The short answer
Ask where the model runs, what is kept, what it is trained on, and whether you need an account. Then ask which of those answers you can check. A policy is a promise about how a company behaves; a model running on your own device is a property of how the thing is built. Both can be honest, and they are not the same kind of assurance.
1. Where does the model run?
Almost always on the company’s servers, which means everything you type is sent there. A few tools run a smaller model on your own device, inside the browser or an app, and then there is nothing to send. Everything below follows from this answer.
2. Is the conversation kept, and for how long?
Look for how long conversations are kept, whether deleting one deletes it everywhere (backups included), and who inside the company can read them. A retention policy can also be overridden: a court can order a company to preserve logs it would otherwise delete, and in 2025 one did, in copyright litigation that had nothing to do with the users whose logs were kept.
3. Is it used for training?
Many services use conversations to improve their models by default and offer a way to opt out. Check which applies to your plan, and whether opting out covers what you have already typed.
4. Do you need an account?
An account ties every conversation to a name or an email address. A service that needs none has nothing to attach your conversations to, whatever else it keeps.
5. What does “encrypted” mean here?
Nearly every site encrypts traffic on its way to the server. That protects it from the network in between, not from the service. A server that answers your question has to read it, so a claim that goes further deserves a careful look at who holds the key, and when.
6. Can you check any of it?
Most of the answers above are statements you have to take on trust. One kind you can test: if the model runs in your browser, open the developer tools’ Network tab and watch what is sent, or turn off your Wi-Fi and see whether it still answers.
How Zero Witness answers them
- In your browser: the model runs on your device. No request carries what you type, so there is nothing for us to keep, train on or hand over. You can check it both ways described above.
- On our hardware: your message is encrypted in your browser before it is sent, read in memory on hardware we own (never rented cloud), and not written to disk, and the model releases it when the conversation ends. Nothing is trained on. That is a promise about how we run our infrastructure, and we can’t prove it to you the way browser mode can be proved. We read your message to answer it, so it is not end-to-end encryption and we don’t call it that.
- No account, in either mode.
- What stays on your device: by default a copy of your conversations is kept in your own browser so you can come back to them. Nothing is kept on our side. You can switch that off, or remove everything, in Settings.
The page always shows which of the two you are in, and the choice is yours.
Related guides: How to run an AI model in your own browser · How to transcribe a recording without uploading it · How to turn a photo into a PDF without uploading it · Can you put client information into an AI chatbot?. All guides.