We claim no ownership of your work
We claim no ownership of what you produce with the service, we do not use your projects, prompts or outputs to train models, and we do not show them in our marketing without your permission.
What we need is a narrow operational permission only: to store your work and show it to you, to send it to the provider that renders the generation, and to keep a backup of it. Nothing more.
What the providers' terms add
Generation runs at outside providers, and each of them has its own terms about what may be done with the output of its models. Your right to use your output is subject to those terms as well as to this document.
No IP indemnity from us
We do not warrant that your output is free of third-party rights, and we do not undertake to defend or indemnify you if somebody claims a right in it. Generative models can produce work resembling existing work, and responsibility for the use of an output rests with whoever uses it.
This is said plainly because some competitors sell an indemnity of this kind. We do not, and we do not want it assumed.
The label inside the file
A file a model generates for you carries, in its own metadata, a standards-registered term saying the file is or contains AI-generated media, the generator's name, and a sentence in Arabic and English. A picture an operation card makes from a generated one carries the same statement, in that vocabulary's term for a composite. The label is in the bytes of the file rather than in a download header, so it stays with the file on your disk and when you upload it anywhere else.
A file you uploaded is never labelled, and neither is a picture made only from files of your own: a label saying a model made your photograph would be a false statement about your work, written by us, in a file you then publish.
What the label does not carry is as deliberate as what it does: no provider's name, no internal identifier, no run number, no price. A file you publish carries no fact about our business or about your account.
There are formats we do not label today, and we say so rather than let you find out: we write the label into PNG, JPEG and WebP, into MP4 and the containers built like it (MOV and M4A among them), and into MP3 and WAV. Anything else leaves without it — GIF, AVIF, WebM, Ogg, FLAC or any format not named here. And even in the formats we do label, a file can leave unlabelled: if its container is arranged in a way we cannot write into safely we write nothing at all, rather than risk damaging the file.
The label is not a signature or a proof of provenance. Anyone who edits the file can strip it, and its presence proves nothing about who made the file. It is a statement we bind ourselves to, not forensic evidence.
What you must do with it
When you publish or share output, make clear that it is AI-generated, and do not strip or disable the label in order to hide that. This is an obligation some providers' terms place on us, passed on to you as it reached us.
Editing a file for production reasons — cropping, converting, compressing — may drop the label without anyone intending it. What is required is that it not be removed in order to hide the fact, not that files go unedited.
Other markings in the file
Some of the providers we render with add an invisible watermark of their own to the output; it is always on and we cannot switch it off. It is not ours and not under our control, and we cannot promise that a file leaving this service carries no trace of which system drew it.
We say it because not saying it would turn supplier invisibility into a promise we cannot keep. What we do undertake is that a provider's name appears nowhere in our interface, nowhere in a file's metadata, and in no response from our service; the one place we name them is the privacy notice, which lists them and says what each of them receives.