
A private AI assistant for WhatsApp that doesn't train on your messages

Your WhatsApp holds the things you'd never paste into a public chatbot: client prices, salary talks, a family health scare, the bank details a vendor just sent. You'd like an AI to help you keep up with it all, but the moment you look at one, the questions start. Who else can read this? Will my messages end up training someone's model? Can I keep some chats out of it?
A private AI assistant for WhatsApp should answer all of those in writing: it reads only the chats you allow, keeps your data apart from everyone else's, never uses it for training, and lets you cut it off in one step. Perisclaw is built that way, and the checklist further down works for judging any assistant, including the ones built into the apps you already use.
What privacy means for an AI that reads your chats
WhatsApp's end-to-end encryption protects messages on their way between devices. Any AI that helps with your chats has to read them, so the real questions are about what happens once it does:
Where your data lives: on whose servers, and whether it sits alongside other customers' data.
Who can read it: the company's staff, human reviewers, or only software acting for you.
What it's used for: running your assistant, or also training AI models.
What it can see: every chat, or only the ones you choose.
How you leave: whether you can revoke access and delete your data, and whether it's really deleted.
No assistant that reads your messages can be "zero access", so look for one that answers each of these specifically, in writing.
How Perisclaw keeps your WhatsApp private
Perisclaw runs on your own number as a linked device, the same way WhatsApp Web does. Like WhatsApp Web, it has to decrypt your messages to show you anything; the difference is that it works on them inside a private environment set up for you alone. You link it from WhatsApp's own Linked devices screen, and you can log it out from the same place.
Linking Perisclaw to WhatsApp with a pairing code on iPhone
Here's what its privacy and security documentation commits to.
Isolated: your assistant runs in its own sealed environment in secure cloud infrastructure. Your data isn't pooled with anyone else's, and the only thing that touches it is your own assistant, acting on what you ask. No one on the Perisclaw team reads it.
Encrypted with your own keys: your data and logs are encrypted with keys unique to you, which are derived when needed instead of being stored, so there's no key file sitting around to steal.
No training: your data is never used to train AI models, Perisclaw's or anyone else's. Perisclaw explicitly opts out of model training on customer data.
No telemetry: Perisclaw keeps no activity logs about how you use it. The one exception is feedback you choose to send, which is logged where you can see it.
Audited by someone else: Perisclaw is ISO 27001 certified and goes through regular vulnerability testing.
Guarded against hidden instructions: a message, email or web page can hide text meant to trick an AI into acting against you, so Perisclaw runs your assistant with safeguards against it and vets every skill and connection first.
On a team plan, nobody on the team, including the owner, can read your chats or use your assistant.
Keep specific chats and messages out of it
Some chats are nobody's business, including your assistant's. Perisclaw gives you four ways to keep them out:
Block a chat: in Settings, then Permissions, search for the chat under Blocked chats and choose Block, or just say "never read" followed by the chat's name in your self-chat. Perisclaw can't read, search, summarise or quote a blocked chat at all.
Use WhatsApp's Chat Lock: lock a chat in WhatsApp and Perisclaw blocks it automatically. While it's locked, Perisclaw can't read it, search it, write to it or put its contents in any log. Remove the lock and access comes back.
Hide one message with /ignore: put /ignore as a separate word in a message you send. Perisclaw reacts with the see-no-evil monkey and won't act on that message, return it in later searches or write it to logs. The message itself stays in WhatsApp.
Restricted chat: WhatsApp's newer Restricted chat setting, still rolling out as of November 2026, blocks linked devices for that chat. That includes Perisclaw, so it won't see a restricted chat at all.
Control what it does, not only what it sees
By default, when you ask Perisclaw to message someone, it writes a draft in your self-chat and sends only when you approve it. Outside your self-chat it does nothing unless you call it with /claw or set up an Automation or Mission for that chat.

A few other controls matter for privacy:
It keeps your context to itself: when it acts in someone else's chat, it treats them as an outsider and won't reveal your other conversations, calendar, email, files or notes, even if they ask. If someone probes, it tells you.
Hard rules: write rules it checks before every action, such as "never share my address with anyone".
Pause everything: say "pause everything" in your self-chat, or set Message other chats to Never, and it sends nothing to anyone until you resume.
How to revoke access and delete your data
Disconnect an app or AI tool: remove Gmail, Calendar, ChatGPT, Claude or any other connection in the web app, and its access stops straight away.
Log out the linked device: in WhatsApp, open Linked devices and log Perisclaw out, the same as any WhatsApp Web session.
Delete your account: go to Settings, then Account, then Danger zone, type DELETE and confirm. This removes your numbers and their data and deletes your login.
How Meta AI, ChatGPT, Claude and Gemini handle your messages
If you're weighing a general AI app for the same job, here is what each company's own help pages say about training, as of November 2026. Business and work plans often follow different rules, so check yours.
