Responsible use

6.7Broader impact and responsible use

The programme aims to make a widely deployed technology work for people it currently serves poorly, wherever they shop. It does not claim a specific reduction in returns. That claim needs live retail trials, which are outside this programme's scope.

Try-on also carries real risks, and live video raises them. It edits images of real people, so it could be misused to alter someone's appearance without consent, to sexualise a photo, or to push a narrow body ideal. FabricVTON has written a guardrail plan for photo and live try-on, with eleven guardrails and a tested decision function, enforced in a layer that every request must pass [111].

6.7.1 The main measures

• Minors are never processed. Two age estimators and a challenge age of 25 mean anyone who may be under 18 is blocked with any garment, in every mode, and no client setting can change this.
• Revealing garments are restricted. Lingerie, swimwear and sheer garments are off for consumers and allowed only for clearly adult, consented catalogue models, with stricter output checks.
• Inputs and outputs are both screened. Nudity checks on the person and garment photos, output moderation, an age re-check, and a check that the output shows no more skin than the garment explains. Research shows harmless-looking garment images can make a try-on model produce nudity, so the output check matters most [119].
• Known abuse material is hash-matched and reported, with evidence preserved in a separate, locked account.
• Consent and provenance. A consent check at every upload and every live session, a block on public figures, and a visible AI label, signed Content Credentials and an invisible watermark on every output, in line with Article 50 of the EU AI Act [69]and India's 2026 rules on synthetic media.
• Access control and takedown. Per-client keys, rate limits, signed jobs so the GPU cannot be reached around the checks, penalties for repeat abuse, and a 2-hour takedown process.

6.7.2Live-session controls

Live sessions add their own controls: the age and consent checks run at session start, sessions have a time limit, output frames are sampled for moderation throughout, and the stream carries the label and watermark. The fairness measures from WP1 are applied to FabricVTON's image, video and live models before every release, and uploads are deleted after each job under a published retention policy.