Someone sitting in the same café cannot read your AI chat. The connection between your phone and the service is encrypted, so what crosses the air is unreadable to anyone else on the network. The Federal Trade Commission puts it plainly in Are Public Wi-Fi Networks Safe? What You Need To Know: most sites now encrypt, so connecting through public Wi-Fi is usually safe, and the thing to check is the lock and the https in the address bar.
The risks that remain are different from the one people picture, and for AI chat specifically the largest one has nothing to do with the network you’re on.
What someone on the same network can actually see #
Not your messages, and not the model’s replies. What leaks is the outline:
- Which service you’re using. The hostname is typically visible during connection setup, and DNS lookups often are too, unless your device is using encrypted DNS. So the network can tell that you connected to a particular AI service.
- When and for how long. Session times, and the rhythm of a conversation.
- How much data moved. Message sizes are inferable from traffic volume, though not their content.
That’s a meaningful amount of metadata on a network run by an employer, a school or a hotel, and none of it is the text of your conversation.
The risks that are real #
A fake network. Anyone can name an access point “Airport_WiFi_Free”. Connect to it and the operator controls DNS and the captive portal, which is enough for convincing phishing pages. The FTC makes the related point that encryption doesn’t help if the site itself is fraudulent: your data reaches the scammer safely encrypted. Get the network name from staff rather than from the list.
Captive portals that ask for too much. A portal that wants an email address is ordinary. One that wants a password, an ID document, or asks you to install a certificate or profile is not. Back out.
Out-of-date devices. Unpatched operating systems and browsers are the gap that turns a hostile network into a compromised phone. Updates are the least glamorous item on this list and the most effective.
Auto-join. Phones reconnect to remembered network names automatically, which means a spoofed name can pull you in silently later. Use “forget this network” after one-off hotspots, and turn off auto-join for public ones.
The person behind you. Shoulder surfing is the genuinely underrated risk in a café or a plane. An AI chat is a wall of readable text, and what you type into it is often more personal than an email would be.
What a VPN does and doesn’t do #
A VPN encrypts everything between your device and the VPN server, which hides the metadata above from the local network and the network operator. That’s a real benefit on a hostile or nosy network.
It does not hide anything from the AI service you’re talking to, which is the other end of the conversation and decrypts your messages to answer them. It doesn’t prevent your chats from being stored against your account. And it moves your trust to the VPN provider, who now sees the traffic the café used to. A free VPN is a business, and you should know what it sells.
For AI chat, the provider is the bigger question #
Encryption in transit protects against the café. It doesn’t protect against retention, because the service must decrypt your message to respond. That’s the distinction covered in what end-to-end encryption means for AI chat: your conversation is readable at the provider by design, stored against your account, and kept under their policy.
Which is why the useful public Wi-Fi rule for AI is about content rather than connection. Don’t type things you’d mind being retained: credentials, ID numbers, client details, someone else’s medical information. Keeping personal data out of AI chatbots lists the categories worth holding back.
The version with no network traffic at all #
A model that runs on your phone removes the network from the question entirely. After the one-time model download, generating an answer involves no connection, so there’s nothing for a café network to observe, no session to correlate, and no stored copy at a provider.
Personal LLM works that way: you download an open model once, such as Qwen 3.5 or Gemma 4, and from then on the chats and any photos you attach stay on the device. There’s no account and no server behind it. You can verify the claim rather than take it: put the phone in airplane mode and keep chatting, which is the test described in how to check whether an AI app is really offline. The app is free on iOS and Android and shows ads while online, and those ads are the one thing that stops loading when you disconnect.
The honest limits still apply. A local app doesn’t protect you from someone reading your screen, from an unlocked phone, or from a device with malware on it.
A checklist for hotel, airport and café Wi-Fi #
- Confirm the network name with a member of staff.
- Treat any portal that wants a password, a document or a profile installation as hostile.
- Keep the OS, browser and apps updated. This is the one that matters most.
- Turn on two-factor authentication for accounts you sign into while travelling.
- Turn off auto-join for public networks, and forget them when you leave.
- Use your phone’s own cellular connection or personal hotspot for anything you’d rather not have observed at all.
- Angle the screen away from the room before opening a long chat.
One practical note about model downloads #
If you use an on-device AI app, hotel Wi-Fi is a good place to download a model. The files run from about 0.8 GB to just over 6 GB, plus a separate 195 MB to 1 GB if you add image support, and pulling that over a metered cellular plan is the kind of thing that eats an allowance. Downloads that can be paused and resumed help when hotel Wi-Fi does what hotel Wi-Fi does. How much mobile data AI chat uses covers the ongoing cost of the cloud alternative.
Frequently asked questions #
Can someone on café Wi-Fi read my ChatGPT conversation? #
No. The traffic is encrypted between your device and the service, so another person on the network sees that a connection happened and roughly how much data moved, not the words. The realistic ways someone reads your chat are over your shoulder or on your unlocked phone.
Do I need a VPN to use AI chat on public Wi-Fi? #
Not for confidentiality of the messages themselves, which encryption already handles. A VPN is useful if you don’t want the network operator to know which services you use, or if you’re on a network you actively distrust. It gives you nothing against the AI provider’s own storage of your chats.
Is hotel Wi-Fi worse than café Wi-Fi? #
The technical risks are the same. Hotel networks are more often heavily monitored and more often serve intrusive portals, and shared conference Wi-Fi tends to be older. Apply the same checks and use cellular for anything sensitive.
Does an offline AI app use public Wi-Fi at all? #
After the initial model download, a genuinely on-device app needs no connection to answer. Ad-supported free apps still fetch ads while online, which is network traffic even though your conversation isn’t. Airplane mode is the quickest way to see which parts of an app actually need the internet.
Should I avoid signing into accounts on public Wi-Fi? #
Signing in is normally fine on a modern, updated device, since the session is encrypted. The precautions worth taking are the boring ones: a unique password, two-factor authentication, and not entering credentials on a page a captive portal pushed at you.