Translation, voice and AI

Hello World translates what you write and what you say. This is what that actually involves.

Where the processing happens

On our own servers. The models that translate your messages, transcribe your voice notes and speak them aloud run on machines we operate. No feature of Hello World sends your messages to an outside AI company, and there is no arrangement under which anybody outside Hello World receives your content to train a model.

Two things outside Hello World can receive parts of your messages. Apple's notification service carries notifications to your phone, including the start of a text message unless the person receiving it has turned message content in notifications off. And when we investigate a problem, or review something that has been reported to us, if the person whose data it is has allowed it, third-party AI services under contract with us — part of Hello World AI Tools, some of them in the United States — can process your data stored on our server, including messages, and they receive what they process and handle it under their own terms. The app asks you when you sign up, and the switch is in Settings › Privacy › Fixing problems; if you say no, or are under 16, your data is left out — including when somebody reports what you wrote, which we then read ourselves. (The rest of Hello World AI Tools runs on our own machines and is not outside Hello World at all.) We commit that Hello World AI Tools do not show the content of messages to any person at Hello World, ourselves included, apart from the two exceptions below. We do not name those services here, because the ones we use change; write to support@helloworldmessenger.com and we will tell you which they are today. The privacy policy describes both in full.

We do not train models on your messages.

What that means for privacy, stated honestly

Translating a message requires reading it. Because the translation happens on our servers rather than on your phone, your messages are not end-to-end encrypted: they are encrypted on the way from your phone to our translation server, and again from that server to the phones of the people you write to, but the server decrypts them, stores them and processes them.

We would rather say that plainly than let the word “private” do work it has not earned. What we can tell you is what we do with that access:

• No person at Hello World reads your private messages. To find and fix problems, and to help us go through what is reported to us, we use Hello World AI Tools, which can process the messages of people who allowed it; we commit that they do not show their content to anyone at Hello World, ourselves included. There are two exceptions: we review a conversation or flash that is reported to us, to keep people safe; and we see what someone chooses to send us themselves, such as a screenshot attached to feedback, which can show other people’s messages.
• There is no automated scanning of message content for moderation or advertising. Our own software does read messages to do its work: to translate them, detect their language, voice a translation and show notification previews. Separately, the suggestion cards use fixed rules on our own server to look through recent messages for a plan, to offer a meeting (tontine cards are switched off).
• We do not sell your data, and there is no advertising network in the app.
• We can reach the server in order to run and fix the service, and that access is bound by the commitment above. That access is what makes a bug fixable, and it is the same access anyone running a messenger that is not end-to-end encrypted has.

Calls are different. Call audio and video go directly between the two phones when the network allows it, and through our relay when it does not. Call media is encrypted between the participants, and calls are not translated — so there is no point at which call audio is processed for content.

Translating text

When you send a message, it is translated for each person who reads it in a different language. Each chat remembers the language you write in and the language you read in.

The translation is produced by a machine translation model running on our servers. Translations are cached so the same sentence is not translated twice, and the original is always kept — you can see what was actually written.

Machine translation gets things wrong. It is usually good and it is never certain, and Hello World shows you how confident it is rather than presenting every translation as fact. For anything that matters — medical, legal, money — do not rely on it alone.

Voice notes

A voice note goes through three steps:

1. What you said becomes text. On iPhone this happens on your phone: Hello World requires on-device speech recognition and refuses rather than sending your audio to Apple.
2. The text is translated, on our servers, as above.
3. The translation is read aloud when the person receiving it plays it, by their own iPhone’s built-in voices.

The recording itself is uploaded to our servers so it can be delivered to the person you sent it to — they hear your actual voice as well as the translation.

Spoken playback exists only for the languages the phone has voices for . In the others the translation arrives as text.

My Voice

My Voice is not available yet. Our server does not generate voices at the moment, so the app says so before asking you for any recording, no sample is recorded and nobody’s voice is modelled. This is how it will work once it is switched on.

This is optional and it is off until you set it up.

If you enroll, you record a few samples of your own voice. From them we build a voice model, and your translated voice notes can then be spoken in a voice resembling yours rather than a stock one — so that a message arriving in a language you do not speak still sounds like it came from you.

• The samples and the model stay on our servers and are used only to speak your own translated messages. They are not shared, not sold, and not used to build anything else.
• Audio generated from your voice carries an inaudible watermark identifying it as synthesised.
• You can delete it. Deleting your voice removes the model and every sample it was built from, from our storage. Deleting your account does the same. See deleting your data.

A voice model is sensitive in a way an ordinary setting is not, which is why it is opt-in, why the deletion is a real deletion, and why we say all of this rather than burying it.

Suggestions in chats

Hello World sometimes offers a suggestion in a conversation — a reply, or an action such as adding an event to your calendar. Suggestions are produced on our servers from the conversation they appear in.

A suggestion is an offer. Nothing is sent, saved or added to your calendar unless you choose it, and you can dismiss any of them.

Suggested photo descriptions

When you choose a photo or a video for a post, or the cover photo of a store listing, and have not written a caption yet, Hello World can offer a description of the picture. The picture (for a video, a single frame) goes to our server, where an image-description model running on our own server describes it, and our translation model puts the description into your language.

The picture is not kept for this, and the description becomes your caption only if you accept it. The photos you send in chats are not described.

What it learns from you, and how to stop it

Corrections you make are kept and applied to your own translations from then on, and words you add to My Words are protected from being translated. This is per-person: your corrections shape what you read and what you send, not what other people see from other senders.

You can remove any word, and you can tell Hello World to forget everything it has learned from your corrections. Both are in Settings › Translation.

What Hello World does not do with AI

• No profiling of you for advertising. There is no advertising network in the app.
• No automated decisions that affect you — an account is never restricted or closed by a machine. AI tools can help us read through a report; what happens to an account is decided by a person.
• No face recognition. The camera effects run on your phone and identify nobody.
• No model looks at the photos you send in a conversation. The one place a model looks at a picture is a suggested description, above.
• No training on your content.

If AI produces something wrong or offensive

Report it in the app, the same way you would report anything else — see safety. That includes a translation that says something you did not, and anything produced by a synthesised voice. Reports are how we find out that a model is behaving badly, and a translation that changes the meaning of a message is worth telling us about even when nobody meant any harm.