AI transparency

Article 50 of the EU AI Act obliges providers of AI systems to inform users when they are interacting with AI and to mark AI-generated content so it can be identified as such. This page is our plain-language fulfilment of that obligation, together with the in-product labelling we describe below. The obligations under Art. 50 apply from 2 August 2026; Sleqk launches designed to comply from day one.

We do not train our own foundation models. We use Google’s Gemini family of models via the paid Gemini API. At the date at the top of this page, the specific models are:

• gemini-3-flash-preview— for text generation (prompt enrichment, outfit rationale, short suggestions).
• gemini-3.1-flash-image-preview— for image generation (try-on renders, background removal, garment visualisation).
• text-embedding-004— for text embeddings used in the recommendation shortlist.

If we change model versions we will update this page. The system architecture (three separate endpoints, each with its own circuit breaker so that one outage cannot starve the others) is documented in our engineering docs.

Concretely, these models power four user-visible features:

Catalogue understanding

When you photograph a garment, the image model removes the background, and the text model infers attributes (category, colour, fabric family, formality) so the item can be tagged and searched.

Recommendations

For a given event, we embed the event description and run a cosine-similarity search against the embeddings of your closet, then ask the text model to rank and explain the shortlist. If the AI path is unavailable we fall back to a rule-based shortlist and, at the last resort, to pre-written suggestions; we log internally whenever we fall back so we can tell.

Photoreal try-on rendering

The image model renders the garments from your closet onto the profile photo you uploaded. The output is an AI-generated image; it is not a photograph of you wearing that outfit, even if it looks like one.

Phenotype extraction

We ask the text model to read your profile photo and extract a compact set of styling-relevant attributes (body shape family, skin-tone range, hair colour family) so the try-on looks right on you. This only runs if you have given explicit consent to phenotype extraction in the profile screen.

Every try-on render carries three layers of marking, so at least one survives wherever the image ends up:

• Visible, on-pixel watermark.A small “AI-generated — Sleqk” mark is burned into every shared or exported render. Most social platforms strip metadata when you post; the on-pixel watermark is our failsafe for that.
• C2PA provenance sidecar. Each render is accompanied by a Coalition for Content Provenance and Authenticity manifest, signed, describing that the image was generated by Sleqk using a specified Gemini model version. Tools that understand C2PA (including some browsers and social platforms) will show this badge.
• In-product and in-web labelling. In the Sleqk app and on share pages like /l/[token]on sleqk.com, every AI-generated image is displayed next to a pill that reads “AI-generated” and, where available, “C2PA signed”.

There are things we will not do with AI, regardless of demand:

• We do not train any model on your photos or your renders. The paid Gemini API tier we use is covered by Google’s no-training policy; we do not operate our own training pipeline. [LEGAL-REVIEW: confirm Google’s paid-tier training carve-out wording at launch and keep the text here in sync.]
• We do not allow try-on renders of another person’s face or body without that person’s explicit consent, and our terms forbid using Sleqk for non-consensual imagery or impersonation.
• We do not generate political, election, or campaigning imagery.
• We do not generate sexual or suggestive content, and we do not allow content that depicts minors in any styling context.
• We do not use AI to make decisions about you that produce legal or similarly significant effects (for example pricing, access to credit, employment). Sleqk is a styling assistant, not a decisioning system.

Every AI output in Sleqk is reviewable by you before it has any effect:

• You choose which events to pass to the stylist, and which garments are in your closet in the first place. Nothing is uploaded or inferred without a conscious action on your part.
• You can reject any outfit recommendation, regenerate any try-on render, and delete any render that you do not like.
• The server-side circuit breakers around our AI provider are rolling-window rate-based: if failure rates exceed a threshold, the system falls back to rule-based shortlists and suggestions rather than continuing to hammer a degraded model.
• We monitor a “fallback used” signal so that if AI paths silently degrade, we see it.

Under GDPR Art. 22 you have the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects you. Sleqk does not make decisions of that kind, but the right still exists — if you believe that Sleqk is doing so, write to privacy@sleqk.com. More broadly:

• You can withdraw AI consents (phenotype extraction, photo enhancement, try-on rendering) in the profile screen at any time.
• You can object to any AI processing we do on the basis of legitimate interests under GDPR Art. 21.
• You can request a human review of any AI output you feel went wrong.

Being honest about what the AI cannot yet do is part of transparency, so:

• Try-on renders work best when your profile photo is well lit, front-facing, and full-length. Complex patterns, very reflective fabrics, and loose flowing garments are the hardest to render faithfully.
• The catalogue tagger occasionally misclassifies garments — you can always correct the tags by hand.
• Recommendations improve as your closet grows and as the product learns which outfits you actually wear. Expect a bland early period.
• AI outputs are probabilistic. Two identical runs of the same prompt can produce different renders.

If a Sleqk-generated image feels wrong (misrepresentation, discriminatory output, safety concern), write to privacy@sleqk.com with the share token or an in-app screenshot and we will investigate. We log serious incidents internally so we can improve guardrails over time.