Assistant | Used to train models? | How to keep chats out |
|---|---|---|
Meta AI in WhatsApp | Meta can read only messages you send it, @mention it in or share with it (plus new messages in a group where it's been added as a member), and may use those prompts, messages and your feedback to improve its AI | Don't share sensitive chats with it; /reset-ai deletes Meta AI's copy of a chat |
ChatGPT (Free, Plus, Pro) | Yes, by default | Turn off "Improve the model for everyone" in Settings, Data controls. Rating a reply with thumbs up or down can still let OpenAI use that conversation |
ChatGPT Business, Enterprise, Edu | No, by default | Nothing to change |
Claude (Free, Pro, Max) | Only if you allow it with the Model Improvement setting | Leave Model Improvement off; incognito chats are never used |
Gemini app (personal Google account) | Yes while Keep Activity is on, which is the default; a subset of chats is read by human reviewers and kept up to three years | Turn off Keep Activity; chats are then kept for 72 hours to answer you |
Perisclaw | Never, for any plan | Nothing to change |
Training isn't the only difference. These apps see only what you paste or forward, and that copy then lives in their system under their rules. An assistant on your own number reads your chats where they are, so what matters is what it's allowed to see and how it's isolated.
How to choose a private AI assistant for WhatsApp: a checklist
Use this whether or not you pick Perisclaw. If a company can't answer a question in writing, treat that as the answer.
How does it connect? A bot number only sees what you forward to it; a linked device on your number can see what you see, so it needs real controls; a Business API bot is built for messaging customers, not your own inbox.
Does it say, in writing, that your messages aren't used for training? Look for a statement that covers the AI providers behind it too, not only the company's own models, and check whether it's the default or something you have to switch off.
Who can read your data? Staff, contractors and human reviewers are all different answers. "Nobody on our team reads it" is the one you want.
Is your data kept apart from other customers'? Ask whether each account runs separately and how data and logs are encrypted.
Does it collect usage logs? Fewer logs means less to leak.
Can you keep chats out? You should be able to block a single chat, and ideally respect WhatsApp's own Chat Lock.
Does it ask before it sends? Drafts you approve should be the default, with a way to pause all sending at once.
How does it handle hidden instructions? Anyone can message you, so ask what stops a stranger's message from steering the AI.
Has anyone checked? An independent certification such as ISO 27001 or SOC 2 means someone outside the company tested the controls.
Can you leave cleanly? Revoking access and deleting your data should take minutes, without emailing support.
What does it risk for your number? Any tool linked to your personal WhatsApp should be honest about account risk.
On that last point, here's ours: Perisclaw connects as a linked device, which WhatsApp didn't build for outside tools, and WhatsApp's help centre warns that linking unofficial apps can lead to a ban. Perisclaw is built for normal personal use, reading your chats and sending replies you approve, and it stops and checks with you after 15 chats if you try to send the same message to many people, because bulk sending is what WhatsApp's spam rules target. Nobody outside WhatsApp can promise zero risk, so we don't.
For a fuller look at what Meta AI can and can't see, read Meta AI in WhatsApp vs Perisclaw. For everything else Perisclaw does, start with what Perisclaw is, or read the short version on our security page.
FAQ
What is the best private AI assistant that doesn't train on my messages?
Look for one that states in writing that your data isn't used for training, keeps each account isolated, and lets you block chats. Perisclaw never uses customer data to train AI models, runs each assistant in its own encrypted environment, and is ISO 27001 certified. Claude's consumer apps also don't train on your chats unless you allow it.
Can Perisclaw's team read my WhatsApp chats?
No. Your data stays in your own isolated environment, and no one on the Perisclaw team reads it. If something needs debugging, your own assistant does it on request.
Are my WhatsApp chats still end-to-end encrypted with an AI assistant?
Your messages stay end-to-end encrypted on their way between devices. An assistant on your own number, including Perisclaw, reads them as one of your linked devices, the way WhatsApp Web does. Perisclaw then stores your data encrypted with keys unique to you.
Can I use an AI assistant on WhatsApp but keep some chats private?
Yes. In Perisclaw, block any chat in Settings or say "never read" followed by its name, lock a chat with WhatsApp's Chat Lock, or add /ignore to a single message.
Does Meta AI train on my WhatsApp messages?
Meta can only read messages you send to Meta AI, @mention it in, or share with it, plus new messages in any group where someone has added it as a member, and it may use those to improve its AI. Your other chats stay end-to-end encrypted and Meta AI can't see them.
An assistant that reads your WhatsApp and keeps it to itself. Start your 7-day free trial →
Try Perisclaw free
Your personal AI assistant on WhatsApp. It reads the chats you allow and gets things done for you. 7-day free trial, no card needed.
Start for free

