I've added Sonic, an AI assistant, to handle reminders, polls and weekly recaps. It only sees this group. Let me know if you'd rather not have it here and I'll take it out.' That covers what it is, what it does, where it works, and how to get it removed. Then keep your word. If the group agreed to rent reminders, don't start asking for opinions on individual members. Once everyone's comfortable, the assistant can handle the reminders for good. Here's how to keep a WhatsApp group on track without nagging, with automatic reminders for standups, timesheets, rent and practice schedules.

What ground rules keep an AI assistant fair to everyone in your chats?

Four rules cover most situations. Tell the group. Keep the assistant to admin tasks. Never forward whole private chats. Choose an assistant that only works in the chats where it was added. None of them needs technical knowledge, and all of them would have softened the backlash against agents that quietly profile everyone in a user's address book. Think of them as WhatsApp manners. You wouldn't photocopy a friend's letter and pass it around the office, and you shouldn't paste a friend's private messages into a tool they've never heard of. The rules below are simple enough to follow on a busy Tuesday, which is usually when these decisions get made.

The admin rule matters most. These are fine: 'Poll the team: Thursday or Friday for the end-of-quarter dinner?', 'Remind everyone on the 28th that rent is due', 'Split the cabin bill between the six of us'. These aren't: 'Who in this group has gone quiet lately, and why?' or 'What does Maya really think of me, based on her messages?' The first set organizes people. The second set analyzes them, which is exactly what critics flagged in agents that build files on users' contacts. One point that may surprise you: the recap, not outreach, is the feature that needs the most care. A summary of decisions is harmless. A summary of who said what about whom is just gossip, nicely formatted.

The last two rules are about scope. Forward only what the task needs: the invoice, not the whole thread with the supplier. The contract, not the side chat where your partner complained about the landlord. If a document includes someone's ID number or medical details and the task doesn't need them, cover them up first. Then look closely at what an assistant asks for when you sign up. A tool that only knows what you send it, plus what's said in groups it was openly added to, is a very different bet from one that wants your mailbox, contacts and desktop on day one. For a fuller list of checks, see our plain-English guide to what an AI assistant can read and do for you before you trust it with anything.

How should an AI assistant handle messaging people on your behalf?

It should message people openly, only after you approve, and give the person on the other end an easy way to opt out. Someone who gets a message from your assistant never signed up for it. They deserve to know who's writing, why, and how to make it stop. That's the standard we'd hold any tool to, ours included. The warning signs are real: agents have already sent messages nobody approved. That's why it's worth reading up on whether an AI agent messaging on your behalf is safe before you let one write to anyone. Good outreach is boring: one clear question, at most one polite follow-up, and a straight answer back to you.

Here's how Sonic's outreach tasks work, limits included. You say who to message and what to find out: 'Ask Dana when the invoice will be ready and let me know.' Before anything goes out, Sonic shows you the exact first message for each person. Reply 'yes' to send it, 'no' to drop it, or ask for changes like 'more formal' or 'shorter, no emoji'. Those notes apply to the whole conversation, not just the first message. Sonic writes from its own number, always introduces itself as your assistant, and never pretends to be you. Anyone who replies 'stop' gets a short apology and is never messaged again by anyone using Sonic. The step-by-step outreach tutorial walks you through it.

The limits are deliberately tight. Outreach is on the Pro plan and in Beta, so it may be paused from time to time. You can run one task at a time and message up to three people a day. First messages only go out between 08:00 and 21:00 in the recipient's time zone, and each person is contacted at most once a day. If someone doesn't reply, nobody chases them; after 24 hours, Sonic simply tells you. Group chats aren't included, only one-to-one chats. Requests to pressure, deceive or spam anyone are refused. You might see those limits as missing features. We see them as the whole point, because they protect people who never chose Sonic. If that's the kind of assistant you want, get access to Sonic and start with a single group recap.

Frequently asked questions