Legal
AI disclosure
How Try-On's results are made, that every model in them is generated rather than photographed, how each file is marked as generated, and who has to say so to whom.
Published in English. English governs this document, whatever translation is provided.
Draft. This text has not been reviewed by a lawyer and is not the agreed version. It says what we intend to agree, so that it can be read and corrected before Try-On goes on sale.
Last changed: 5 October 2026
Everything Try-On makes is made by a machine. This page says so plainly, in one place, so that a merchant can link to it and a shopper can read it.
1. What the results are
A Try-On result is a still image or a short video generated from what a merchant sends: a photograph of a garment, the values of a body profile, and the length and quality they choose.
The person wearing the garment does not exist. They are generated, not photographed, and they are not meant to depict any real person. Any resemblance to a real person is unintended. The one exception is where a merchant sends a photograph of a real person, which they may do only under that person’s written release, as the content policy and clause 9 of the terms require.
The garment is real. It is the merchant’s own product, from their own photograph. A result shows a generated person wearing a real garment, and a generated person is not a photograph of a fitting.
2. What a result does not prove
A result is an illustration, not a measurement. It does not show how the garment will fit a particular shopper, and it is not evidence of fabric, drape, weight or colour accuracy. A merchant who publishes one is responsible for it representing the product they are actually selling.
3. How we mark a generated file
Every result we deliver is marked as artificially generated, in the file, in a form a machine can read, and we provide a way to detect that marking. The marking travels with the file: it stays when the file is downloaded, stored in a media library and served from a storefront.
Where the Artificial Intelligence Act (Regulation (EU) 2024/1689) applies, this is the duty Article 50(2) places on the provider of a system that generates synthetic images or video, and it has applied since 2 August 2026. The Regulation sets the result — effective, interoperable, robust and reliable as far as is technically feasible — and not the technique: the Commission’s guidelines of 20 July 2026 and the code of practice on marking and labelling published on 10 June 2026 describe how it is read, and no particular standard is mandated. Marking and detection are two duties, not one, so a marking nobody can check is not enough.
California asks for the same two things, and for one more. The California AI Transparency Act (Business and Professions Code section 22757 and following) has been operative since 2 August 2026, and on 30 September 2026 the threshold that had limited it to systems with a million users was deleted, so it reaches a service of any size that is publicly accessible in the state. It requires the marking to carry particular information — who provided the system, which system and version, when the content was made, a unique identifier, and whether the system created or altered the content — and it requires a free, publicly accessible tool that anybody can use to check a file and get that provenance back. That tool is the same detection duty the Artificial Intelligence Act implies, so one piece of work answers both.
Honest about the state of it: the marking and the detection are being built into the render pipeline, and they are a condition of the service being offered in the European Union or to California merchants, rather than something we will add afterwards. Until this page says otherwise without this paragraph, treat it as what we are committed to rather than as a description of a finished system.
Marking is not the same as a label a shopper sees. It is for software: a platform, a tool or a regulator checking whether a file was generated.
4. Who has to tell whom
| Duty | Whose |
|---|---|
| Marking each file as generated, in machine-readable form | Ours |
| Telling shoppers, where a published result has to carry a disclosure | The merchant’s |
| Telling us the truth about a photograph of a real person | The merchant’s |
A merchant publishing a result is the one speaking to shoppers, and the law puts the disclosure on whoever publishes. We cannot do it for them, and our marking does not discharge it. In particular:
- Article 50(4) of that Regulation requires the deployer of a system producing a deep fake to disclose that the content is artificially generated or manipulated.
- New York requires an advertisement to disclose conspicuously that a synthetic performer appears in it, where the advertiser knows one does (General Business Law section 396-b, in force since 9 June 2026). A generated model is a synthetic performer within that definition.
- California requires the same from 1 January 2027, in wording substantially similar to “this performance features a synthetic performer” or “no human performer is depicted”, where the synthetic performer is prominent in the advertisement, whether or not the advertiser knew (Business and Professions Code section 17610 and following).
- More places are legislating the same way. We say what we know; a merchant sells where they sell, and keeping to the law there is theirs.
Saying a result is AI-generated does not make it lawful to have made it. Where a real person’s likeness is involved, a label cures nothing: the question is whether that person agreed, and in several places the answer must be a written release obtained before the likeness was made. The disclosure duties on this page and the consent rule in the content policy are separate obligations, and meeting one says nothing about the other.
A generated person must never be presented as a customer, a reviewer or anyone with an opinion about the product. A generated model wearing a garment is an illustration. The same figure saying the garment fits well is a fabricated endorsement, which is a different thing and is not something Try-On may be used for.
5. On this site
We are not only the provider of the system; where we publish results of our own, we are a deployer of it too, and the disclosure duty is ours. So: the sample results on try-on.co are generated, and no human performer is depicted in them. The pages that show them say so where they show them, which is what a disclosure at first sight means.
6. What we do not do
- We do not train models on a merchant’s images, body profiles or results, ours or anyone else’s. Clause 11 of the terms says so, and the providers who run the generation for us are bound to the same.
- We do not identify anyone. Nothing in the service compares a face against a database or builds a template to recognise a person by, which is why body measurements and ordinary photographs are not special categories of personal data in our hands. The privacy policy explains what we do hold.
- We do not decide anything about a person. No result affects anybody’s access to a service, a price they are offered, or any other decision about them.
7. Telling us something is wrong
If a result misrepresents you, your product or a person you are responsible for, write to [email protected]. Rights holders in the United States may also use the notice procedure in clause 13 of the terms. We act on reports, and we would rather hear about one early.