1. Where AI Is Used
AI is not one feature on OnFire. It appears in several places and they behave differently, so it is worth being specific.
| Feature | What it does | Runs where |
|---|---|---|
| Fire assistant | Conversational assistant you can chat with, and which can carry out actions in your account. | Our servers & a third-party model |
| Voice assistant | Spoken conversation — speech to text, a model reply, and a synthesised voice. | Our servers & third parties |
| AI music | Generates music tracks from a text description. | Third-party service |
| AI image and ad media generation | Generates images for listings, adverts, and posts. | Third-party service |
| Writing help | Suggests captions, product descriptions, and profile text. | Third-party model |
| Transcription | Converts voice messages and calls to text. | Third-party service or your device |
| Safety scanning | Assesses conversation text for safety risks. See the Community Guidelines. | Your device, with our own server as fallback |
Under the EU AI Act, OnFire is a provider of these features, not merely a user of someone else's model. Building the assistant on a third-party model does not change that; putting it into service under the OnFire name does. So the transparency duties in Article 50 fall on us directly rather than on the model vendor.
Article 50(1), which requires that you are told when you are interacting with an AI system, applies now. Article 50(2), which requires generated audio, images, video and text to be marked in a machine-readable form, applies to our generative features from 2 December 2026. Section 4 sets out honestly where we currently stand against it.
2. What Happens to Your Content
When you use an AI feature, the input it needs leaves your device:
- When you message the assistant, the recent history of that assistant conversation is sent so it can answer in context — not just your latest message.
- When you use the voice assistant, your audio is sent for transcription and the transcript is sent to a model.
- When you generate music, an image, or text, your prompt is sent to the provider that produces it.
- When safety scanning cannot run on your device, the conversation itself — up to 200 messages, both sides, the words as written, with timestamps — is sent to a server we run ourselves and stored there. Only participants' names are replaced with “Speaker A/B”; the messages are not redacted. This is triggered by your device being unable to run the local model, not by anything being flagged.
Your ordinary chats are not sent to third-party AI providers. Messaging a friend does not put that conversation through an external model. The assistant sees an assistant conversation; it does not read your other chats unless you ask it to act on them.
They may still be sent to us. The safety-scanning fallback above transmits and stores whole conversations on our own servers, and it is not conditioned on anything being flagged. “Not shared with third parties” is not the same as “stays on your device”, and this page should not be read as saying the second.
Retention of the data we hold, including safety-scanning transcripts, is covered in our Privacy Policy.
3. Training
OnFire does not train its own AI models on your content. The models we use are built by other companies and used as supplied.
Whether those providers may use what we send them to improve their own models depends on our contract with each of them. If you would prefer your prompts not to reach an external provider at all, do not use the generative features — nothing in ordinary messaging sends your conversations to them.
4. Labelling AI-Generated Content
Music generated with AI music, and media generated for adverts, carry an AI Generated label in the app where they are created.
That labelling does not survive sharing. Once AI-generated media is posted, sent in a chat, or downloaded, there is no marker telling the person receiving it that a machine produced it. Text written with AI writing help is not labelled at all, and we do not currently embed provenance data or a watermark in generated files.
So the accurate statement today is that generated content is labelled at the point of creation, not that it is labelled. Persistent, machine-readable marking is what Article 50(2) will require of our generative features from 2 December 2026, and building it is work we have ahead of us. In the meantime, if you share AI-generated media, please say so.
5. What the Assistant Can Do on Your Behalf
This deserves its own section because it is the part people most often misunderstand. The Fire assistant is not only a chatbot that answers questions. It is connected to tools that let it take actions in your account — reading and changing your data in the same way you could yourself.
Those actions run with your own permissions. The assistant cannot see or change another person's data that you could not see or change. But within your account, an instruction to the assistant can result in a real change, not just a reply. Treat “delete that” or “send it” as instructions that may be carried out.
If you would rather it did not act, ask it questions rather than giving it instructions, and check the result of anything consequential.
6. Limits — What Not to Rely On
AI output on OnFire is generated by statistical models. It is often useful and sometimes confidently wrong. In particular:
- It can state things that are simply untrue, including about real people, businesses, prices, and availability.
- It is not professional advice. Do not rely on it for medical, legal, financial, or safety decisions.
- Transcription makes mistakes, and a mistaken transcript can invert the meaning of what was said.
- Safety scanning both misses things and flags things wrongly. It is an aid, not a guarantee that what reaches you is safe.
- Generated images and music may resemble existing works or real people without that being intended or apparent.
Where an AI feature produces something you are going to publish, sell, or act on, check it first. You remain responsible for what you post, including when a model wrote it.
7. Ownership and Third-Party Rights
Generated output can resemble existing copyrighted work, and generated likenesses can resemble real people. Neither we nor the providers can guarantee that an output is free of third-party rights, and in several countries purely machine-generated output may not attract copyright protection at all.
Using AI features to imitate a specific living person's face or voice, or to pass generated media off as a genuine recording, breaches our Community Guidelines.
Ownership of generated output, and who bears the risk of an infringement claim, are dealt with in the Terms of Service. This page describes what the features do; the Terms allocate the rights.
8. Your Choices
The generative features are opt-in by use: if you do not open the assistant or ask for a generated image, no prompt of yours is sent to a generative provider.
Safety scanning is different. It is not opt-in, and there is no setting that turns it off. The switches governing it are fixed when the app is built, so there is nothing you can change. We think scanning for serious harm without an opt-out is defensible, but you are entitled to know that the choice does not exist rather than to look for a setting that is not there.
There is also no control today to delete your assistant conversation history separately from the rest of your data, and no per-feature AI toggle. To delete your data generally, see the Privacy Policy.